For most of the 20th century, getting something done that you didn't fully understand meant finding a person who did, and letting them operate the interface on your behalf. Take booking trips before travel aggregators: you didn't compare hundreds of fare classes across a dozen airlines; you just called a travel agent, named the destination, and they navigated a reservation system you never had to see. Before the web, booking a flight meant visiting a travel agent who consulted a printed schedule book on your behalf — speed improved over decades. Still, the shape of the interaction didn't change: you spoke your intent, and someone else operated the tool.

That model held until self-service software made it economically pointless. Online travel agencies let ordinary consumers reach the same reservation systems agents once monopolised, and the effect was brutal on the intermediary layer — U.S. travel agent employment fell from a peak of roughly 340,000 around 2000 to roughly 60% below that by the mid-2020s, driven by airlines cutting agent commissions and by Expedia-style platforms proving people would rather click than call. Banking followed the same arc on a different clock: ATMs first made tellers more productive by taking over routine cash transactions, but mobile banking finished the job — commercial bank branches per capita have fallen by nearly 30% since peaking in 2009, and Bank of America alone cut headcount from 288,000 to 204,000 between 2010 and 2018 as branch visits stopped being necessary at all.

Life and Work through a Single interface

My life, just as much as yours, is split between a bunch of different interfaces where I keep things and get things done. Booking a weekend getaway via a hotel booking app, flights via a flights app, keeping notes, ideas, tasks, and thoughts in various documents on my Google Drive and some of them in the Notes app, and my business and personal schedule in Google Calendar. Hours of my life are spent just keeping track of everything and making sure that no important e-mail got missed, an important call got scheduled, and all that wouldn’t collapse or conflict with one another. A while ago, I wrote about context stitching being an outcome of the absence of one single product that would be solving the issue from top to bottom. Now, with AI racing through various sets of problems with enormous speed that makes everyone feel vertigo, we finally get to witness a gleaming solution to one of the nastiest problems — we figured out a way to manage everything through a single interface.

Grok Bot

Grok Bot is xAI’s answer to this, launched into public beta on August 11, 2026, as part of the newly formed SpaceXAI group — Elon Musk’s merger of xAI, SpaceX, and Cursor into one vertically integrated stack. It isn’t a chatbot you prompt. It’s a roster of persistent, named AI agents, each with its own cloud computer, that you hire, train, and hand real work to.

Why Grok Bot’s “Let someone else handle it” actually lives up to expectations?

Computer use

Grok Bot has its own computer, where it does things for you - just as if any human personal assistant that you hired, who would have their own personal computer and be there to get things done for you, with 24/7 availability and unlimited capacity.

Every single time you request it to run a task, it does it through what we now call 'computer use', when an agent can execute a task, like finding you plane tickets or buying you groceries - it actually browses the web and navigates an interface like a human would, scrolling through pages. It's very amusing to watch Grok going through a desktop window. Any other cloud agents are mostly accessing other apps through MCP or CLI, which works like an API call - it only works where a structured integration for that exact app and action already exists, and it breaks or simply can't reach the destination the moment it hits a site or workflow nobody's wired up yet. On top of that, you can show how you do things to Grok and teach him a task using his computer - your screen will be recorded, analysed, and Grok will execute the steps the same way you normally do it. It feels just like onboarding a new employee.

Wide range of integrations

Grok Bot has 20-plus integrations with other platforms like Figma, Notion, Gmail, Google Calendar, GitHub, and Salesforce, which allow seamless and cohesive navigation of all sources required for executing a task and closing a full loop without leaving “one last thing” that you need to do yourself manually.

It does payments on your behalf

We finally saw something that ChatGPT was trying to launch within Operator in January 2025, but couldn't quite make it work in a full end-to-end manner. What the SpaceX team was focused on is figuring out the last bottleneck, which is making a payment - by partnering with Stripe, they were able to set up a workflow that allows Grok Bot to make a payment on your behalf after your authorising it in the Grok Bot interface. Stripe’s Link issues a one-time virtual card for that specific purchase (merchant and amount) once you approve it, which Grok Bot enters at checkout without ever holding your real card. This works well with tasks like recurring routine shopping that you don’t want to babysit end-to-end (like buying one particular set of one-day contacts every 3 months or a cat litter refill every single month delivered to your address — you still approve each charge) or a one-off purchase that you don’t want to deal with yourself (like buying train tickets on a Spanish website for one particular day of your Euro trip leg).

Churn risk:

1. Context-insufficiency execution errors.
   The most common case is when a user hands off a task without enough context (a name spelling, a contract clause, which "John" they meant) and the Bot executes anyway, confidently, because nothing told it to pause and ask. The counter-lever is a context-completeness check before anything leaves the system, not a review of the output after it's gone.

2. Payment autonomy without a way back. The approval tap only catches an intent mismatch, not an execution mismatch — a user can approve “buy the flight” and still get charged for the wrong date or a duplicate booking, and because the charge runs through a single-use virtual card, there’s no issuer on the other end to reverse it the way a chargeback would. Confirm it by watching the ratio of “wrong but technically approved” refund requests against total Grok Bot-initiated transactions — this category has no fraud protection to fall back on, since the transaction really was authorised. The counter-lever is a cooling-off window above a set dollar amount, where the charge holds instead of settling before the user has a real chance to catch it.

3. Site friction and auth handoffs.
   The chances are high for a Bot to get blocked mid-flow by a CAPTCHA, a “verify you’re human” wall, or a login / 2FA / SSO / passkey step only the user can finish. The task was going fine until the screen demanded a human, and every handoff is a chance they drop it and don’t come back. The counter-lever is designing for graceful pause-and-return (clear “your turn” handoffs, preserved cart/state, one resume path) rather than treating a blocked page as a failed agent.

I lived with Grok Bot for 2 weeks and handled various tasks I would have handled myself previously - either fully manual or partially done in Claude. I tried to live all my life’s admin through one single interface of Grok Bot, and here’s what I learned…

Breakdown of Grok Bot is coming next week!

Claude Cowork

Claude Cowork was born when Anthropic noticed people bending Claude Code, its terminal coding agent, into a general-purpose task tool - so it shipped the non-coding version in about ten days, built using Claude Code itself, and launched it on January 12, 2026.

The connector-first architecture

Cowork doesn’t reach for computer use by default. When you hand it a task, it follows a strict priority order: connectors first, computer use only as a fallback. If a connector already exists for what you’re asking — Gmail, Google Drive, Slack, GitHub, Google Calendar, Microsoft 365 — Cowork uses that connector to complete the task through the app’s own API, the same way any integration would. Only when there’s no connector available for the job does it fall back to interacting with your screen directly.

Beyond the built-in connectors, Cowork can also reach custom or remote MCP servers, which route through Anthropic’s cloud infrastructure rather than your local machine — true even when you’re running Cowork locally. Through a broader integration layer, that extends to more than 1,500 managed third-party integrations beyond what ships natively.

Where the infrastructure won’t go

A few boundaries are built in deliberately:

• Consequential, unattended actions. Anthropic’s own safety guidance is explicit: don’t schedule tasks that make purchases, send messages on your behalf, or take other hard-to-reverse actions. This applies in every approval mode, including “Skip all approvals.” Permanent file deletion is the one action that always requires confirmation, regardless of mode.

• Anything outside your own access. Since permissions are inherited rather than granted independently, there’s no path for Cowork to reach a file, channel, or account you personally can’t already open.

• Unattended screen or file work. As above — no desktop app open, no computer use, no local files, full stop, regardless of what mode you’ve set or how the task is scheduled.

Custom-built agents

Trailblazer companies, like Lovable, Stripe, and AI-first companies that had a chance to think of all their workflows and processes reimagined and had an obvious advantage of being able to build everything from scratch, thought about a set of business problems that are similar in nature to what consumers are struggling with. Context stitching between several sources (design lives in Figma, comms live in Slack and email, finance lives in Stripe… list goes on and on), pouring data from one interface to another, having to type everything into each dedicated field, or having to delegate research tasks to someone with taste and judgement several times a day. The fact is that for them these problems are way more urgent, and business and financial incentives push them harder to fix those. Time is precious, and the price of taking too long to execute literally kills startups in this excellence race. Those companies are building internal systems, with the key execs shipping those themselves, that are customized to what their day-to-day workflow is like. And they're small enough to be able to do it fast and get everyone on board with that.

1. Lovable Head of Design, Felix Haas () has built custom agents that bring all his daily stack down to a single interface with several.

2. Department of Product () shared how Uber is running more than 8,000 different agents internally, with McKinsey, Figma, and Ramp cited alongside it as companies building bespoke agent stacks instead of buying an off-the-shelf assistant.

3. Sharadh Krishnamurthy's engineering team at Stripe built what they internally call a "company brain" - shared context, governance, and a skills platform baked directly into how the whole org works, not a personal hack but a company-wide system.

Agents Talking to Agents

I’ve been experiencing this part for the past two weeks, where there’s an AI agent doing things for me: opening support tickets, scrolling the internet, and placing orders on my behalf. The obvious direction once you've watched it happen enough times: the next stop is my agent talking directly to the business’s agent, with neither of us watching the exchange happen in real time. That layer already exists in early form. Google unveiled the Universal Commerce Protocol at NRF in January 2026 - an open standard, backed by more than 20 partners including Shopify, Target, Stripe, Visa, and Mastercard, built specifically so an AI agent can discover, compare, and complete a purchase without a human ever loading a product page. OpenAI and Stripe built a competing version, ACP, sitting behind ChatGPT’s Instant Checkout, already processing something like 50 million shopping queries a day. And on the enterprise side, agents built on Microsoft Copilot or Salesforce Agentforce are already reordering office supplies and renewing SaaS subscriptions without a human approving each one - business to business, agent to agent, no browser in the loop at all.

It certainly raises a question nobody in marketing or growth has fully answered yet: if a meaningful share of shopping and browsing moves onto an agent that a business’s website never actually renders a page for, what happens to the businesses that don’t adapt - or that actively block agent traffic to protect their own analytics? Do they just quietly stop showing up in the consideration set, the way a vendor with no product feed on Google’s or OpenAI’s protocol becomes invisible to a shopper who’s delegated the comparison entirely?

And on the churn side, the problem gets wilder: how do you even notice a customer going cold when they never log in, never open the app, never generate a single session - because the entity doing all of that is their agent, running on a schedule, and the human hasn’t touched the product directly in weeks? The usual churn signals - declining logins, shrinking session length, fewer page views - stop meaning anything the moment the agent is the one logging in. Somebody is going to have to invent agent customer experience as its own discipline: what does it mean to delight an agent, keep it coming back, and notice when it’s quietly switched to a competitor on your behalf without your user ever making a conscious decision to leave?

For the past three years, the entire AI-product industry has been centered around one belief: access to a model is the advantage. Teams ‘tokenmaxxed’, burned as many frontier-model tokens as they were able to afford (until they couldn’t afford any longer), often encouraged by management, as proof that they were being ambitious and AI-pilled. The model was the moat (at least it felt like it).

The Fable 5.1 drop on Sept 1 shows that belief is outdated. Anthropic shipped its new flagship, and the same day it was live on AWS Bedrock, Google Vertex, Azure Foundry, the Vercel AI Gateway, OpenRouter, LiteLLM, Cursor, GitHub Copilot, Microsoft 365 Copilot, Amp, Kilo Code, Perplexity, Harvey, and Lovable. The best model on earth had a competitive half-life measured in hours. When every product can offer the identical frontier model as fast as it’s available, "we have the best model" is no longer a unique selling feature or a moat.

While I was writing this article, right when I was 80% through, I got an email that landed in my inbox with a headline stating that OpenAI just released Astra, a new model that is the closest thing to AGI we've seen yet.

It has been called "the world's most intelligent and aligned model" - the closest thing to AGI anyone has shipped. And here's the tell: it wasn't consumer-available on release. It rolled out to a limited set of organizations first, with access for ChatGPT Plus/Pro/Business/Enterprise users promised: “over the coming days."

You'd expect the closest-to-AGI model to be the thing everyone would be willing to get on board with. For a regular consumer, it mostly won't be, because what actually keeps an everyday user isn't the raw intelligence ceiling; it's the interface gravity of the workspace they already live in.

For someone whose work and life are already set up inside one agent window - where they don't pick a tool and go build, they just keep talking, and the thing to build materializes in the same chat when there's context for it (a rendered HTML artifact, a PDF, a doc that persists, work that lives on in a Cowork-style workspace) - a marginally smarter model behind a different door might not be an incentive to move.

Valuemaxxing is the new flex

As frontier tokens got expensive, the flex inverted - from “I run the latest, most powerful model” to “I know which model to run for each task.”

The category’s center of gravity is moving off the model and onto the layer that decides, routes, remembers, and integrates. Now, the experience layer is the new moat.

Whenever you open Instagram, you don’t want to think or prompt that you like watching videos with silly dogs or cooking recipes; you just want it to be shown to you.

Just like this, retention belongs to whoever makes the model invisible - the product that routes to the right model, at the right cost, inside a workflow the user won't need to rebuild - because the user stops being able to tell (or care) which model they're on, and starts staying for everything wrapped around it.

A model can produce an impressive answer and still leave you with a broken app. It can write convincing code that never runs or create a finished-looking checkout page that saves the wrong data. On a benchmark, the response may look excellent. In the browser, it may be another story.

Here, three products solve the problem of maximizing value for money, while also easing decision fatigue.

Cursor

AI-native code editor
Founded 2022 in San Francisco by four MIT students (CEO Michael Truell). Cursor is an AI-first IDE - autocomplete, chat, and agentic multi-file editing built into the editor — and became the default AI coding tool for a huge swath of professional engineers. Revenue ramp is historic: ~$100M ARR (Jan 2025) to roughly $4B ARR by May 2026. Funding went $900M Series C at $9.9B (June 2025) → $2.3B Series D at $29.3B (Nov 2025).
Anysphere was reportedly acquired by SpaceX in a ~$60B all-stock deal that closed in mid-August 2026, with Cursor now operating inside the SpaceXAI division. They released a Cursor Router (July 2026): automatic per-request model selection across Intelligence/Balance/Cost modes, trained on 600k+ live requests, claiming ~60% lower cost.

Job it does: "I want the right model for this exact task. And I don't want to read a benchmark leaderboard every morning to get it."

It translates to:
You write a request
The Router reads the task, not just the prompt length
It routes to the model that clears the bar for the lowest cost
It can switch models mid-task as the work changes
You never touch the dropdown.

The router was trained on 600,000+ live requests - proprietary usage data no rival has. The model belongs to 3d party, but the routing intelligence is what Cursor owns.

Best users:
Engineering teams with a real token bill
Eng managers tired of “which model should I use?”
Startups optimizing burn
Enterprises standardizing a model policy

Churn watch: The risk arrives the first time a senior engineer suspects “Cost” mode quietly downgraded a task that mattered — a subtly wrong refactor, a shallow answer on a hard bug. The metric is override rate: how often power users flip Auto back to a named model. Once that climbs, the router has stopped being a convenience and become a thing to babysit — and a router you babysit is worse than no router. The counter-lever isn’t a smarter model, it’s transparency: showing which model ran and why, so an automatic choice stays legible. (Cursor’s ~60%-cheaper number is its own and third-party-unverified — that gap is the trust it has to earn.)

Lovable

AI app builder for non-engineers.

Founded 2023 in Stockholm by Anton Osika and Fabian Hedin, growing out of their open-source GPT Engineer project. It turns a plain-language description into a working full-stack web app - auth, database, APIs - with no traditional coding, aimed squarely at non-technical builders. It's been one of the fastest-scaling software companies on record: ~$100M ARR (July 2025) → $200M (Nov 2025) → $500M ARR by May 2026, with ~8M users. Funding tracked the curve: $200M Series A at $1.8B (July 2025), $330M Series B at $6.6B (Dec 2025), and a $400M Series C at a $13.3B valuation in August 2026 (led by Menlo Ventures) - ~$950M raised total. Latest product moves are the heart of its strategy: its August manifesto "The model picker is a dead end" and a "control plane" that auto-assigns build steps to different models, plus adopting Claude Fable 5.1 same-day on Sept 1.

Job it does: “I want the app I pictured to exist by tonight, and I don’t want to know, choose, or think about which model is building it.”

Primary intent: Build the thing without getting into the weeds of tech.

It translates to:
You describe what you want
A control plane breaks the build into parts
Each part goes to whichever model is best at it, invisibly (to a user)
Failures feed the eval loop, not your lap
You get an app, not a model decision.

Best users:
Non-technical founders
PMs prototyping without engineering
Designers shipping real products
Solo operators who’d never open a model dropdown

Churn watch: Lovable’s risk is the vibe-coding cliff - the wow of the first generation fades, and retention lives or dies on the second and third session, when the app has to keep working and evolving. Because the model is hidden, every failure is charged to Lovable, not to “a bad model” - the user can only say, “Lovable didn’t make it work.”

The metric is return-to-build rate: if fewer than ~30–40% of first builds come back for a second within the first couple of weeks, the magic was a demo instead of a habit. The counter-lever is making the output improvable in place, so the user never wishes to look under the hood.

Zapier

workflow automation / the "glue" layer

Founded in 2011, a Y Combinator (S12) company, by Wade Foster (CEO), Bryan Helmig, and Mike Knoop. Zapier connects thousands of SaaS apps so non-developers can automate multi-step workflows ("when X happens, do Y") without code - the original no-code automation layer. It's famously capital-efficient: roughly $310M+ ARR on about $1.3M ever raised, last valued around $5B (2021 tender offer), and still independent. Its recent direction is a full lean into AI: Copilot (describe a workflow in a chat window, and it drafts the Zap, which you then build from or tweak), Zapier Agents, and a pointed "valuemaxxing" positioning - automate the deterministic work and only spend a model where a task actually needs judgment, claiming ~71% average cost savings across modeled agent workflows.

Primary intent: Judgment only where it’s needed.

Job it does: “I want AI on the parts of my work that actually take judgment.
And I don’t want to pay a model to ‘think’ about a task that has exactly one right answer.”

It translates to:
Map the process into steps
Deterministic steps run as plain automation, no tokens
Only the judgment steps call a model
Swap the model behind a step later without rebuilding
The whole thing runs cheaper than pushing all of it through AI.

Best users:
Ops leads watching an AI bill climb
Automation builders wiring real workflows
Small teams with no ML budget
Anyone who’s watched an agent burn tokens on a rules problem

Churn watch: Credit where it's due - Zapier already attacks the blank-canvas problem. Its Copilot opens the interaction in a chat window, drafts the workflow from a plain-language description, and gives you the choice: build from zero, or take the draft and tweak it. The remaining friction is subtler, and it adds to the churn score: the draft still hands you a workflow you have to understand, verify, and own - the copilot writes the first pass, but you're still accountable for whether the logic is right and which steps quietly needed AI. So the churn moment comes when fully-agentic tools that "just do the whole thing" make even reviewing and maintaining a draft feel like overhead. The signal is how far users get from "described it" to "trusted it in production" without hand-holding. The counter-lever is more proof: closing the gap between drafting the workflow and trusting it runs correctly, so tweaking reads as control instead of extra homework.

Decision fatigue has never been more real

Being a human and a professional in 2026 is the most exciting thing that ever existed.

The abundance of opportunities is dizzying. One can build a career in any space with an option to work from anywhere in the world. You have access to all the books ever published and the smartest AI models and tools right from your computer. There’re thousands of ready-to-go ideas flowing on social media for how to make money, get a job, build a startup, or land a promotion than ever before. It has never been harder to avoid getting lost among a thousand ideas and opportunities, to decide on what you want to work on, to commit and stay focused without drifting off topic for long enough before actually seeing the compounding results.

All that creates enormous decision fatigue, which decreases our efficiency by the same proportion that AI amplifies it. In this context, having fewer things to decide on while interacting with a product is something that could make a significant difference. I am sure that when you open your laptop in the morning to get something done - create a report, prepare an SOW for a client, test a new acquisition channel - you don’t think to yourself: I want to become an expert in AI. All you want is just to get things done in the fastest, most energy-efficient way possible.

Tara Seshan, Product Lead at OpenAI, says that their North Star is that users do not have to decide between Chat or Work (which is Codex undercover), but at this moment it’s still on the user to understand what they’re trying to do and match it with the capabilities and limitations of both products.

The AI-product stack didn't converge into one super-tool. It split along the lines of what each player decided to own when the model itself stopped being a competitive advantage.

When the best model is available to everyone within hours, having it stops being a moat, and how you decide to spend it - whether a task needs a model, which model it needs, and whether the user ever sees that choice - becomes the entire game.

In 1966, a graduate student in psychology at MIT confided in a therapist named ELIZA. She typed that she was having problems with her boyfriend, and the program asked her to share more. She kept talking for several minutes, in real detail. Underneath the terminal, ELIZA was doing nothing more than scanning her words for keywords and handing them back as questions. It had no memory of what she'd said thirty seconds earlier, and no model of her, her boyfriend, or "boyfriends" as a concept. If it found no keywords, it just offered a generic response.

To someone who's spent much of their career in CX, this evokes the first customer-facing support chatbots. They all had a bit of ELIZA in them: too many questions, no memory, infinite patience, zero opinions. In the 1960s, that combination may have passed for a rough model of empathy. It obviously leaves a lot to be desired.

In 2005, IKEA was the first company to ever put a support chatbot in front of one of its customers. A lovely lady, this time named Anna, gladly accommodated user questions with a variety of answers about IKEA products, prices, sizes, delivery, spare parts, and opening hours. This was enough to ease the anxiety of customers in their post-purchase journey as they awaited quirky yet minimal furniture to arrive at their door.

However, customer expectations began to outpace reality extraordinarily fast.

Silicon Valley began scaling product capability faster than support could add capacity.

One-click purchases meant that a billing dispute, a fraud flag, or a wrong address was no longer a problem to be solved for ten customers an hour, but a problem which needed to be solved for twenty thousand customers per minute.

This turned the entire experience of being an online consumer into the worst one you could possibly have. If you’ve ever been a customer of Amazon, Uber, Airbnb, any cellular provider, any airline, any bank — any business operating at a scale where you’re one face among millions — you know the specific dread of getting on the phone with their support.

According to the latest Parloa research, when presented with alternatives to being put on hold, customers are willing to make remarkable sacrifices.

29.9% respondents claimed they would switch brands entirely if the company were to put them on hold, while 21.1% would be even willing to wake up at 5 a.m. instead of being left hanging. In March 2017, Servion Global Solutions predicted that by 2025, AI would power 95% of all customer interactions, and that the experience would be so seamless that customers would be "unable to spot the bot."

Now, let’s take a moment to do a reality check. Today’s state of AI is indeed at the height of its capabilities and beyond what we imagined it would be able to do by 2026. One single prompt run by Claude’s Fable 5 can now deliver passable product launch videos. Long-running agents can autonomously execute a single complex task for hours at a stretch. At the 2025 International Mathematical Olympiad, AI models sat the real contest. No internet, no tools, two 4.5-hour sessions back to back. Both Google DeepMind’s Gemini Deep Think and an OpenAI research model earned gold-medal scores. And yet, as recently as 2018, the people closest to this technology were telling us to expect all of it would take decades longer to arrive.

We definitely didn’t see it coming.

Subscribe now

In a 2018 study, a consensus of machine learning scholars put “high-level machine intelligence” (AI that outperforms humans at every task) about 45 years out, roughly 2061. Even that came with only a 50% probability attached.

That wave of AI transformation hit customer support hard. It had to. Communication and issue resolution is what call centers and support teams do all day, at huge and growing volume, 24/7. On paper, that’s exactly the kind of task AI should replicate easily.

Your own experience as a consumer probably tells a different story.

Talking to a chatbot too often feels like a waste of time. “Press 1 to talk to a human” is most people’s first choice if they actually want the problem solved. Self-service menus work fine for a password reset. They don’t work when you’re standing on the shoulder of a highway waiting for a tow truck.

The data backs this up. Just 7% of respondents said that chatbots consistently resolve their issue, and 50.7% assumed their problem would be too complicated for automation before they even tried. That consistent frustration led many customers to draw the wrong conclusion: that the technology itself was bad.

So if AI capabilities keep outpacing the timelines we set for them, why haven’t we figured out how to make voice and chat bots reliable, or even just useful? Where’s the gap that remains unclosed?

Well, there’s a short answer and a long answer.

More than ever, humans desperately need other humans to solve their problems. Do users actually need human judgment? Or do they just want a better experience?

After a decade being on the other end of CX, and a lifetime of being a customer, the bad reputation of chatbot support is a signal of two things:

1. Chatbots have now become a synonym for low-quality, get-outta-here, ‘we-don’t-actually-want-to-talk-to-you’ kind of service.

2. What customers actually need is to have a positive AI experience, one which doesn’t suck - to rebuild the trust we lost.

I was wrestling with the bad chatbot problem while wearing the Head of Support hat at a $400M dating and social app holding company, serving 500M users, back in 2024. At the time, there were virtually no companies that could consistently deliver something that felt even remotely close to talking to a friend on the phone.

Leave a comment

The few companies working in this space were still building old-school voice trees: four buttons to reach a human. That’s frustrating enough when your app crashes mid-chat. It’s much worse when someone’s had a safety issue: matched with a person, met them, and felt physically unsafe, only to get a phone robot on the other end.

Understandably, users say they hate AI, hate bots, hate automation, and demand a human instead.

What users are actually saying is: we’re tired of bad CX. There just hasn’t been enough proof that AI can do the job well.

If we decode this message, what the consumer asks for is:

Speed — my request gets processed quickly.

Comprehension — my request is understood the first time.

Reliability — whenever I leave and come back, it’s stable, and I don’t have to repeat myself or re-explain the context.

Resolution — my issue actually gets solved, and I’m happy!

And all of that should happen within a single interaction.

In short, consumer expectations of what an AI voice agent should deliver are higher than ever, maybe a little higher than what current companies can actually deliver. What's missing, at least for now, is trust: not the glamorous kind tied to model intelligence, but the boring kind, the kind that has nothing to do with how smart the model is and everything to do with whether the thing on the other end remembers you, understands you the first time, and actually finishes what it started.

Whether a company like Parloa closes that gap first is a separate question from whether it's worth trying. The capability ceiling is already proven. Gold medals, marathon coding runs, models that can reason their way through a graduate exam: none of that is in question anymore. What's still unproven is whether anyone can turn that capability into something a stranded, angry customer on the side of a highway actually trusts. That's not a model problem. It's a company-building problem, and right now almost nobody is building for it. Someone will close this gap. The only question is who gets there first, and how much longer customers have to keep pressing 1 for a human while they wait.

For ten years, I ran support teams whose accumulated knowledge was the only reason nobody ripped out the costly CRM underneath them, and not once did anyone in those rooms call that a moat. Between November 2025 and June 2026, four enterprise software companies sold at a premium because that particular moat stopped holding, and every one of them was still growing on the day the deal was signed.

Two routes back to a barrier

Hamilton Helmer’s 7 Powers names seven conditions that produce persistent differential returns: Scale Economies, Network Economies, Counter-Positioning, Switching Costs, Branding, Cornered Resource, and Process Power.

Acquisition is not among them, and the omission is deliberate. Power in Helmer’s model requires two things at once: a Benefit that materially improves cash flow and a Barrier that stops competitors from arbitraging the Benefit away, and his own non-arbitraged test disqualifies any company that gains preferential access to a desired resource but pays a price for capturing the value it was buying. In an efficient market for companies, the premium goes to the seller. Which is what makes a $3.6 billion cheque interesting: nobody writes one unless the Barrier needs attention.

When a Barrier starts to erode, there’re two routes back to it.

The Rebuild Path means constructing the new layer yourself, on your own roadmap, with your own engineers, and accepting that the market will move underneath you while you ship. Salesforce took this path with Agentforce, which reached $1.2 billion in ARR in Q1 FY27, up 205% year over year, and which asks enterprises to build their own agents on a customisable platform. Adobe took it with LLM Optimizer, a product aimed squarely at brand visibility inside language models. Both worked. Neither was fast enough.

The Repurchase Path means buying the layer customers are already anchoring their habits to and absorbing the Barrier wholesale. What is striking about the last eight months is that the same companies walked both paths in sequence, rebuilding first and repurchasing second, which tells you the rebuild was the thing they tried before the clock ran out.

Switching costs were never built by the software. They were built by the effort users spent learning it. Agents absorb the effort, and the barrier leaves with it.

The test any operator can run on their own product this week: name what a customer would have to relearn if they left you this quarter, then ask honestly whether an agent could do that relearning for them over a weekend. If the answer is yes, what you have is not a switching cost. It is a habit, and habits are arbitraged.

Fin, acquired by Salesforce for $3.6 billion

Fin is optimised for what I’d call packaged resolution and does its job extremely well.

Primary intent: Resolution without deployment.

Job it does:

“I want tickets to stop reaching my team at all. And I don’t want to spend two quarters building the agent that makes that happen.”

It translates to:

Ingest the knowledge base
Resolve across every channel
Escalate only the residue
Report cost-to-serve

Best users:

• Mid-market support leaders

• Post-Series B ops teams

• Companies with high ticket volume and thin headcount

• Anyone whose CFO has asked about cost per contact

Fin’s agent, running on a proprietary model called Apex, resolves an average of 76% of support volume end-to-end across live chat, email, WhatsApp, SMS, phone and Slack, and handles more than two million conversations weekly across 30,000 organisations. Salesforce did not buy the capability because Agentforce had the capability. It bought the packaging, and packaging is where habit forms, because the mid-market never builds what it can deploy on Tuesday.

Intercom was founded in 2011 by Eoghan McCabe, Des Traynor, Ciaran Lee and David Barrett, did roughly $200 million in revenue by 2021, and then did something almost nobody with a decade of brand equity does: it renamed the entire company after one product, the agent, and pushed past $400 million in ARR on the strength of it. In March 2026 it raised $250 million in venture debt from Hercules Capital. In June it sold.

Churn watch: the risk arrives at the resolution ceiling. Fin looks extraordinary against a world of all-human support, but once 76% automated becomes the floor, the buyer takes it for granted and starts judging Fin on the hard quarter it cannot close, the ambiguous 24% where every vendor is weakest. And because everyone now clears roughly the same 76%, the headline number stops being a reason to pick Fin over a rival; the lead rests on a model that is only briefly ahead. Category leadership, measured in resolution rate, has a shelf life exactly as long as the gap between frontier models.

Semrush, acquired by Adobe for $1.9 billion

Semrush is optimised for what I’d call visibility measurement and does its job extremely well.

Primary intent: Discoverability assurance.

Job it does:

“I want to know whether anyone can still find us now that our customers ask a model instead of a search box. And I don’t want to infer it from analytics that only count the traffic which already arrived.”

It translates to:

Crawl the index
Benchmark against competitors
Diagnose the visibility gap
Hand the fix to the publishing stack

Best users:

• In-house SEO teams

• Independent agencies

• Enterprise brand marketers

• Content operations leads

Adobe’s portfolio was built to act rather than to see, with Experience Manager to publish, Commerce to sell, Experience Platform to unify and Brand Concierge to converse, and none of it able to tell a brand whether that work was being surfaced at all. Semrush closed the loop, which is why Adobe’s own framing points at customers running LLM Optimizer alongside Semrush AIO to move from visibility intelligence to deployed optimisation to measured outcome, against a backdrop of AI traffic to US retail sites rising 269% year over year as of March 2026. The price was $12.00 per share in cash, a premium near 77%, against $471.4 million of ARR. Roughly four times revenue for the instrument panel of the agentic web.

Churn watch: Semrush was not losing customers in any dramatic way, and that is exactly why the risk is easy to miss. The clearer tell is that the customers it already had were barely spending more than the year before, so even though total revenue grew, almost all of that growth came from signing brand-new accounts and nudging existing ones onto more expensive plans, not from the core base becoming more valuable on its own. The fast growth was concentrated at the top, among the biggest spenders. At the same time, the ordinary middle of the customer base stayed flat, which is really a bet that a small number of large corporate accounts will keep carrying the whole company, and that bet only holds while those big accounts keep buying. The deeper problem is what Semrush is actually for: its entire business is helping brands get found when people search, and people are increasingly asking an AI assistant instead of searching at all, so when Semrush launched its own app inside ChatGPT, it was planting itself inside the very shift most likely to make its original job disappear.

Semrush was founded in Boston in 2008, went public in 2021, and by the end of 2025 had grown AI product ARR from $4 million to more than $38 million in twelve months while enterprise ARR quadrupled from $9 million to $37 million across 579 customers. Profitable on a non-GAAP basis, growing 18%, category-leading. It sold anyway, to a buyer whose $20 billion Figma bid had collapsed under regulatory pressure in December 2023 and who therefore had to buy something small enough to survive review.

Cursor, acquired by SpaceX for $60 billion

Cursor is optimised for what I’d call delegated authorship and does its job extremely well.

Primary intent: Engineering throughput.

Job it does:

“I want the tedious sixty percent of this codebase written without me. And I don’t want to hand my repository to a vendor my security team has never heard of.”

It translates to:

Index the codebase
Suggest inline
Run multi-step agents in parallel
Review before merge

Best users:

• Enterprise platform teams

• Fast-shipping startups

• Staff engineers running large refactors

• Anyone whose backlog grew faster than their headcount

Cursor’s revenue curve is the most violent in modern software, moving from $500 million to roughly $4 billion annualised inside a year, $2.6 billion of it enterprise. Underneath that curve, its competitive position was slipping: Claude Code, launched only in May 2025, overtook Cursor on developer preference through late 2025, and by May 2026 Ramp corporate spending data put Cursor at roughly 26% of AI coding spend with Anthropic at about 13% and closing. Revenue was doubling while the thing that predicts revenue, who developers actually reach for, was moving the other way.

Churn watch: the moment the agent stops being the reason and starts being the interface. When a developer’s loyalty attaches to the model doing the reasoning rather than the editor framing it, the editor becomes a skin, and skins are switched in an afternoon. Cursor could not build the one Barrier that would have held, its own frontier model, so it rented one.

Anysphere was founded in 2022 by four MIT students, Michael Truell, Sualeh Asif, Arvid Lunnemark and Aman Sanger, launched Cursor in March 2023, raised an $8 million seed led by the OpenAI Startup Fund, and by November 2025 priced a Series D at $29.3 billion. In April 2026 SpaceX paid roughly $10 billion for exclusive collaboration rights plus an option to buy the company outright at $60 billion, pre-empting a round pricing near $50 billion. SpaceX exercised on 16 June 2026, four days after the largest IPO in US market history, in all stock, at about fifteen times revenue, for the coding data that trains Grok. The Cornered Resource was never the editor. It was the corpus.

Six mechanics behind these deals

1. Effort is the barrier, not the software. Every switching cost these three companies enjoyed was denominated in user effort, the ticket macros a support team wrote, the audits an SEO lead ran weekly, the shortcuts a developer memorised, and an agent that absorbs that effort dissolves the barrier without ever competing on features.

2. Packaging beats capability in a commoditising category. Agentforce could already do what Fin did and still could not reach the mid-market, because a platform that asks you to build the agent loses to a product that arrives having already built it, and the $3.6 billion was the price of that distinction.

3. Distribution is the terminal moat. Fin defined a category, hit $400 million in ARR, renamed itself after the product, and still sold, because when the category commoditises the surviving Barrier belongs to whoever owns the customer relationship rather than whoever owns the better model.

4. The measurement layer outlives the execution layer. Adobe could rebuild publishing, commerce and personalisation faster than it could rebuild Semrush’s decade of crawl history, which is why the instrument panel cost $1.9 billion while the things it measures are Adobe’s own products.

5. Revenue growth is a lagging retention signal. Cursor quadrupled ARR while losing a third of its market share, and the only reason we can see that at all is Ramp’s spending data, because ARR is the number a company reports and share is the number that predicts what it will report next.

6. The exit multiple prices the Barrier, not the Benefit. Semrush went at roughly four times ARR, Fin at about nine, Cursor at fifteen, and the spread has almost nothing to do with product quality and everything to do with how defensible each company’s position looked to the person writing the cheque.

   
   

   Thanks for reading Do Not Churn! Subscribe for free to receive new posts and support my work.

If you found this useful, here’s what else Do Not Churn has covered:

• The collapse of earned activation: Why agents that perform the work destroy the attachment the work used to build, and the one motion worth protecting from automation.

• Churn is just graduation you didn’t design for: The churn event that never shows up in cancellation data, because the user outgrew the tool without ever leaving it.

• AI takes the thinking out, how do we get it back?: What three AI tutors reveal about the difference between delivering a result and building the model that makes the result stick.

For a decade, I helped engineer activation in dating and social apps, and it always came down to getting users to do the work themselves. Swipe enough that the algorithm learned your taste, and the faces started matching what you actually wanted. Start three or four conversations, so there is someone worth coming back to. Keep circling back to look for someone better.

Even if that was not an entirely evil mission, the arrival point for a user happened to be a permanent search that gave the sense of progress - searching was the product, and a user who stopped searching was one the product had already lost. Every motion in that loop was the user’s own, and the doing was the attachment. I wrote in an earlier issue about how three decades of dating apps expired their users’ trust, rewarding motion over outcomes until people stopped believing the product was moving them anywhere and slowly disengaged.

At least the user did the searching themselves, and that effort built a layer of real attachment as a byproduct, the profiles they wrote, the conversations they tended.

I have now had the same conversation a dozen times with teams wiring AI agents into that loop, and it always opens the same way: our activation numbers have never looked better, so why is retention falling? What AI agent that takes over the doing that builds attachment, is now manufacturing the feeling of value with the user barely doing anything.

The pattern repeats.

The activation chart flashes green, the onboarding team takes the QBR victory lap, and a few weeks later, the same cohort is retaining worse than the clumsier flow it replaced. Teams reach for the usual explanations: a messaging problem, a pricing problem, a notification-cadence problem, and spend a quarter tuning the wrong dial.

The problem isn’t that your activation is no longer working.

Put plainly:

the user got the result without ever doing the thing, so there is nothing for them to come back to.

For fifteen years, activation worked as a proxy for retention because the early action it measured was an action the user performed, and the performing was the attachment, so when the agent does that action for the user, that proxy breaks.

Before you can fix activation, you have to answer a simpler question:

How much of the work does your product still leave in the user’s hands?

That single distinction predicts almost everything that matters for retention: whether the user learns anything, whether they own anything, whether a habit forms, and whether your activation number means what you think it means.

If you’re staring at a great activation number and a sinking retention curve, here’s what you’ll have by the end of this post:

1. A one-sentence test and a five-minute audit prompt to locate activation debt in your funnel.

2. The earned-to-delivered spectrum, so you can place any AI feature and predict its grip before you ship it.

3. Six install-ready levers to rebuild an earned moment without adding extra work back in.

All by recognizing that activation is not one event. It is a spectrum of how much the user earned, and only the earned end of it predicts retention.

Why this matter

RevenueCat’s 2026 State of Subscription Apps report found AI-powered apps lose subscribers about thirty percent faster than non-AI apps at the median, with refund rates roughly twenty percent higher, even as they convert trials better, and a16z now tells teams to measure AI-product retention from month three because the early weeks are inflated by what they call AI tourists. High intake, fast leak. The acquisition engine works; the retention engine is missing a part, and the assumed part is earned effort.

The framework: three kinds of activation

Every AI feature you ship lands somewhere on one spectrum, from work the user still does to value that is simply gifted to the user’s hands with no effort.

I’m breaking them down into three profiles to make it simpler to approach them and find a way to treat. All have a different retention profile, a different way to measure it, and a different collapse reason.

Type 1: Assisted activation
The agent accelerates the work that the user still performs.

What this is: the user stays the operator. The agent speeds up, completes, or drafts, but the user reviews, decides, and integrates every cycle, and something the user owns accumulates underneath them as they go.

How to spot it: the user could not get the outcome by doing nothing; their judgment is in the loop on every pass; an asset such as a codebase, a document, or a board compounds as they work.

The live example: Cursor.

Built by Anysphere, founded in 2022 by four MIT students and led by Michael Truell, Cursor forked Visual Studio Code and rebuilt it around AI, becoming the fastest software company ever to reach $1B ARR, roughly eighteen months, and $2B ARR by early 2026 at a $29.3B valuation.

The agent writes code and finishes your edits, but you read every diff, accept or reject it, and commit it to your own repo. It is delivered at full speed, and it still retains, because the developer is doing the work, and the codebase is the kept switching cost.

How to measure it: the share of agent output accepted without a single edit. A stable, moderate number is healthy; the leading risk indicator is that share is climbing over time.

Warning sign and the fix: when the user stops reviewing and starts accepting everything, they slide from operator to passive reviewer, the deskilling that controlled studies have already measured sets in, and the grip erodes. The fix is to keep the review an active step rather than letting the user hand off all decision power to an agent.

Type 2: Ambient activation
Value arrives with no user action at all.

What this is: the product is wired into data the user already has, so it produces value unprompted. There is no first action, no setup, nothing the user assembled.

How to spot it: the user did nothing to start it; the value appears without being asked for; if it disappeared tomorrow, the user might not even notice.

The live example: Gemini Personal Intelligence.

Google’s, launched in January 2026 and wired straight into the accounts you already have, it draws on your Gmail, Photos, YouTube, and Search history, your bookings, your purchases, and what you watch, without being asked. There is no first action to complete. The value simply starts arriving in users’ hands.

How to measure it: the gap between reach and protected behavior. Wide reach with nothing the user would fight to keep is the signature of maximum debt.

Warning sign and the fix: invisible value cannot be missed, so it builds no loss aversion and no switching cost, and a rival assistant wired to the same data is a flat substitute. The fix is to make the user curate what it learned, confirm it, correct it, shape it, so a sense of authorship forms after the fact and leaves them something of their own to lose.

Type 3: Earned activation
The user does the work and owns the result.

What this is: the user assembles, configures, or contributes something themselves, and that effort is what produces both the value and the attachment at the same time.

How to spot it: setup takes real work; the user supplies or builds something; what they end up with feels like theirs.

The live example: OpenClaw.

An open-source agent built by developer Peter Steinberger that you run on your own machine with your own API keys, OpenClaw is the hardest tool here to start and the one its users abandon least.

A running USER.md accumulates your preferences and patterns until, after roughly thirty days, the agent behaves in a way no fresh install could imitate.

How to measure it: setup-completion rate against thirty-day-plus retention. Expect brutal front-loaded drop-off and an unusually flat curve for everyone who crosses the threshold.

Warning sign and the fix: the risk is that setup is so hard most people never activate at all. The fix is not to make setup easier everywhere; it is to preserve one act of real ownership, because a 2012 study by Norton, Mochon and Ariely, the IKEA effect, showed that labor invested in something is exactly what converts into how much it would hurt to abandon.

The prompt to detect an activation debt

The day after an agent finishes setting up your product, can the user explain in one plain sentence what it does for them? If they cannot, you delivered value and didn't develop a habit, and that gap is your activation debt.

1. To find where it is building up, paste your real onboarding into this:

You are a retention strategist. Here is my product's onboarding,
step by step: [paste every step from signup to first real use].

1. Score each step for USER effort (none / light / real work) and
   value returned (none / partial / core payoff).
2. Mark the step where the user first feels the product worked.
   Tell me whether that earned moment exists at all.
3. Flag every step an agent already does, or could do within 12
   months. For each, name what the user stops learning when it does.
4. After an agent runs this flow, could the user explain in one
   sentence what the product does for them? If not, name the habit
   that never forms.
5. Give me three task cards to re-introduce ONE earned moment the
   agent does not take: a choice, a contribution, or an ownership
   act the user keeps. Each card: title, the change to the flow,
   the retention behavior it should create.

Be specific. If there is no earned moment even today, say so first.

2. Pull two numbers for the same cohort and put them side by side: your activation rate, or for an agentic product, your signup-to-first-output rate, and your month-three retention, since week one is noise for AI tools.

If activation is climbing while month-three retention dips > the spread between the two lines is your activation debt.

Segment by whether the user hit any earned moment in their first session, the kept decision, the contributed asset, the owned artifact, and the earned-moment cohort is the one whose curve flattens instead of falling.

That single split tells you which steps to stop automating.

The split

Activation didn’t stay one thing; it split along how much of the work each product leaves in the user’s hands. Assisted tools like Cursor deliver at full speed but keep the user in the loop and let an owned asset compound underneath them, which is why they hold. Ambient tools like Gemini Personal Intelligence deliver value the user never asked for and therefore cannot miss, which is why they leak. Earned tools like OpenClaw make the user do everything and keep them through the effort that doing required.

Churn emerges when value arrives without ever being earned, because nothing learned and nothing owned is nothing to come back for, and retention in the agentic era grows when a product protects one motion the agent could have taken and deliberately leaves it to the user.

Five takeaways

• Activation is a spectrum of how much the user earned, and only the earned end predicts retention; the number alone tells you nothing.

• Assisted tools retain because the user still does the work and an owned asset compounds underneath them; speed was never the problem.

• Ambient tools leak hardest, because value the user never assembled is value they never feel and never miss.

• The hardest onboarding on a list is often the stickiest, because invested effort is what the IKEA effect converts into attachment.

• Retention follows earned effort; churn begins the moment the agent performs the action that was supposed to be the user’s.
  
  PS: this article was born under the June Portuguese sun, the week I turned thirty.

  

I remember when I was a high school kid myself, in the 2000s, online learning, which was on the rise, was mostly content delivery. I’d watch a video, or read a module, and do a quiz where the system told me whether I was right.
The intelligence in the loop was the teacher, who wasn’t in the room, and the parent, who was trying to help with long division at 9 pm using methods that no longer matched the curriculum. Learning technology existed to scale content, not to close the gap between content and comprehension.

When ChatGPT launched in November 2022, educators didn’t know whether to panic or celebrate. Here, finally, was the thing every edtech founder had been promising since the MOOC (massive open online courses) era - a genuinely intelligent system that could answer any question, explain any concept, and meet a student exactly where they were. Writing for The Atlantic, English teacher Daniel Herman declared it “The End of High School English”; in the same publication, Stephen Marche declared the college essay “dead.”

The argument spread fast in the other direction too: technologies had made rote memorization unnecessary in most cases, and our educational system simply hadn’t caught up - why spend a lesson on knowledge retrieval when access to any fact was three seconds away?

The optimists argued that we were finally separating the low-order work (remembering things) from the high-order work (thinking about them).

About this newsletter:

I’m Daria Littlefield
I spent a decade running customer operations at a $400M ARR portfolio
of 35+ consumer apps.
That taught me more about why products lose users than any framework ever could.
Do Not Churn is where I put that to use. Every week, I break down the retention mechanics behind the products we all use - what keeps people, what loses them, and why the difference is rarely what you’d expect. If that’s useful to you, subscribe below.

Leave a comment

It sounded reasonable; however, the response from institutions was less convinced: Los Angeles Unified, the second-largest school district in the US, immediately blocked access to ChatGPT from its network, and by January 2023 districts across the US, UK, and Australia had followed.

It took about two years to realize both sides had missed the real problem.

The optimists were wrong about what high-order thinking requires - you cannot think critically about something you have no internal model of, and offloading retrieval to a machine doesn’t free up cognitive capacity, it atrophies it. The pessimists were wrong about the threat - the danger wasn’t cheating, it was something way harder to measure. MIT’s Media Lab made it concrete in June 2025: across 54 students monitored over four months with EEG equipment, those relying on ChatGPT showed measurably lower prefrontal cortex activation than those using search or no tools at all, with cognitive effects that lingered after they stopped using the model - the students who never had to retrieve information failed to build the memory traces that retrieval practice produces. The brain stops practicing what it no longer needs to do. The problem was never the model, but the design pattern: give the answer, skip the struggle, get it done for me.

A small number of products looked at that interface decision and made the opposite choice.

On the surface, this looks like a feature upgrade on top of existing edtech products. In reality, it reflects something more structurally significant:

The democratization of the moment that used to require a human.

The moment of productive struggle - the point in a learning session where you’ve encountered something your existing model can’t resolve - has always been where learning actually happens. And it has always been where self-directed digital learning falls apart, because the options are: give up, search the internet, or ask someone. AI tutors are the first technology that can intervene precisely at that moment with something more than a search result. Whether they intervene well - whether they guide or give away, whether they scaffold or surrender - is the design question the entire category is now splitting along.

Brilliant

Founded 2012 by Sue Khim, Silas Hundt, and Sam Solomon. Khim - a University of Chicago mathematics dropout and Forbes 30 Under 30 honoree - built Brilliant after a previous edtech startup, Alltuition, pointed her toward the gap between access to education and genuine comprehension. 10M+ learners globally.

Brilliant is optimized for what I’d call conceptual mastery with visual guidance, and it does that job with a degree of technical precision that most edtech products can’t reach.

Primary intent: Understanding, not performance

Job it closes:

“I want to actually understand this concept so that next time I see it, I’m not starting from zero. And… I don’t want to work through yet another explanation that tells me the answer instead of showing me how to think.”

It translates to:
Enroll in the course
Attempt an interactive visual lesson
Encounter the moment of being stuck
Invoke Koji
Koji asks guiding questions and can annotate the live lesson on screen
Return to attempting the problem
Progress tracked against mastery

Optimized for:
Conceptual math and coding, Adult learners returning to STEM, Older students approaching material their school hasn’t taught well, Parents buying for curious kids, not struggling ones

Brilliant’s structural advantage is the one that took a decade to build. Koji can see everything on the learner’s screen - the exact problem, which choices they’ve made, where they’ve paused - because the lesson infrastructure was built as code from the beginning. Koji doesn’t just read a problem description; it reaches into the interactive itself and annotates it, highlights a region, poses an intermediate question. The blog post announcing Koji makes this explicit: the goal is for Koji to eventually make himself unnecessary, gradually handing the thinking back to the learner until they’re asking themselves the questions Koji used to ask. That’s the only tutoring philosophy that structurally compounds value rather than creating dependency.

This is a product with real philosophical backbone.

Churn Watch: The risk lives in the acquisition promise versus the product reality. Brilliant’s founding audience was curious, self-directed learners - people who wanted to go deeper on things they already cared about. Koji widens the aperture to include students who are stuck and need rescue. Those two populations have very different tolerances for productive friction. A learner who opened Brilliant because they wanted to understand calculus will love that Koji won’t give them the answer. A learner whose parent subscribed to them because they’re failing pre-algebra will experience that same design choice as abandonment. The churn signal to watch is session-level abandonment before Koji is invoked - which would suggest the user hits the wall and exits rather than asking for help.

There’s a second structural exposure: Brilliant courses are Brilliant’s own, not curriculum-aligned to what a student is actually studying in school. That makes the product a supplement rather than a tool with urgency, and supplements are vulnerable to “I don’t need this this week” attrition.

Khanmigo

Khan Academy founded 2006 by Sal Khan - MIT triple-degree graduate and former hedge fund analyst who started tutoring his cousin Nadia in math over the phone in 2004, posted the sessions to YouTube, and quit his job in 2009 when the demand outgrew his evenings. Khan Academy is a nonprofit; Khanmigo launched in March 2023 as its AI tutoring layer. 150M+ registered users across Khan Academy globally. Backed by Google, Bill and Melinda Gates Foundation, and others.

Khanmigo is a Socratic tutoring layer built directly on top of Khan Academy’s content library - the biggest free educational content archive ever assembled - and it is the only AI tutor on the market that has categorically refused to give students answers since its launch in March 2023.

Primary intent: Guided reasoning, curriculum-aligned

Job it closes:

“I need help right now, in the exact subject I’m studying for class tomorrow, and I want to understand it — not copy it. And… I don’t want my kid to just get better at asking AI for shortcuts.”

It translates to:
Log in via Khan Academy
Navigate to active subject
Invoke Khanmigo on exercise or reading
Khanmigo asks probing questions rather than resolving the problem
Student reasons through each step
Progress recorded in dashboard

Optimized for:
K-12 students working through structured curriculum Teachers needing lesson planning, rubric generation, differentiation School districts evaluating AI readiness Students working in their native language (Khanmigo operates in over 34 languages)

Khanmigo grew from 68,000 users in the 2023-24 school year to over 700,000 in 2024-25, expanding from 45 to more than 380 US school district partners in a single year, with projections to pass one million students in 2025-26 (K-12 Dive, 2025). That growth rate tells you something: institutional buyers don’t purchase products they believe will make their students dependent. They purchase products they believe will make their students think.

Churn Watch: The engagement gap is the most honest data point in this category. Industry reporting has confirmed that only around 15% of students in Khanmigo-enabled classrooms use it regularly, a number Sal Khan has acknowledged publicly and that is driving a product redesign shipping in summer 2026. The structural problem is not the pedagogy, but an activation model. Khanmigo waits to be asked. Most students in academic distress want fast relief, not a guide who will make them figure it out themselves; they want the answer and they will find it elsewhere if Khanmigo won’t provide it. The product’s most reliable retention vector is the teacher side - where Khanmigo for lesson planning and rubric generation has sticky, habit-level daily use. The redesign needs to solve for proactive student-side engagement; the Socratic method is only as good as the student’s willingness to start the conversation.

Synthesis Tutor

Founded 2022 by Josh Dahn, who previously ran Ad Astra - the experimental school Elon Musk commissioned for SpaceX employees’ children in 2014. Dahn built Synthesis to take the game-based, mastery-driven philosophy from Ad Astra and make it available to any family, not just those with a rocket scientist parent. Headquartered in Austin, Texas. Consumer subscription, $300/year or $95/year family plan. On pace for $10M+ revenue in 2025.

Synthesis Tutor is a K-5 AI math tutor built around the insight that confidence is the prerequisite to learning, and that for young children especially, the feeling of being capable and making progress is what everything else runs on.

Primary intent: Math confidence and fluency, ages 5-11

Job it closes:

“I want my kid to actually enjoy math and feel good at it before school teaches them they’re not. And I don’t want to drill them on worksheets they already associate with failure.”

It translates to:
Parent subscribes
Child opens colorful adaptive interface
AI tutor assigns age-appropriate math content with visual manipulatives
Voice-guided instruction walks the child through the problem without triggering frustration
Difficulty adapts in real time based on the child’s responses
Parent receives progress dashboard

Optimized for:
Young children aged 5-11, building foundational numeracy
Neurodiverse learners who need multi-sensory input
Homeschooling households wanting structured math progression
Parents who want measurable acceleration, not just engagement

When Synthesis went public in 2022, it brought a specific design philosophy with it: that schools are structured to produce average performance, and that the right technology could produce genuinely exceptional outcomes for ordinary children. When a child gets a wrong answer on Synthesis, the AI doesn’t say “try again” - it adjusts the visual representation, breaks the problem into a smaller step, and tries a different sensory approach. The system is optimized for the moment just before frustration, which is exactly the moment most kids exit.

Churn Watch: Synthesis has graduation churn written into its architecture. The product covers K-5 only, which means every family with a child approaching age 11 is on a countdown to forced attrition - as the product runs out of content for their child. The company has stated plans to expand to K-12, but until that ships at scale, subscriber tenure has a ceiling.

The second churn vector is parental confidence: without visible academic improvement within roughly 60-90 days, the subscription shifts from “investment in my child” to “expensive educational entertainment.” Unlike Khanmigo, Synthesis has no institutional safety net - every subscriber is a direct consumer decision, renewed or canceled on a parent’s subjective read of whether their child is measurably ahead.

Six Retention Mechanics in AI tutoring
for you to steal

1. Context access as a moat. The retention mechanic that most of this category can’t replicate is Brilliant’s: Koji has full visibility into the live lesson state - every choice made, every pause taken - which allows it to intervene with precision no generic LLM can match. When an AI tutor can see exactly what’s on your screen and reach into the interface to annotate it, the tutoring session is irreplaceable. This is technically non-replicable for anyone who didn’t build the underlying curriculum infrastructure.

2. Institutional lock-in as a substitute for habit formation. Khanmigo’s 380+ district partnerships mean the product’s retention is often not about the student’s relationship with the tool at all - it’s about whether the district renews. This is a fundamentally different churn model from consumer edtech, and it explains why a 15% individual engagement rate is survivable. Institutional churn operates on annual cycles and is driven by administrator satisfaction, not student activation.

3. Emotional confidence as the actual product. Synthesis has identified something the rest of the category under-invests in: for young children, the affective experience of the learning session is the product, not the curriculum. A child who finishes a Synthesis session feeling capable will return without being asked. A child who finishes a session feeling inadequate has already churned - they just haven’t cancelled yet. Building the AI to intervene before frustration, rather than after it, is a retention lever disguised as a pedagogical choice.

4. The MIT study as institutional tailwind. In June 2025, MIT’s Media Lab published research showing that students who relied on ChatGPT showed lower brain activity in the prefrontal cortex compared to those who used search or no tools, with cognitive effects that lingered even after they stopped using AI. This study did not hurt this category - it turbocharged it. Every institutional buyer now has a citation that justifies choosing a guided AI tutor over a generic LLM. Products whose design philosophy predates the study suddenly have peer-reviewed wind in their sails.

5. The anti-dependency frame as an acquisition mechanic. All three products lead their positioning with some version of “we don’t give you the answer.” This is now the category’s primary conversion message, targeted directly at the anxiety every parent has about AI making their child intellectually lazy. But it cuts differently depending on who’s hearing it: a parent choosing Synthesis is reassured; a student in distress choosing between Khanmigo and ChatGPT at 11 pm will choose ChatGPT. The frame that wins on acquisition may not be the frame that wins on retention if the student population skews toward distress rather than curiosity.

6. Progress visibility for the buyer, not the user. In consumer edtech, where the payer (parent) and the user (child) are different people, retention depends on keeping both satisfied - and they measure different things. The child needs to feel competent and not bored. The parent needs to see evidence that the money is working. Products that give parents rich, specific, visible progress signals buy themselves substantially more time to prove value, because the parent’s renewal decision is not purely based on the child’s enthusiasm for opening the app.

Read more:

Every time I get to learn an unfamiliar piece of software, I find myself in the same loop. I open the app, I get confused by something within the first ninety seconds, and close the window, rumbling to myself: “Okay, it's harder than I thought”.

I open another tab, search “how to do X in Y,” scrub through three YouTube tutorials at 1.5x speed, watch a TikTok of someone’s hand pointing at a button I can’t find in my version of the app, lose my place twice trying to bring the original window back into focus, and return to the project with a partial memory of what I was supposed to do.

It multiplies by every prosumer tool the modern creator (or solo-founder) is supposed to know - DaVinci, Blender, Figma, Logic, After Effects, CapCut, Notion, Linear - and the cumulative cost of self-learning becomes the reason most of us know one or two tools well and twenty others badly. It appears that the friction isn’t in finding the answer, but in holding two windows open in your head at the same time.

That wasn’t always the case. It was, actually, worse.

In the 1990s, learning unfamiliar software meant buying a 600-page Wiley manual or watching a corporate training VHS at half speed in a fluorescent-lit conference room. The friction was so high that most professional software was learned in cohorts - you took a class, you sat next to a colleague who already knew Photoshop, or you didn’t learn it at all.

Then the long pivot began. Manuals became PDFs; PDFs became help websites; help websites became Stack Overflow threads; Stack Overflow became YouTube tutorials; YouTube became 90-second TikToks of someone’s hand showing you where the bevel modifier lives in Blender. The cost of finding the answer collapsed at every step. The cost of interrupting your work to find it did not.

The volume of available help exploded, but the shape of the friction stayed exactly where it was for thirty years.

That’s the gap Clicky closes. Open DaVinci Resolve, hit Control + Option, ask the small AI named Clicky what the color wheels are for, and a glowing blue triangle flies across your screen and points at the exact button you need to press. Six days after launch, the $9/month product hit $1,000 in monthly recurring revenue. The launch tweet pulled 2.9 million views. The day after that, the founder open-sourced it.

Farza Majeed, raised in Pembroke Pines, Florida, started his first company at 13 - selling blank DVDs and t-shirts on eBay, hitting $100K in revenue by 15. CS at the University of Central Florida; then deep learning engineer at Mayhem, CTO at Kanga, and in December 2019, he founded Buildspace - the YC- and a16z-backed school that raised $10M to teach people how to build their own ideas. Clicky is what happens when the founder of a learning school decides the school should live next to your cursor.

The companion product in this story has a different shape. Screenpipe is the local-first memory layer that records everything happening on the user’s screen and microphone, 24/7 - built by Louis Beaumont, a former French intelligence engineer and Techstars alum based in San Francisco. Founded in 2024 under Mediar, Inc. with co-founder Matthew Diakonov and backed by LG Ventures, Screenpipe went from weekend project to 15,000 GitHub stars, 100,000 users, and $100K in revenue inside its first weeks. Same category as Clicky on the surface. Opposite contract underneath.

Hard to imagine we ever lived without a teacher who sees what we see and points at what we need.

Subscribe

Clicky
the teacher who lives next to your cursor

Clicky is optimized for what I’d call companionable real-time teaching and does its job extremely well.

Primary intent: Learning by doing, with company.

Job it does:

“I want to figure out the unfamiliar piece of software I just opened, in real time, while I’m using it. And… I don’t want to leave the app I’m in, watch a 12-minute YouTube tutorial, and try to remember what the person said by the time I get back.”

It translates to:
User hits Control + Option

Clicky screenshots the screen

User asks the question

Claude reads the screen, ElevenLabs voices the response

A glowing blue triangle flies across the screen and physically points at the UI element the user needs to interact with

The optional “clicky agent” command spawns a background agent to research, build, or automate something autonomously.

Best users:
Solo creators learning unfamiliar professional software (DaVinci, Blender, Logic Pro, Figma)

Founders researching competitors and wanting a second pair of eyes on what they’re reading

Hobbyists picking up a new tool they’ll use once and then put down

Kids building games who need someone to explain what they’re looking at

Churn watch: After the third or fourth time the triangle points at a button, the magic compresses into utility - and utility competes on price. The 10-message memory window means longer learning arcs cannot be sustained inside one session. Because Clicky runs on Claude, Anthropic could ship a first-party version of this UX and compress the moat overnight. The defensible position is not the screen-pointing mechanic - it’s the agentic layer the user discovers on the second visit.

Screenpipe
the memory layer that watches in the background

Screenpipe is optimized for what I’d call silent total recall and does its job extremely well.

Primary intent: Continuous memory of everything that happened on the user’s screen.

Job it does:

“I want to be able to ask ‘what did that client say about pricing in our call three Thursdays ago’ and get the actual answer. And… I don’t want to take notes during the call, or invite a bot to the meeting, or trust that I’ll remember what mattered later.”

It translates to:
Screenpipe runs locally in the background

Captures screen and audio 24/7

Indexes everything into a searchable local database

The user queries it later in natural language

Developers extend it with “pipes” - custom modules that act on the captured data.

Best users:
Founders and operators who live in back-to-back calls and need cross-meeting recall

Researchers who watch hours of recorded interviews across sessions

Developers building agentic tools on top of their own activity stream

Privacy-conscious knowledge workers who want a memory layer without the cloud

Churn watch: The shape of churn is the opposite of Clicky’s. Adoption friction is high - installing a local daemon, granting screen and microphone permissions, watching disk usage climb. The user has to commit before they have proof of value, because the value compounds only after the first month of accumulated history. If the user doesn’t query Screenpipe within the first 30 days, they uninstall before the corpus is big enough to be useful.

Subscribe

Why the category keeps reinventing itself

The dream of building a small mechanical companion is older than software.

What Rifkin names is not a technology but forever present, recurring human appetite - the desire for a small, attentive, slightly mechanical helper-figure at the edge of one’s work has shown up in every century where the tools to build one existed. Hero built automata for temples. The clockmakers of 18th-century Europe built mechanical boys who wrote out poems. Microsoft built Clippy and Cortana. Now Farza has built Clicky.

The category keeps reinventing itself because it keeps finding it’s way to work. The technology gets close enough to a real assistant that users get excited, and then close enough to a real annoyance that users uninstall. Clippy was retired in 2007; Cortana was shelved in 2023. The April 2026 Tom’s Hardware retrospective on Clippy’s 25th anniversary called it “Microsoft’s hapless office assistant” - a face users hated because the assistant pretended to be a friend without earning the right to be one.

Clicky understands what Clippy didn’t: the triangle is not a face, and the buddy framing lives in the copy rather than in a pair of cartoon eyes. The contract - I summon you, you appear, you help, you leave — avoids the failure mode Clippy died of. Screenpipe understands what Cortana didn’t: it makes no pretense of personality. It is a database with a daemon, and the contract - I trust you in the background, I consult you on my schedule - avoids the failure mode of an ambient assistant that keeps interrupting to remind you it exists.

Two thousand years after Hero, the lesson is the same: users want the helper to know its place.

The six retention mechanisms inside screen-aware AI

1. Summoning ritual (Clicky’s Control + Option hotkey) - Every interaction is opt-in, which prevents the uninvited-help failure mode that killed Clippy. Why it prevents churn: the user never feels watched when they didn’t ask to be watched.

2. Visual pointing (Clicky’s blue triangle) - The AI doesn’t describe where to click; it points. This compresses the overhead of translating a verbal instruction into a physical action. Why it prevents churn: the post-wow plateau is delayed because the gesture is genuinely faster than any alternative form of help.

3. Voice-first interaction (Clicky’s ElevenLabs layer) - The user speaks the question and hears the answer, never breaking focus on the screen they’re trying to learn. Typing would force a return to the keyboard and a context switch; voice keeps the eyes where the work is.

4. Compounding personal corpus (Screenpipe’s local indexed history) - Every additional day of recorded screen and audio makes the next query more valuable. The switching cost is not technical; it’s historical. Why it prevents churn: the longer the user stays, the more expensive leaving becomes.

5. Local-first architecture (Screenpipe’s no-cloud commitment) - For a user about to grant 24/7 screen and audio recording, the only acceptable trust frame is “this never leaves my machine.” Local-first is not a feature; it’s the precondition the use case requires. Why it prevents churn: a privacy breach, real or rumored, would collapse the entire user base overnight. Local-first defuses that risk structurally.

6. Agentic delegation layer (Clicky’s “clicky agent” command, Screenpipe’s “pipes”) - The product earns its second visit by being more than a teacher. Both products turn the screen-aware base into a launchpad for something stickier. Why it prevents churn: graduation churn - the student outgrows the teacher - is defused when the teacher becomes an assistant.

I have spent a decade inside the rooms where retention gets engineered. Most of what happens in those rooms is tactical - cancellation flows, re-engagement emails, paywall friction tuned to the nearest decimal. The interesting stuff is structural: the decisions made years before a churn problem surfaces that determined whether users could even leave. The skill marketplace is one of those decisions, and it’s happening right now, in public, and most people watching it are reading it as a developer story, when the real story is a retention one.

Two companies using free, open infrastructure to create the kind of lock-in that doesn’t feel like lock-in, because users aren’t trapped. They’re invested.

The most durable retention mechanism is not a feature. It’s a library of work the user has already done inside your ecosystem - where the files are portable but the institutional investment that produced them is not.

Two approaches to the same strategic move

Platform companies have always built retention the same way: the more a user deposits into a system - workflows, templates, accumulated output - the higher the switching cost becomes, and the higher it becomes without the user ever feeling pressured.

The Accumulation Approach is what Notion ran with their template gallery - templates solved the blank-canvas problem for new users while creating a secondary behavior: users who found a template they loved started building on top of it, customizing it, making it theirs. The template was the acquisition hook; the customization was the retention mechanism.

Figma ran the same play with community files. Shopify ran it with their app store, which grew to 8,000+ apps and became the most-cited reason mid-market merchants chose Shopify over alternatives even when the platform was more expensive.

The underlying mechanic is always the same: make it easy to start, make it rewarding to build, make the output of building portable only in theory.

The Standard Approach looks more generous - you publish not just a library but an open specification, and explicitly invite competitors to adopt it. Anthropic did this with MCP, donated it to the Linux Foundation, watched it become the de facto standard for how AI agents access tools, then ran the same play in December 2025 with Agent Skills. When the standard wins, the company that authored it earns ecosystem gravity regardless of which tool any individual deploys to. The accumulation happens at the format level, not the platform level.

That wasn’t always the case with AI tooling. The current form of skills is the third attempt to solve the same problem, and each failed attempt makes the strategic logic of this one clearer.

OpenAI’s Custom GPTs were the first attempt. Three million were built within two months of the November 2023 launch; the benefit was real.
The barrier wasn’t: no persistent memory, lives only inside the ChatGPT interface, resets to zero with every conversation.

A switching cost with no accumulation mechanic is just a habit, and habits break.

Anthropic’s October 2025 Agent Skills fixed the container - skills were persistent, versioned, composable - but kept them Claude-only, which meant a barrier bounded by one vendor’s market share, fragile the moment a viable alternative arrives.

About Do Not Churn

Do Not Churn is a newsletter about why products retain or lose users - written by , who spent a decade leading customer operations across a $400M+ ARR portfolio of 35+ apps. If you work on retention, product strategy, or the structural design decisions that determine whether users stay, this newsletter is sent out weekly so you can stay ahead of the curve on the best anti-churn practices and apply them. Subscribe not to miss the next issue.

Subscribe now

The December 2025 open standard changed which power was being built: portability across 32 agent environments activates Network Economies - value that scales with the whole ecosystem, not one platform’s subscriber count - while the proprietary governance layer on top captures Switching Costs for organizations that build their security processes around it. The open standard is what makes the proprietary layer worth owning.

On the surface, this looks like developer tooling - open source, community-driven, a rising tide that lifts all agents. In reality, it reflects something more structural:

The skill library is the new template gallery. The platform that owns the library owns the switching cost.

Anthropic
the standard as platform lock-in

Anthropic’s skill play is optimized for what I’d call enterprise workflow capture - and the design choices make the commercial logic plain.

“I want Claude to know our company’s processes, brand guidelines, and internal workflows without me having to re-explain them every session. And I don’t want to rebuild that context every time we onboard a new tool or upgrade the model.”

It translates to:
Build internal skill library
Deploy to team via admin settings
Agent invokes skills contextually across Claude.ai, Claude Code, and API
Organizational knowledge compounds inside the platform

Best use cases:
Enterprise teams with repeatable, high-judgment workflows
PMs and operators who need consistent agent behavior across a large org Partners building skills that integrate their product into Claude’s ecosystem Individual contributors building personal skill libraries across products

The enterprise management layer - launched alongside the open standard - is where Anthropic’s retention mechanics actually live. Team and Enterprise administrators can now manage which skills are provisioned org-wide and which are enabled by default. That admin layer is not in the open standard. It is a paid plan feature. When an organization’s skill library is administered through Anthropic’s enterprise tooling, migrating to a different AI platform means not just porting the SKILL.md files but rebuilding the entire governance layer, the distribution logic, and the institutional knowledge about what each skill does and when it should fire.

The partner skills program completes the picture. Atlassian, Canva, Cloudflare, Figma, Notion, Ramp, and Sentry all shipped skills within months of the spec going live. Each of these integrations deepens the graph of dependencies between Claude and the tools organizations already use - meaning every Atlassian skill installed is a joint retention mechanism for both Anthropic and Atlassian, where switching away from Claude now also means losing the deeply integrated Jira-to-Claude workflow the team has built around.

Churn risk emerges at the point where the portability promise is tested and found partially true: skills transfer in format but not in governance, not in partner integrations, and not in the accumulated behavior tuning built up across hundreds of real invocations. Over time, teams may feel the files move but the institutional investment doesn’t.

Vercel
the distribution layer as an acquisition funnel

Vercel’s skill ecosystem is optimized for what I’d call developer workflow gravity - the accumulation of deployment habits that make Vercel the natural endpoint for anything built with AI.

“I want my AI agent to deploy my projects without me switching contexts to a terminal or a dashboard. And… I don’t want to manage deployment infrastructure myself every time I spin something up.”

It translates to:
Install Vercel skills to agent environment
Agent handles deployments conversationally
Each deployment creates a claimable Vercel project
Developer’s deployment history and configuration accumulates on Vercel

Best use cases:
Frontend developers using Claude Code or Cursor as their primary environment Individuals and small teams who want deployment handled inside the agent workflow
Developers evaluating infrastructure who want the lowest possible friction to first deploy
Teams building AI-powered applications who want Vercel’s AI SDK as part of the skill stack

The Vercel deploy skill is the most commercially transparent artifact in this entire ecosystem. It creates a preview URL and a “claim” URL in the same output - every time a developer uses the skill, they have a live Vercel deployment to claim. The npm analogy that circulates on X is exactly right: npm is free; npm is also how you end up deploying to Vercel.

Churn risk for Vercel is speed-dependent: every deployment a developer makes through the skill before a native alternative exists is one more project, one more configuration that creates switching inertia. The counter-lever is habit formation, and habits form fast.

Vercel, founded by Guillermo Rauch, is a $3.25B infrastructure company built on one consistent thesis: developer experience is the acquisition strategy, deployment scale is the business. Skills.sh is that thesis applied to the agentic layer.

Six retention mechanics the skill ecosystem reveals

1. Accumulated workflow investment as a structural switching cost.

A team that has refined its sprint-planning skill across six months of real invocations has an asset that is technically portable but practically irreplaceable - the refinement is encoded in the file but the judgment that produced it is not documented anywhere else; porting the files is easy, rebuilding the institutional knowledge is not. The users simply wouldn’t want to leave.

2. The open standard as ecosystem gravity, not altruism.

When the standard wins, the author earns Network Economies proportional to total ecosystem scale, not just their own platform’s share; Anthropic’s decision to open-source SKILL.md and donate MCP to the Linux Foundation is the move that makes the proprietary governance layer worth building on top of.

3. Partner skill integrations as joint retention mechanisms.

Every Atlassian, Figma, or Notion skill installed through Claude creates a bilateral dependency - the user is now retained by both platforms simultaneously, and switching away from Claude requires renegotiating the integration value of every partner skill the team has adopted, not just the Claude workflow itself.

4. The leaderboard as quality signal ownership.

Vercel’s skills.sh generates anonymous telemetry from every skill install and surfaces it as a ranked directory, owning the reference layer developers consult before adopting a skill - the same position npm holds in the JavaScript ecosystem, where download counts became the default proxy for package reliability regardless of whether that proxy is actually valid.

5. Free infrastructure as top-of-funnel for paid deployment.

The Vercel deploy skill produces a claimable deployment as its primary output, converting a productivity behavior (using Claude to ship code) into a commercial behavior (acquiring a Vercel account and accumulating project history) without the developer experiencing the transition as a sales interaction.

6. Governance infrastructure as the non-portable moat.

The admin layer Anthropic built for Team and Enterprise plans - org-wide skill provisioning, default skill settings, centralized management - is the part of the skill ecosystem that the open standard explicitly excludes, and it is precisely the part that creates the deepest switching cost once an organization’s IT and security processes have been built around it.

Synthesis

A library that only runs on one platform is a cage. A library that is technically portable while practically invested is a home. Anthropic is betting that knowledge governed through enterprise tooling and integrated with commercial partners creates switching costs deep enough that Claude becomes the default by assumption, not by force. Vercel is betting that the developer who deploys once and accumulates project history on Vercel is the developer who stays - not because they’re locked in, but because starting over costs more than staying.

Churn emerges when the library stops growing, or when the governance layer fails its implicit promise that skills are safe to depend on.

Retention follows investment depth.

If you found this useful, here’s what else Do Not Churn has covered:

• Is ‘app-of-apps’ becoming the new interface for work and life? - The daily assistant category betting that the switching cost is the product, and what that means for every traditional SaaS tool sitting inside the same workflow.

• Why AI note-taking tools no longer compete on features - When a category splits along user intent, the tools that remain are not the ones with the most capabilities. They are the ones who stay closest to the job they were hired to do.

• Why buying a plane ticket feels harder than it did 30 years ago - A category that added complexity every decade in the name of user control, and produced the worst retention profile of any consumer product category I’ve analyzed.

• You didn’t churn. You graduated. - The PKM category and what happens to retention when the product solves its own problem so completely that the user no longer needs it.

  Leave a comment

In February 2026, - General Partner at , former growth lead at Uber, and one of the most-followed operators in consumer tech - wrote a 700-word post on X describing the AI inbox tool he wanted someone to build.

He said he would pay $150,000 a year for it. He listed eleven specific capabilities, one of which was the line that made the post hit 108k views in 48h:

45+ products surfaced in the replies - Cora, Lindy, Fyxer, Serif, Jace, Poke, Tasklet, Superhuman, Extra, Notion Mail, Inbox Zero, Town, and three dozen more.

That matters right now because the AI productivity category is at a moment when all these products are claiming to solve the same problem, and the answer to which one will retain users has nothing to do with capability. When they all do roughly the same thing, then only one question remains: which user are they solving for? On that axis, the category has already split.

That wasn’t always the case.

In the early 2010s, email management was a feature inside Gmail, and Gmail was a feature inside the browser. The serious productivity work happened in plugins. Boomerang shipped scheduling and reminders. Streak turned the inbox into a CRM. Mailbox raised eight figures, sold to Dropbox, and was shut down. Inbox by Google launched, attracted a fanatical base of power users, and was killed by its own parent company in 2018 for reasons that had nothing to do with the product.

The pattern was always the same: every five years a new entrant tried to displace Gmail, raised a round, got acquired or quietly died. The graveyard is full and the gravestones are nearly identical.

What changed in 2023 was not that anyone solved email. What changed was that the AI exposed a structural mismatch nobody had named — the inbox is a queue, and a meaningful slice of users had always wanted a feed. A queue assumes you will process every item. A feed assumes the system will surface what matters and let the rest scroll by. For two decades the productivity industry had been trying to make the queue faster. The new entrants started asking whether the queue itself was the bug for half their audience.

Market at a glance

The AI productivity tools market is scaling from $8.9B in 2025 to a projected $34.99B by 2034, growing at a ~16% CAGR, driven by generative AI maturity, knowledge-worker tool fragmentation (the average professional now switches contexts seven times before they have started their day), and the willingness-to-pay shift from seats to outcomes. The broader gravity includes calendar intelligence platforms and cross-app agent runtimes, making this a structurally expanding category rather than a feature war inside Gmail. (Grand View Research, 2025)

The willingness-to-pay ceiling moved from $30 a month in 2024 to a comparison set that now includes Cora at $20, Superhuman Mail at $40, and Andrew Chen’s stated $12,500 a month if someone actually built what he wrote down.

Email no longer competes on speed.
It competes on whether the user wants to stay in the inbox, leave the inbox, or never see the inbox again.

These three positions are not variations of the same product. They are answers to three different questions, and the products that win each one will retain users for entirely different behavioral reasons.

Superhuman
the inbox, accelerated

Superhuman is optimized for what I’d call velocity inside the existing metaphor, and it does its job extremely well.

Rahul Vohra is the longest-running operator in this category. Born in Birmingham, Cambridge CS graduate, taught himself to code at eight to make video games, interned at Jagex on RuneScape in 2004 - the gaming pedigree is where his now-famous obsession with how users feel about software comes from. He founded Rapportive in 2010 with two Cambridge classmates, built it into the first Gmail plugin to scale to millions, sold to LinkedIn for $15M in 2012, then started Superhuman in 2014 and has been at it ever since. More on what that sixteen-year arc reveals about the category in the section below.

Primary intent: Inbox throughput.

“I want to get through 200 emails a day without losing the thread on any of them. And… I don’t want to leave the muscle memory I’ve built around triaging my own inbox.”

It translates to:
> open Mail
> scan Auto Labels (response needed / waiting on / meetings / marketing / cold pitches)
> review Auto Drafts that adapt to your voice and tone per recipient
> Auto Archive sweeps the marketing and cold pitches before you see them
> Ask AI handles cross-references across inbox, calendar, and the public web in a single query
> Smart Send ships replies at the optimal time. Founder Rahul Vohra reports hundreds of his own emails are auto-archived daily. Ninety-four percent of weekly active users engage with at least one AI feature.

Ideal Customer Profile:
Founders and operators processing 200+ emails a day
Sales professionals running CRM-integrated workflows
Senior executives whose response time is a load-bearing variable in their role
Power users who have built keyboard-shortcut muscle memory over years

Churn watch: Superhuman’s risk is not in the product - net revenue retention sits above 120%, the highest in the category - it is in the audience size of users who genuinely want a faster inbox versus users who say they want a faster inbox but actually want no inbox.

Cora
the inbox, replaced with a feed

Cora is optimized for what I’d call the inbox as a once-twice-daily ritual, and it answers a question Superhuman explicitly refuses to answer.

Built by Every (Dan Shipper), launched June 2025, $15-20 a month, growing on word-of-mouth.

Cora is one of several products inside Every - built with Kieran Klaassen, Brandon Gell, and Nityesh Agarwal - and its design philosophy comes directly from Shipper's writer-first identity: he treats email as content to be summarized, not a system to be optimized. The product looks like an oil-painting morning briefing because the founder is a writer who happens to ship software, not a software person who happens to write.

Primary intent: Email as background.

“I want my inbox to disappear from my day-to-day attention. And… I don’t want to feel guilty about not opening it for six hours.”

It translates to:
> connect Gmail
> Cora reads your historical email to learn your patterns and writing style
> only humans needing a reply land in your inbox
> Cora pre-drafts replies in your voice for common requests
> everything else is summarized into a Brief delivered twice a day.

Best users:
Operators who would otherwise check email reflexively every fifteen minutes
Founders whose inbox is mostly noise punctuated by a few real signals
Writers and creatives whose work suffers from constant context-switching
Anyone who has tried Inbox Zero and found it made them feel worse, not better

Churn watch: Cora’s risk is the day a user actually needs to act on something in real-time and discovers the Brief came too late. The metric is whether the user starts opening Gmail directly between Briefs. If they do, the ritual has broken. The counter-lever is push notifications for genuine emergencies, but Cora has resisted this - emergency-channel inflation is the fastest way to turn a feed back into a queue.

Lindy
the inbox, dissolved into a work agent

Lindy is optimized for what I’d call email as one channel inside a bigger work graph, and it does not really think of itself as an email product at all.

Flo Crivello is French, joined Uber as a PM in 2015, and has been writing publicly about technology, philosophy, and business for almost a decade. Before Lindy, he founded Teamflow in 2020, raised $50M, built it into a virtual-office product, and watched growth flatline the moment COVID ended. He fired two-thirds of the team and pivoted into AI agents in 2023. The breakthrough was a positioning move he calls “the Notion head fake” - Lindy 1.0 was sold as “AI employee,” which fell flat because the product was too broken for the promise; he repositioned as “Zapier of AI,” which made the platform legible to mainstream users while the product caught up to the original vision. That instinct - to ship something familiar on the way to something radical - is the entire Lindy strategy in one sentence.

Primary intent: Cross-tool action.

“I want an agent that knows everything I know — my inbox, my Slack, my calendar, my CRM, my notes. And… I don’t want to be the integration layer between my tools.”

It translates to:
> install Lindy in iMessage
>connect 3,000+ integrations including Gmail, Outlook, Slack, Notion, HubSpot, Salesforce, Linear
> describe what you want in plain English (”if an email comes from my manager, mark it urgent and ping me on Slack - if it’s a newsletter, archive it”)
> Lindy builds the agent and runs it 24/7
> The agent talks back to you over text. Founded in 2023 by Flo Crivello, $50M+ raised through Series B, 5.5x growth in twelve months.

Best users:
Founders who run their day across iMessage, not across apps
Solo operators who need an agent that handles email, calendar, and CRM in one breath
Engineering leaders who already think in workflows and triggers
Anyone whose actual job is happening in five tools, not one inbox

Churn watch: Lindy’s risk is the onboarding cost. The agent has to ingest enough of your work context before it becomes useful, and most users churn in week two - when context is still being built and the value has not compounded. The counter-lever is a templated agents that deliver value before personalization kicks in (shipped recently).

Six retention mechanics for anyone building in this category

If you’re shipping into the AI inbox space, these are the six mechanics the data suggests will decide who retains and who churns. Treat them as advice, not laws - the category is moving fast enough that the next twelve months will rewrite some of them.

1. Train for voice fidelity before you ship anything else.

(Superhuman’s Auto Drafts adapting tone per recipient; Serif’s 64% sent-without-edits rate.) Ingest at least 90 days of historical email before letting users see a draft. Below 50% sent-without-edits, you have a writing tool. Above 70%, you have a switching cost that compounds every week the user stays.

2. Pick one daily ritual and own the time slot.

(Cora’s twice-daily Brief.) If your product is in the inbox-replacement camp, the brief has to arrive at the same time every day without exception. Users who read at 7:42 a.m. for six weeks will not switch to a competitor delivering at 8:15 - but they will churn the day your brief is late twice in one week. Set the SLA tighter than your engineering team is comfortable with.

3. Make context depth feel useful in week one or you will not get to week three.

(Lindy’s templated agents; Tasklet’s starter recipes.) The biggest killer in the cross-tool agent camp is the gap between “the agent has ingested everything” and “the agent does anything useful.” Ship templated workflows that deliver value before personalization compounds. If less than 40% of week-one signups reach week three, your onboarding is the bug.

4. Let users escalate autonomy at their own pace.

(Superhuman’s Auto-Send opt-in per category; Cora’s resistance to push notifications.) Andrew Chen named this himself - “for now, don’t send… in the future maybe a YOLO option.” Users want to grant the AI more authority over time, not all at once. Ship every action behind a per-category toggle: auto-archive marketing first, auto-reply to scheduling second, auto-send to inner-circle last. Products that auto-everything on day one churn the user the first time the AI sends something they wouldn’t have.

5. Treat the metaphor as a product decision, not a default.

(Superhuman keeps the inbox; Cora replaces it with a feed.) The biggest unforced error in this category is shipping without picking. If your product is “AI for Gmail” without a point of view on whether the user should still see Gmail, retention collapses the moment a competitor with a sharper position takes the slot. Pick the metaphor. Defend it on the landing page. Build the rest of the roadmap downstream.

6. Build for the user behaviour you have, not the one your roadmap was written against.

The 2023 user opened Gmail twelve times a day and triaged manually. The 2026 user opens Gmail once and edits AI drafts. Run a behavioural audit every quarter - do your retained users actually use the product the way your personas describe? When the answer is no, the roadmap is the problem, not the product.

***

Key takeaways

• Email as a category no longer competes on features or speed. It competes on which user psychology is being catered to.

• Inbox-acceleration tools (Superhuman) retain through voice fidelity and muscle memory built across years.

• Inbox-replacement tools (Cora) retain through ritual capture — the daily Brief becomes the new behavioral anchor.

• Cross-tool agents (Lindy) retain through context depth that compounds with usage and is impossible to rebuild after switching.

• Retention follows user psychology alignment; churn begins when the product solves a different relationship with email than the one the user signed up to have.

I have spent ten years inside the rooms where churn gets engineered on the UI and product level.

Sitting next to the heads of product who green-lit five-step cancellation flows because the data said each step saved seven figures a year, sitting through QBRs where a 0.3% lift in conversion from a darker shade of red on the upgrade button was treated like a strategic victory, listening to retention specialists explain why the default-on auto-renewal toggle was actually pro-consumer if you squinted at it the right way.

The entire D2C and SaaS economy of the last 15 years rests on decisions like these. Most of them are not malicious, of course, and made by smart, well-intentioned people inside companies whose growth model requires that users not be too easy to lose. There are many examples out there on websites like Deceptive.Design and Confirmshaming - you’ll actually be surprised to see companies whose services you use every day.

But something has been shifting underneath this whole structure for a while now as users stopped tolerating dark patterns. They started recognizing the cancellation maze, the default-on toggle, the artificial scarcity timer, the "are you sure you want to leave us?" spiel, and they learned to screenshot it and post it. They learned that the friction was a tell - a sign that the product knew, on some level, that it could not retain them on value alone. I wrote about this in The Retention Wall Is Dead issue: the entire architecture of friction-based retention has been losing legitimacy after the loud FTC case that happened with Amazon.

And underneath that legitimacy crisis, a second thing has been happening - a thing that is harder to see from inside a company because it does not show up in dashboards. User expectations have grown faster than any single interface can keep up with.

The people using your product in 2026 are using ten other products in parallel, all of them iterating weekly, all of them reshaping what “good” feels like.

> The support agent using Zendesk is also using a custom Notion dashboard their team configured in two hours.
> The marketer using HubSpot is also asking ChatGPT to draft their entire campaign brief in chat.

Every one of those parallel experiences resets the user’s sense of what their interface should feel like - and the gap between what their primary SaaS can show them and what their workflow actually requires has gotten wider every quarter for five years. No static UI is capable of reflecting 100% of how a real user actually works. The shape of work is now too specific, too cross-product, too personal. The interface that the vendor designed three years ago for an idealized user who does not exist is not going to catch up.

Both of these pressures - the legitimacy collapse of friction-based retention, and the impossibility of designing one screen that fits a real user’s actual workflow - point to the same conclusion: something has to replace the static, vendor-controlled interface. The question that has been live in product circles for the last two years is what.

There have been two main guesses, and they are very different from each other. The first guess is that the interface gets replaced by voice. This is the death-of-the-UI thesis - that the next paradigm of computing is intent-first, that users describe what they want and a voice agent does it, and that the screen becomes a vestigial layer or disappears entirely.

Voice Control?

It is an aesthetically clean idea. It is also, on closer inspection, almost certainly wrong as a primary modality. Voice input is fast, but voice output is slow - the user has to listen to the response in real time, in serial, with no ability to scan or skim. A 200-word answer takes 90 seconds to listen to and 12 seconds to read. Multiply that across every interaction in a workday, and voice becomes punishing for any task more complex than setting a timer. Voice will absolutely have a role for ambient computing, hands-free contexts, and accessibility - but it is not going to be the primary way most knowledge workers interact with their software, because nobody has the patience.

Picture courtesy of

Jakob Nielsen on UX

No More User Interface?

Summary: AI products have changed from invisible enhancement of classical user interfaces to soon become the main avenue for users to engage with digital features and content. This may mean the end of UI design in the traditional sense, refocusing designers’ work on orchestrating the experience at a deeper level…

Read more

a year ago · 58 likes · 2 comments · Jakob Nielsen

The second guess is that the interface becomes fluid - composed on the fly, per user, per query, per session, by an agent that knows what the user is trying to do and what the underlying software can do.

Subscribe now

Dynamic interface thesis

Two signals that are currently setting the fashion here:

Y Combinator put Dynamic Software Interfaces on its Spring 2026 Request for Startups and named it as a category.

MIT Technology Review put generative coding on its 10 Breakthrough Technologies list for 2026, which is the supply-side enabler - the moment users (or their agents) can compose software faster than vendors can ship it, the lopsidedness of who-controls-the-screen has to break.

The institutional consensus formed quickly, faster than any tech-trend consensus I have seen since mobile-first. Within six months of the YC RFS, every major SaaS vendor had made some public gesture toward primitives, agent readiness, or MCP integration.

So far, so unanimous. And here is where I have to stop nodding along, because almost everybody discussing this shift is discussing only half of what it actually is.

The story being told right now about dynamic interfaces is about freedom - users finally getting software that fits their workflow instead of bending around it. That story is half the picture. The other half is that an entire generation of retention machinery is about to stop firing, and most of the people who built that machinery are pretending not to notice.

I have been thinking about this shift for two years, since the moment it became obvious that the speed at which someone could vibe-code a working app had collapsed below the speed at which an incumbent SaaS team could run an A/B test on a paywall variant.

Dynamic interfaces don’t just change how products look. They expose which products were retention-by-design and which were retention-by-manipulation.

Two cycles are running at the same time, pulling in opposite directions. They describe the same shift, but they produce opposite consequences.

The Authorship Cycle is what everyone talks about. AI gets good enough to compose interfaces from primitives, vendors ship those primitives instead of finished screens, users (or their agents) reshape the product into the form they actually need, and the long tail of niche workflows finally gets served.

The Trap Cycle runs underneath the same shift in the opposite direction. When the interface composes itself per session, the upgrade banner doesn’t render. The cancellation flow has nowhere to put its five “are you sure?” steps. The pre-checked subscription box never appears in the form because the agent didn’t think to ask for it. The cross-sell sidebar doesn’t exist because the user didn’t request a sidebar. The streak counter, the unread badge, the activity feed driven by in-app state, the “you have 3 unread items” notification, the weekly recap email triggered by engagement metrics that only mean something inside the vendor’s UI - all of these are interface manipulations that depend on the vendor controlling the screen, and they stop firing the moment the vendor stops controlling it.

The thing that has been quietly disturbing to watch is how few of the incumbent SaaS companies are willing to talk about this honestly. Look at the launch partners listed in the Anthropic MCP Apps press release - Salesforce, Asana, Slack, Box, Canva, Figma, Hex, monday.com. Every one of them signed up. Not a single one has publicly addressed what it actually means for their growth model when a user’s agent composes the interface inside Claude rather than inside their app. That silence is not a coincidence. Signing up was defensive - the alternative, being absent from the launch slide while every competitor was on it, would have been worse. But every interface surface they let Claude render is a surface where their own retention machinery does not fire. They are renting out their data to a chat window with the UI traps removed. The internal numbers will eventually make this visible. The public conversation has not caught up yet.

The dynamic interface era is not a UX trend. It is an audit of which SaaS businesses had real product value and which ones were getting away with it.

Here is the test I keep running in my head, on every product I have observed up close, every product I see in the market, every product I am asked to evaluate. If the user’s agent rendered your product into the exact view that user needed - no upsell modal, no cancellation friction, no cross-sell, no streak counter, no notification dot, no default-on toggle, no “are you sure you want to leave?” - would your retention numbers hold?

For most products I have watched up close, the honest answer is no. The companies were not retaining users. They were retaining the conditions under which users couldn’t leave. The dynamic interface era takes those conditions away.

It means a reckoning is coming, and most of the companies it is coming for cannot see it from where they are standing.

THESYS

Thesys was started by Rabi Shanker Guha and Parikshit Deshmukh - two engineers who came out of Google, Stripe, and Salesforce having watched dozens of enterprise AI deployments stall on the same boring problem. The model worked, the data was clean, the agent did what it was supposed to do, but every customer wanted a different interface and nobody could afford to build them all by hand. Guha and Deshmukh quit and built the missing layer themselves. Today Thesys sells an API called C1 that turns LLM responses into live, streaming, interactive UI components - forms, charts, dashboards, tables — composed in real time inside someone else's product. The customer is a SaaS company building an AI agent that needs a different interface for every possible query state. C1 is the layer that makes those interfaces appear without anyone hand-coding them. The whole company is optimized for what I'd call infrastructure-grade UI generation.

Primary intent: Vendor adaptation.

Job it does:

“My AI agent has infinite output states. I cannot pre-design a screen for every query my users will ask. I want a layer that generates UI from the agent’s response, in production, matching my brand.”

It translates to:
Developer integrates C1 in two lines of code
User queries the AI agent inside the SaaS product
C1 streams structured UI components into the React frontend per query
User interacts; actions feed back to the agent
Next query produces an entirely new composition.

Best users:

Engineering teams building AI copilots inside existing SaaS
AI-native startups whose product is the chat interface
Internal tools teams replacing six static dashboards with one dynamic surface
Vertical AI agents in legal, medical, or financial domains

Churn watch: Thesys becomes infrastructure debt the moment a customer ships it to production, which is its strength and its trap. The strength is real, though - ripping it out means rebuilding the entire UI layer, which nobody wants to do. The trap is that the same primitive is being standardized at the model-provider level, for free, by Thesys’ biggest suppliers. Anthropic shipped MCP Apps in January with the same capability. OpenAI shipped Apps SDK in October. Vercel open-sourced json-render in January and it accumulated 13,000 GitHub stars in three months. That conversation is coming inside twelve months, and Thesys’ best defense is to become so design-system-aware, so brand-protective, so embedded in customer workflows that it earns its keep on top of the commoditized layer beneath it.

FLIPBOOK

Flipbook was built by Zain Shah - a former OpenAI researcher who’d worked on AI and robotics, then spent time at Samsung as a Creative Technology Expert - together with Eddie Jiao and Drew Carr. Shah is a YC S13 alum, which means he has been around the cycle once before and chose to come back in 2026 with a thesis: that the next paradigm of computing is not better software but a different rendering layer entirely, one where pixels come straight from a model rather than from a layout engine interpreting HTML. They launched Flipbook in April 2026, and the founder's thread went viral within forty-eight hours.

The product they shipped is optimized for what I’d call interface-as-medium, and it does its job in the most maximalist possible way. There is no HTML. There is no layout engine. There is no DOM. Every page is a full-screen image generated live by a model, streamed via WebSocket from a serverless GPU running an optimized video diffusion model. You click anywhere on the image, and a new image generates that dives deeper into that part of the previous one. The page is the medium, and this medium is alive. There is nothing else.

Primary intent: Visual exploration.

Job it does:

“I’m curious about something and I do not want to read a wall of text or click through ten Wikipedia tabs. I want the medium itself to feel shaped to my curiosity.”

It translates to:
User types a topic or uploads an image
Flipbook generates a full-page illustration in real time
User clicks any region of the image
A new image generates that goes deeper into that region
Navigation history saves as a tree of generated nodes
Exploration continues until the user closes the tab.

Best users:
Visual learners who hate reading
Designers and product thinkers looking at the future of the web
Curious consumers who fall down rabbit holes for fun
Educators experimenting with media as pedagogy

Churn watch: Flipbook has no functional job-to-be-done. The novelty wears off in three sessions. It cannot replace search (slower, less accurate, no citations). It cannot replace Wikipedia (you cannot quote a generated image, you cannot trust it on facts). It cannot replace YouTube (the video is hallucinated, not real). The team has said publicly that the product is limited by design and that the set of useful applications will expand as the models get more accurate and stateful. That is a confession, not a roadmap - it concedes that today’s retention depends on the user accepting a demo, not a product. Flipbook is the most intellectually honest thing in the entire dynamic interface conversation, because it commits to the maximalist version of the thesis (the page itself should die) without pretending to have figured out what people will actually do with it.

[VACANT]

This is the slot that does not have a product yet. It is the slot Y Combinator’s RFS is pointing at. It is the slot Rhys Sullivan’s prototype gestures toward, the slot Peter Steinberger’s openclaw canvas hints at on a Friday night demo. And it is the most valuable empty seat in software in 2026.

The product that fills this slot does one thing: it lets a user’s agent reshape an existing SaaS into the form that user actually needs, while the vendor keeps the data, the integrations, the auth, and the billing. The user does not build a new app. The user modifies someone else’s app in place. Salesforce, but only the seven fields the salesperson actually uses. Notion, but with the structure your team’s coding agent thinks makes sense for your workflow. Zendesk, but with the ticket interface generated per support agent based on their case mix and tenure. Linear, but with the project view restructured around how your specific team thinks about scope.

Primary intent: Substrate extension.

Job it would do:

“I use this SaaS every day. Sixty percent of the interface is irrelevant to my workflow, and I have spent two years building elaborate workarounds to ignore the parts I don’t need. I want my agent to rebuild the parts I actually use, in the shape I actually need, while keeping everything underneath that I depend on.”

Six retention mechanics in a world where the vendor doesn’t pick the screen

If you are a PM, founder, or Marketer watching this shift unfold, here is what to actually do about it. Not next year, not when your CRO finally takes the threat seriously, but now.

1. Invest in the substrate, not the surface.

Stop treating your UI as the moat. The moat is your data model, your integrations, your permission system, your API speed, the accuracy of your search index - the parts of your product that survive even if a user’s agent renders the entire interface from scratch. Reallocate engineering hours away from polish on screens that may not exist in two years and toward making the underlying capability genuinely best-in-class. The question to ask in every roadmap review is no longer “will this look good in the demo?” but “would a user’s agent still pull from us if it could pull from anyone?” If the answer to the second question is no, your roadmap is wrong.

2. Run the trap audit before the market runs it for you.

Take every UI surface in your product right now and sort it into two columns: surfaces that serve the user, and surfaces that serve the business at the user’s expense. Be honest. The cancellation flow with five “are you sure?” steps goes in column two. The pre-checked email subscription box goes in column two. The “upgrade to Pro” banner that fires on the seventh page view goes in column two. Then look at how much of your retention math depends on the column-two surfaces. That number is your exposure. Every dollar of retention that comes from column two is a dollar that disappears the moment the user’s agent starts composing the screen. Plan for the day that revenue line goes to zero, because it is going to.

3. Rebuild your cancellation flow before regulators or your users force you to.

The legal cover for dark-pattern cancellation is collapsing - the FTC went after Amazon Prime, the EU passed the Digital Services Act, and a generative interface that lets users (or their agents) skip your friction layer entirely makes your “are you sure?” flow look indefensible. Get ahead of it. Make cancellation a one-click action. Replace the friction with a genuine why are you leaving survey that you actually use to fix the product. You will lose some short-term saves. You will gain something more durable: a churn number that reflects reality, which is the only kind of churn number you can actually fix.

4. Redesign engagement to be requested, not injected.

Streak counters, notification dots, unread badges, activity feeds, in-app state-driven recap emails — these survive only as long as you control the screen, and you are about to stop controlling the screen. Start building engagement loops the user can opt into and configure. Let them choose which signals they want pushed to them and where. Build the version of your product that still works when the user has filtered out everything you wish they wouldn’t filter out. The Duolingo-Snapchat-Slack engagement model is on a clock; the products that move first to user-requested engagement will define what the new model looks like.

5. Treat cross-sell and discovery as products, not placements.

Stop relying on the sidebar, the “customers also bought” rail, the homepage promo slot. They will not exist in two years for the segment of your users whose agents compose the interface. Build discovery as a feature the user actively wants to use - a recommendation surface they open intentionally, an exploration mode they trigger, a question they ask their agent that pulls from your catalog. The companies that make discovery feel useful rather than inserted will extract real expansion revenue from the new era. The ones still relying on placement will watch attach rates fall by half with no clear path to rebuild what they had.

6. Make your substrate carry the brand the screen used to carry.

When the user’s agent decides what the interface looks like, your brand stops being carried by the visual layer. It gets carried by the data quality, the trustworthiness of the API, the speed of the response, the ethics of how you handle customer information, the way support handles a tough ticket. Audit each of these the way your design team audits visual consistency - because they are now doing the work the visual layer used to do. The substrate is the brand now. Make sure yours is one a user’s agent would choose.

Today, you can open Claude, type “prototype a pricing page for a Python automation SaaS, three tiers, dark mode,” and, within a minute, have a clickable, on-brand prototype you can hand to Claude Code. Without Figma, a product manager and a designer, with no back-and-forth with anyone who lives in a Bay Area zip code. One release has changed the order of how things get from the idea to the product.

I’d been planning this article for months and kept setting it aside. Every design tool I looked at was solving the wrong endpoint - idea to prototype, when the job was always idea to product. Claude Design is the first release that picked the right finish line.

In the early 2020s,

shipping a product meant a chain of specialists: PM writes the spec, designer opens Figma, and engineer translates the file into code. It took three weeks if everyone stayed aligned.

Then the chain started collapsing. Vibe coding took the engineering step first - Lovable, Bolt, v0, Claude Code turned “describe an app” into “working app.” By early 2026, 25% of Y Combinator startups were shipping products where 95% of the codebase was written by AI.

But the chain had a stubborn middle. Somebody still had to design the thing. That judgment call - taste - was the last step outsourced to humans. Figma owned the tool those humans lived in.

On April 17, 2026, Anthropic shipped Claude Design. The middle link folded in too.

The moat wasn’t beaten on features. The workflow stopped needing the step Figma owned.

Market at a glance: the global design, editing & rendering software market was $66.95B in 2024, grew to $70.59B in 2025, and is projected to reach $87.84B by 2029 at a 5.6% CAGR (Research and Markets, March 2025).

Figma held an estimated 80–90% share of UI/UX design tooling going into 2026. Since its summer 2025 IPO, FIG stock is down more than 80% from its post-IPO high.

Claude Design launched to every paying Claude subscriber. Anthropic's traffic is up roughly 5x year-over-year, valued at $380B as of February. Eight of the Fortune 10 are customers.

Claude Design is optimized for what I’d call outcome delivery for the non-designer. Launched April 17, 2026, powered by Claude Opus 4.7. The product everyone is talking about because Figma’s stock moved.

Primary intent: Shipped prototype without a designer in the loop

Job it does:

“I want to turn my idea into a branded, interactive, production-ready prototype - without opening a design tool, finding a designer, or knowing any of standard design practices.”

It translates to:
Point Claude at codebase, Figma file or screenshots of your website/app
Claude extracts design system
Prompt for screen or flow
Iterate via chat, inline comments, custom sliders
Hand off to Claude Code

Best users:
Solo founders who have never opened Figma
PMs who want to show, not describe, what they mean
Marketers producing on-brand assets without a design queue
ccount executives building pitch decks from a rough outline

Churn watch:
Token economics - early users on the $200 Max plan report rate-limiting after 20 iterations on a single project. If the cost-per-design exceeds hiring a junior designer, the product becomes a luxury most non-designers can’t afford.

Skill-graduation churn - Claude Design is built for the user who doesn’t know what good looks like. The moment a user develops taste, they notice the token burn and start looking for the cheaper loop. I’ll come back to this.

Figma is optimized for what I’d call multiplayer design as the source of truth. Dylan Field and Evan Wallace met at Brown in 2011, where Wallace was Field’s TA. Field took a Thiel Fellowship, dropped out, and the two spent four years building something heretical at the time: a full design tool that lived in the browser and let people edit together in real time. They didn’t launch a beta until 2015, didn’t charge a dollar until 2017. By the time Adobe tried to buy them for $20B in 2022, Figma was the design tool - so dominant the deal was blocked on competition grounds. It IPO’d in August 2025.

Primary intent: Collaborative design as the organizational source of truth

Job it does:

“I want my team to design, critique, and iterate together in one place - and I don’t want to lose what we built when we hand it to engineering.”

It translates to:
Build components
Edit with teammates in real time
Comment and iterate
Hand off via Dev Mode

Best users:
Product design teams of 5–50 at venture-backed companies
Design system owners maintaining component libraries
PMs who annotate designs alongside their designers
Agencies shipping client work through shared files

Churn watch: Figma’s churn risk is almost entirely external. The product hasn’t gotten worse - the workflow around it has gotten shorter. When a PM generates a good-enough prototype in a conversation with Claude, hands it to Claude Code, and ships it, Figma’s step becomes optional. Its moat (multiplayer, components, libraries) defends against competitors doing the same thing better. It was never designed to defend against the chain getting one step shorter.

Anthropic’s CPO Mike Krieger stepping down from Figma’s board three days before the launch wasn’t a coincidence; it was a market signal everyone read at the same time.

Google Stitch is optimized for what I'd call owning the format, not the canvas. . Galileo AI was founded in 2022 by Arnaud Benard and Helen Zhou - a two-person startup that hit 100,000 design generations in days during its public beta and raised $15M on the premise that designers wanted production-grade UI from a prompt. Google acquired them in May 2025, folded the product into Google Labs, and relaunched it as Stitch running on Gemini. For most of the last year, Stitch looked like the weakest of the three players in this space - a sketchpad people opened for ten minutes, got bored of, and exported to Figma. Then, four days after Claude Design launched, Google made a move nobody expected.

Primary intent: Defining the portable file format for design context

Job it does:

“I want my design system to live in a file that any agent, any IDE, any design tool can read - and I want to stop being locked inside one vendor’s canvas.”

It translates to:
Define your design system in DESIGN.md
Validate it with the open-source CLI
Use the same file in Stitch, Claude Code, Cursor, Paper, Figma
Update once, propagates everywhere

Best users:
Design system owners who want portability across tools
Agencies that move design context between client stacks
Developers tired of re-describing the brand to every new AI tool
Teams hedging against AI-tooling lock-in

On April 21, Stitch open-sourced a formal specification for DESIGN.md - the text-file format that describes a design system to agents. The spec includes a token section (every design decision expressed both as prose reasoning and as a machine-readable value), a components section with role-based references instead of hex codes, and a CLI validator that agents can run on their own output to catch WCAG violations before shipping. Published on GitHub, accepting pull requests.

The community reaction was immediate and telling.

Stitch stopped trying to win the design-tool product fight. It moved up the stack and defined the substrate everyone else has to read from. If every IDE, every coding agent, every design canvas learns to parse the same DESIGN.md format, Stitch becomes infrastructure. Claude Design becomes a client of that infrastructure.

File formats outlast their creators - Adobe learned this with PSD, Autodesk with DWG. Google just tried to claim the same position for AI-era design context.

The six retention mechanics of the vibe coding chain

1. Workflow ownership, not step quality.

Claude Design doesn’t win on design quality - by most serious reviews, Figma still produces better output at the step level. It wins because it owns more of the chain. When the job is “ship the thing,” the tool that takes the user further down the chain wins, even with worse step-level output. Stitch owned the shortest span. Figma’s is longer, but bounded.

2. The design system as context, not artifact.

For a decade, the design system was a human-maintained artifact living in Figma. Claude Design reads your codebase and infers it automatically. Taste used to be a skill. Now it’s a file. Once taste is machine-readable context, it’s an input to a model, not a specialist you hire.

3. Cold-start-friendly onboarding for the non-designer.

Claude Design is built for the user with nothing - no codebase, no Figma file, no vocabulary. It generates something generic-but-coherent on the first try, and generic-and-shipped beats nothing-at-all. Figma was built for the user who was already a designer. Canva built a ~$40B business serving the opposite user. Claude Design inherits that user with a better handoff.

4. The real product isn’t the canvas - it’s the handoff.

Claude Design to Claude Code is a single click. The designer-to-developer handoff has been the most expensive coordination cost in software production for twenty years. Anthropic didn’t improve it; they collapsed it.
Figma isn’t competing with a better design tool. It’s competing with a pipeline: Opus 4.7 → Claude Design → Claude Code → Cowork. When your competitor owns the model, the design surface, and the code surface, no single feature fight matters.

5. Open formats as anti-churn insurance.

The DESIGN.md spec is a different kind of retention mechanic - it retains users to the format rather than to any single tool. If your design system lives in a portable file that every agent can read, you’re insulated against any one vendor rate-limiting you, raising prices, or shutting down. Stitch wins users not by keeping them inside Stitch but by making sure the next tool they use still speaks its language. This is how infrastructure beats product in categories that move fast: you don’t compete for the user’s session, you compete for the user’s artifact.

6. The pricing lever disguised as a feature.

The launch-week complaint wasn’t “Claude Design is bad” - it was “Claude Design repackages capabilities we already had for 20x the token cost.” That’s a distinct churn pattern. When a platform bundles cheap primitives (MCP + Claude Code + a third-party canvas) into a premium wrapper, power users notice the arbitrage instantly and post about it publicly. The product looks less like a new feature and more like a monetization surface. Trust with the most visible segment of the userbase erodes first - and those are the users whose tweets set the narrative for everyone else.

The power-user twist

Designer-developers already running Claude Code loops don’t need Claude Design. They have Paper - a canvas built on HTML and CSS instead of a proprietary vector format - connected to Claude Code via MCP. Because Paper’s canvas is the DOM, LLMs read it natively. The agent manipulates it in real time while the designer watches.

@kloss_xyz, posted the day after launch: Opus 4.7 + Paper MCP + five years of design files generated hundreds of on-brand assets at 3% of her weekly Max quota. Same model, Claude Design, rate-limited after 20 iterations on one project. The read is: “All Anthropic is doing here is throttling our usage with a new Design UI and calling it a feature.”

It’s right and wrong in the same sentence. Right, the power user had a cheaper path first. Wrong that Claude Design is aimed at people like her. Anthropic built it for the user who doesn’t have five years of design files, who’s never heard of MCP, who can’t set up Paper from scratch.

For that user, the inference of a design system from an empty codebase is the product. The power-user complaint is the sound of a tool being built for someone else.

On Lenny’s Podcast this month, Nikhyl Singhal - former CPO at Meta, Google, Credit Karma - said what a lot of product leaders were already thinking:

“There’s a split in the PM market where the information mover archetype is going to go obsolete. But the folks who like to build stuff are going to be wanted. Today, everyone wants a builder.”

His prediction: companies will shed 30,000 PMs and rehire 8,000 AI-first builders. Triple the wages, a fraction of the headcount.

The information mover - the PM who coordinates between engineering, design, and leadership without direct authority - becomes a liability in the same chain collapse that killed the designer’s gatekeeping role.

Claude Design is what the builder’s tool looks like. It isn’t a design app. It’s a way for one person with a thesis to ship a product without assembling a team. Figma’s decline isn’t only Figma’s story, but every middle-level specialist's story.

Design software didn’t converge into a single super-tool but split along lines of who owns the chain.

Some tools own the multiplayer canvas - Figma’s bet, still real, still shrinking. Others own the chain from idea to shipped code — Anthropic’s bet, locked to their surface. Others own the portable file format everyone else has to read from - Google’s bet, placed four days after Claude Design launched.

Churn emerges when a product’s moat defends the wrong layer of the stack. Figma defended its step and watched the workflow collapse around it. Anthropic defended its pipeline and watched the power users defect to the open-MCP path. Google stopped defending a product altogether and tried to claim the substrate. In categories that collapse into AI-native pipelines, retention grows when the product owns something that survives the next generation of tools. A canvas remains valuable only if the workflow still needs a canvas. A pipeline remains valuable only if users stay inside it. A file format survives both.

Key takeaways

• Products that deliver a step churn to products that deliver an outcome. Figma delivers the design step. Claude Design delivers the shipped prototype. When “ship the thing” is the job, chain length beats step quality.

• The non-designer was always the larger user. Canva proved it. Claude Design inherits that user with a better handoff to production code.

• Taste became context — and now context has a spec. Google open-sourcing DESIGN.md turned a tacit convention into a portable, validatable file format. The file survives the tool.

• The power user has the cheaper path first, and it doesn’t save the category. Paper + MCP is architecturally cleaner than Claude Design. It also doesn’t reach the non-designer. Both things are true.

• The Problem Cycle beat the Opportunity Cycle in the same week. Anthropic saw a market opportunity and shipped a product. Google saw a behavior already happening in the community and formalized a standard. Within 96 hours, the market named the difference.

  Thanks for reading Do Not Churn! This post is public so feel free to share it.

  Share

If you’ve spent the last few months building a personal knowledge base, you probably know this feeling. The first week is extraordinary. You drop a few documents in, you ask a question, the AI synthesizes an answer that would have taken you an hour to produce manually. You restructure your whole capture workflow around it. It feels like you’ve finally solved the information problem.

Then six weeks in: your archive grows, your use cases multiply. And somewhere around the third month, the thing you built for yourself stops scaling with you - turning into a bloated mess over time - and you’re not sure if the problem is the tool or the system or just you. Hard to imagine we ever built personal knowledge management systems that felt this good to start, and this frustrating to maintain.

That wasn’t always the case.

In the 2000s, personal knowledge management wasn’t a product category. It was a practice - and a fairly elite one. The people who built serious personal archives were academics, researchers, and a small group of methodological obsessives who had discovered Niklas Luhmann’s Zettelkasten method and committed their careers to replicating it. Everyone else kept bookmarks, email folders, and the occasional Word document labeled “notes - important.” Memory was personal and fragmented, and nobody expected technology to fix it.

The first wave of mass-market change came with Evernote. Capture became easy. The problem, quickly obvious, was that easy capture creates a different crisis: retrieval. By 2015, the average Evernote user had thousands of notes they couldn’t find. Evernote had solved the input problem and created the organization problem in its place.

Notion and Obsidian arrived with better answers to organization, and built devoted communities around the idea that a well-structured personal archive was worth investing in. The PKM (personal knowledge management) community that grew up around these tools in the early 2020s was genuinely influential - YouTube channels dedicated to note-taking systems attracted millions of views. The practice became a philosophy. The philosophy became a market.

Then AI note-taking arrived and seemed, briefly, to solve the problem from a different angle. Tools like Granola and Coconote stopped asking users to organize at all - they attended your meetings, transcribed your conversations, and surfaced structured summaries automatically. If you haven’t read the piece on why AI note-taking tools no longer compete on features, the short version is this: those tools solved the capture problem elegantly, but treated knowledge as infrastructure rather than something that compounds. Once the meeting ended, the value was mostly extracted.

And yet the failure mode persisted. The people who maintained elaborate systems were the people who had time and inclination to do so - a narrow subset of knowledge workers. For everyone else, the system became a second job at exactly the moment their first job got busier.

The real shift is happening now: when Andrej Karpathy posted a GitHub gist in April 2026 that described what he called “LLM Knowledge Bases” - a pattern where the AI doesn’t just answer questions from your archive, it compiles and maintains the archive itself. The post reached 16 million views in 48 hours. He basically named the actual failure mode everyone had been living with: knowledge systems fail not at setup, but at scale.

In reality, that failure reflects something more structural:

The churn event in personal knowledge tools is not abandonment. It is the moment the tool stops being able to grow with the problem it was hired to solve.

Personal knowledge base tools replaced bookmarks, then replaced folders, and now they’re replacing the act of organizing knowledge at all.

Cabinet, Mem, and NotebookLM are not variations of the same product. They are answers to different theories about where in the scaling curve a user actually needs help - and each one reveals something distinct about how churn operates when the problem grows faster than the tool designed to solve it.

Cabinet is optimized for what I’d call autonomous compounding - a system where the knowledge base writes itself, continuously, without the user’s active participation.

Hila Shmuel is a former Engineering Manager at Apple who left to build Cabinet in public, with the open-source community. The thesis was personal: every time she started a new Claude session, it forgot everything - her project context, her decisions, her research. Cabinet was her answer to that. Launched in 2026, it’s free, self-hosted, and built on the belief that your AI context should live on your machine - not in a vendor’s cloud. The tagline says it plainly: one knowledge base, AI agents that remember everything.

Primary intent: Persistent operational memory

Job it does:

“I need my AI to know what I know, remember what I’ve decided, and keep building on it while I’m doing other things. And… I don’t want to re-explain my context every time I open a new session.”

It translates to:
Install locally
Onboarding wizard builds agent team
Agents run on cron schedules
Agents write structured entries to the KB
Context compounds across all sessions
User queries a KB that grows without prompting

Best for:
Technical founders running lean
Solo operators managing multiple products
Builders who want agents as coworkers, not assistants
Anyone whose AI sessions currently feel like amnesia

Cabinet nailed that the problem isn’t accessing a knowledge base - it’s maintaining one. Every other tool in this category implicitly requires the user to do something: clip this, tag that, organize the other. Cabinet’s design thesis is that a system you have to tend is a system you will eventually abandon. The agents do the tending. You do the steering.

Cabinet intentionally ignores mobile experience, non-technical users, and cloud-first deployment. It treats the knowledge base as infrastructure that agents read and write to, not an interface that humans browse. The human’s relationship to the KB is interrogative, not editorial.

The strongest thing Cabinet does is eliminate the maintenance bottleneck entirely. In every other PKM system, the failure mode is predictable: users set up the system in a burst of enthusiasm, use it well for three to six weeks, get busy, stop maintaining it, and return six months later to an archive that feels stale and untrustworthy. Cabinet’s automated agents break this cycle structurally — the KB continues updating regardless of whether the user is paying attention. This is not a feature. It is a different theory of what a knowledge system is.

The compounding mechanism is also real in a way that most tools only promise. A Reddit scout running every six hours doesn’t just collect links — it deposits structured, synthesized intelligence into a KB that cross-links with everything already in it. At month six, the KB contains institutional memory that no single document could hold.

Churn Watch: When agent output quality degrades, or agents start producing noise rather than signal, the trust in the KB erodes. Unlike manual systems, where the user controls what goes in, Cabinet’s automated writes mean garbage-in-garbage-out is invisible until the KB feels unreliable. Cabinet needs a quality signal - some way for users to see whether agent outputs are being used, or quietly accumulating unread.

When retention depends on automated compounding, the winning strategy is not more automation - it’s auditable automation.

Subscribe now

About this newsletter:
I’m Daria Littlefield. I spent a decade running customer operations across 35+ dating and social apps - the kind people delete in a rage and reinstall a week later. That taught me more about why products lose users than any framework ever could. Do Not Churn is where I put that to use. Every week, I break down the retention mechanics behind the products we all use - what keeps people, what loses them, and why the difference is rarely what you’d expect. If that’s useful to you, subscribe below.
If you’ve ever watched your metrics show healthy retention while your best users stopped doing the thing they originally signed up for - share it with me in comments.

Leave a comment

Mem is a zero-folder knowledge management product built for people who capture more information than they can organize. Mem nailed that the output of note-taking is not notes, but accessible, connected knowledge - and that the act of organizing is the thing most people will not do.

Kevin Moody and Dennis Xu met as college freshmen in San Francisco in 2014, eating Korean fried chicken and talking about why digital storage systems seemed to do the opposite of human memory - information captured into them fragmented instead of compounding. Seven years later, they launched Mem with backing from a16z and, in 2022, a $23.5M Series A led by the OpenAI Startup Fund. The founding mission hasn't changed: build the world's first self-organizing workspace, where the AI handles organization so the human never has to.

Primary intent: Frictionless accumulation

Job it does:

“I capture everything and organize nothing - and I need an AI to make that archive useful anyway. And I don’t want to build a taxonomy that I’ll abandon in three months.”

It translates to:
Capture freely, any format
AI auto-organizes and tags
Related Mems surface contextually
Mem Chat synthesizes on demand
Smart Write drafts content from your history

Best for:
Consultants and knowledge workers who capture in bursts
Researchers who accumulate faster than they process
Writers who want a second brain that writes back
Professionals who’ve abandoned every PKM system they’ve tried

Mem fixes what Notion and Obsidian leave behind: the organizational debt. Both of those tools give you a powerful structure - and then require you to maintain it. Mem bets that the AI should absorb the organizational burden entirely. The user’s job is to capture. The system’s job is to make the capture retrievable.

Mem’s strongest feature is the one that’s hardest to describe in a demo: the feeling, after three months of active use, that the archive is smarter than you are. Mem Chat doesn’t just retrieve notes - it surfaces conceptual bridges between ideas you captured months apart and never consciously connected. The user who asks “What do I know about why early SaaS products fail?” gets a synthesis that draws from five different notes taken in five different contexts. That synthesis is genuinely novel. It’s not a search result.

The Smart Write function is Mem’s most underrated retention anchor. When the AI can draft content that sounds like you — because it’s drawing from your accumulated voice, your preferences, your past positions — the switching cost becomes real in a way pure note-taking tools never achieve. Your voice is in the KB now.

Churn Watch: Mem’s churn risk arrives precisely when the KB reaches critical mass. The zero-structure philosophy works elegantly at low volume. At high volume - thousands of notes, years of capture - users start to feel that the AI organization is serving the system rather than serving them. The problem that its organizing logic becomes opaque at scale. Users who built elaborate Notion systems and felt the organizational debt creep in have a muscle memory response to that feeling: they start looking for the exit.

The second churn vector is data trust. Mem can only be as useful as what you’ve put into it - which means gaps in capture create gaps in synthesis. Users who go through three weeks without capturing anything return to an archive that feels stale, and they blame the tool rather than the habit. Mem needs a gap-awareness layer - something that surfaces what’s missing from the archive, not just what’s in it.

When retention depends on accumulated capture, the winning strategy is not more intelligent retrieval - it’s making the cost of not capturing visible.

NotebookLM is a source-grounded research assistant designed to let users interrogate a defined body of material without reading every word. It doesn’t try to build a long-term archive, it doesn’t try to remember across projects, it doesn’t try to write to itself.

NotebookLM started as a research project inside Google Labs in 2023, originally called Project Tailwind. The core idea came from author Steven Johnson, who joined the team early: build an AI that helps you understand things - not the internet, not general knowledge, just the specific documents you’ve chosen to work with. It launched publicly in 2023, reached tens of millions of users without a marketing budget, and became Google’s quiet answer to the question of what AI-native research actually looks like. By 2026 it connects directly to Google Gemini, turning your notebooks into a persistent context layer across your entire Google workspace.

It tries to be the fastest path from a pile of documents to a usable understanding.

Primary intent: Project-scoped synthesis

Job it does:

“I have forty PDFs and twelve links and a week before this matters. Help me understand all of it without reading all of it.”

It translates to:
Upload sources (PDFs, URLs, YouTube, docs)
AI generates source map and summary
Natural language query against sources
Generate Audio Overview, mind map, flashcards
Export insight for use elsewhere

Best users:
Students preparing for exams or papers
Researchers synthesizing literature for a project
Analysts doing competitive or market deep-dives
Operators onboarding into a new domain quickly

NotebookLM fixes what document overload leaves behind: the two days of reading you don’t have. It is source-grounded by design - responses are anchored to what you uploaded, not what the model knows generally, which means answers are traceable and hallucinations are structurally constrained.

NotebookLM’s best design decision is also its most unusual: it doesn’t try to be a permanent brain. Each notebook is a project. The sources are scoped. The value is extracted. And then you move on. This is the exact opposite of the Karpathy pattern - and it is the right answer for a significant portion of knowledge work that isn’t about compounding but about sprinting through a defined body of material.

The Audio Overview feature has no real equivalent in the market. The ability to generate a conversational podcast from your uploaded documents - and then interrupt the hosts mid-sentence to ask clarifying questions - is a genuinely different cognitive interface. Users who learn through listening, or who process information while commuting, have found something in NotebookLM that nothing else offers.

Churn Watch: NotebookLM’s churn risk is structural and, in an important sense, by design. The notebook model means that knowledge doesn’t compound — each project is an island. Users who start with a scoped research task and then want to build a growing personal KB will hit the ceiling of what NotebookLM was built to do. The tool doesn’t grow with them. It doesn’t remember what they uploaded last month or cross-reference the current notebook against previous ones.

This is not a failure, but a positioning decision. But it creates a predictable churn moment: the user who graduates from “I need to understand this report” to “I need a system that builds on everything I’ve ever learned” will churn from NotebookLM toward Cabinet or Mem. NotebookLM’s retention strategy should not try to prevent this transition - it should own the project-scoped lane completely and build deeper into it.

When retention depends on task completion, the winning strategy is not accumulation - it’s becoming indispensable to the sprint.

6 retention mechanics to steal

1. Build compounding, not resetting. If your product’s value resets when a session ends, your retention depends entirely on the next session starting. That’s a fragile place to be. The tools that keep users longest are the ones whose value at month six is structurally higher than at month one - not because of new features, but because of what accumulated. Design for accumulation. Every interaction should leave something behind.

2. Solve maintenance or lose the busy user. The knowledge management graveyard is full of tools people loved during onboarding. The failure mode is always the same: maintenance debt builds during a busy period, the user stops tending the system, returns to find it stale, and leaves. If your product requires active upkeep to stay valuable, you’re asking users to work hardest for you exactly when they have the least capacity. Automate the maintenance, scope it away, or watch it become your churn trigger.

3. Know where your product stops working. The most invisible churn event isn’t frustration - it’s graduation. Users don’t cancel; they quietly stop using your product for the job they’ve grown into, while technically remaining subscribers. You won’t see it in your cancellation data. You’ll see it in engagement collapse three months before the cancel. Every product has a scope ceiling. The question is whether you know where yours is and whether you’ve designed anything for the moment users hit it.

4. Trust in AI-organized data is fragile and non-negotiable. The moment a user catches your AI getting something wrong - a synthesis that feels off, a gap that should have been filled - the trust relationship with the entire archive cracks. This is not a bug report. It’s a churn signal. Users don’t need perfection, but they need to believe the system is reliable enough to depend on. One bad retrieval at a critical moment undoes months of compounding goodwill.

5. The users hardest to lose are the ones who’ve gotten the most value. Switching costs in compounding systems don’t scale with subscription length - they scale with accumulated value. A user who has two years of captures in Mem isn’t just paying $15/month; they’re sitting on an archive that would take two years to rebuild. That asymmetry is your deepest moat, and most products never deliberately design for it. Ask yourself: what does a user have after one year that they couldn’t easily recreate somewhere else?

6. Tell users before they hit the ceiling. The churn event at the scale transition is usually not inevitable - it happens because the product never warned the user it was coming. If you know your product works brilliantly up to a certain scale and starts to strain past it, build the signal that says so. Offer the path forward before the user discovers the wall in frustration. The product that says “you’re outgrowing this - here’s what’s next” retains more than the one that quietly fails and waits to be abandoned.

Subscribe now

Personal KB tools didn’t converge into a single super-tool but split along lines of where in the user’s scaling curve they were designed to be useful.

Some tools exist to synthesize a bounded problem fast. Others exist to accumulate and organize without demanding maintenance. Others exist to compound institutional memory autonomously, indefinitely.

Churn emerges not when a product fails at its original job, but when the user’s job grows beyond what the product was designed to handle - and the product has no answer for what comes next.

In personal knowledge management, retention grows when the tool scales with the problem it was hired to solve. A knowledge base remains valuable only if the depth it offers today is not the ceiling it will hit tomorrow.

Key takeaways:

• Personal KB tools compete on scale tolerance, not intelligence - the question is not which tool is smarter, but which tool can grow with the user who hired it.

• Project-scoped tools like NotebookLM retain users by becoming indispensable to the sprint; their churn event is structural graduation, not dissatisfaction.

• Accumulation-first tools like Mem retain users through compounding capture value, but face a trust collapse when the AI organization becomes opaque at high volume.

• Automation-first tools like Cabinet eliminate the maintenance bottleneck entirely - their churn risk lives at setup friction and agent output quality, not at feature gaps.

• The most important churn event in this category is invisible in standard metrics: the user who silently stops using a tool for the job they’ve scaled into, without ever formally churning from the job they started with.

Most founders I’ve seen build great products stumbled into them - not through a market map, but through a moment they couldn’t ignore.

Most of the products that failed at retention were built by founders who understood the opportunity better than the problem. The distinction sounds subtle, but the churn data makes it obvious.

In this issue:

There are two ways to decide what to build.

The Opportunity Cycle starts with the market. You identify the gap, size the audience, map the competitive landscape, and build toward the largest addressable segment. The revenue case for this approach is real and legible: because the product is designed for a broad audience, acquisition scales quickly - paid channels work, the positioning is clear enough to convert, and the early growth curve looks healthy. Many of these products cross their first million in ARR faster than anything built from personal obsession, because they’re engineered to capture demand that already exists and is easy to reach.

The problem is what happens next. The cohorts slope downward at a rate no growth spend can outrun, because something in the product is subtly wrong in a way the founder cannot see from the outside. It’s not broken - it works, it converts, it reviews adequately - but it was designed for the category of person experiencing the problem rather than for the specific, urgent, 2am version of it that the people who need it most are actually living. Users feel that difference even when they can’t name it. When retention drops the Opportunity Cycle founder improves onboarding; when onboarding improves and retention still drops they add features; when features don’t hold users they adjust the paywall or run a re-engagement campaign. The product becomes more complex and the churn problem doesn’t move, because the product was always solving the stated version of the problem rather than the real one - and the founder never had the intimacy with the problem to know the difference.

The Problem Cycle runs differently, and it is not a faster or more elegant one. It is slower, messier, and mostly invisible from the outside.

It starts not with a market but with an experience - something that happens to the founder personally, in a way they cannot ignore. They go deep on it not to validate a business idea but because they need to understand what is happening to them. They find everything written about it, speak to professionals, develop a precise personal vocabulary for what the problem actually feels like at its worst - the kind of understanding that no brief could give them and no user interview could replicate. Then they look for solutions and find only approximations: tools that work for the average case but not for the person who needs help most urgently. So they build something themselves - scrappy, value-first, technically imperfect - where execution is secondary to the feeling of finally getting the thing right. They launch quietly, often while holding down another job, with no paid acquisition budget and no growth playbook.

Early revenue is slow because they are finding the people for whom this is the only thing that works, and those people are harder to reach but far less likely to leave. Then something compounds: the users who stay leave reviews that read like personal testimonials, the founder responds to every one, treats each as a product signal, ships the improvement, and the product becomes more precisely right for the person who needs it most. Word of mouth does the work that marketing budgets usually have to. The organic growth that looked like a disadvantage turns out to be the proof - products found by people specifically looking for them retain at a structurally different rate than products discovered through a paid channel.

The timeline is long. Building takes months; growing to a meaningful business takes one to three years before the recognition arrives and the revenue crosses into seven figures. The founders who make it through that window are the ones who were never building from a brief - they were building the answer to a problem that had been theirs for long enough that they couldn’t let someone else get it wrong.

What separates the products that last from the products that don’t isn’t - is the depth of the founder’s relationship with the problem before the product existed.

There is a useful test for whether you have this. Build something that - if you found out someone else was building it - you would get genuinely upset, not competitive but upset, because you’d know they might not understand it well enough to do it right. That feeling is the signal: you are not building an app, you are building the answer to a problem that has been yours for long enough that you cannot let someone else get it wrong.

As Paul Graham emphasizes in his essays, founder market fit is foundational in building a company that is solving a problem the founder understands intimately.

Three products, three founders, three very different problems - all built by people who had no business building apps except for the fact that they couldn’t stop thinking about a problem no one else was solving well enough.

Ania Wysocka experienced her first panic attack in her final year of university, reached for her phone, and found nothing useful - just a medical app and a hypnosis tool, neither of which spoke to someone alone and in crisis. What followed was years of immersion in anxiety research, work with mental health professionals, and a self-taught design background, all of which gave her a precise personal vocabulary for what relief actually needed to feel like. By 2017, she partnered with a student developer to launch Rootd on World Mental Health Day -- a single red panic button at the centre of a deliberately un-clinical design - and crossed $1M in revenue six years later.

Rootd is optimized for what I’d call crisis utility - and it does its job better than anything in its category.

Primary intent: Immediate relief

Job it does:

“I’m in the middle of a panic attack, and I need something that understands what’s happening to my body right now - not a meditation, not a hotline, not a breathing exercise that assumes I have 10 minutes.”

It translates to:

✔️ Panic button triggered
✔️ CBT-backed guided interruption
✔️ Breathing and grounding tools
✔️ Stats page (panic attacks survived, warrior points)
✔️ Long-term lessons for anxiety management

Rootd intentionally ignores the broader wellness category - sleep coaching, journaling ecosystems, mood tracking dashboards - treating the panic attack as the moment that matters, not a data point in a larger health picture.

Churn watch: The product’s success creates the disengagement, just as in dating apps - you delete it once the goal is achieved. Users who feel meaningfully better start to deprioritise anxiety management, which means the strongest signal of Rootd working is also the earliest signal of potential churn. The structural answer is long-term lessons that shift value from crisis response to lasting change and a new crafted routine - a job that doesn’t end when the panic attacks become less frequent.

Travel Diaries is a memory architecture tool built for people who experience things worth keeping. Veerle nailed that photos and notes are not the output - the permanent, shareable, printable record is.

Veerle Witte was a travel journalist writing for National Geographic Traveler when, during a four-month trip through South America, she searched for a tool that could hold long narrative text alongside photos and route maps, allow sharing as a blog, and eventually produce a printed book - and found nothing. She built Travel Diaries in eight months while continuing to write, ran it as a side project for eight years, and only went full-time after external capital arrived in 2019. By then: 200,000 users, 15,000 printed books ordered, a 4.9 App Store rating.

Primary intent: Memory preservation

Job it does:

“I want to document this trip - or this season of life - in a way that actually survives. Not a camera roll. Not scattered notes. Something I’ll still want to open in ten years.”

It translates to:

✔️ Write long-form entries
✔️ Add photos and maps
✔️ Arrange layouts
✔️ Share as a blog and order as a printed book

Travel Diaries ignores the social feed and the engagement loop - it’s not competing with Instagram but replacing the printed photo album and the years of guilt about never quite making one.

Churn watch: The window between trips. Users who built a beautiful diary of their Portugal trip and don’t travel again for six months may not open the app for that entire period. Veerle’s expansion into Baby Diaries - and planned cookbook and journal apps - is the direct answer: widen the events that trigger the job so users never have to wait for the next trip to have a reason to return.

Ahana Banerjee had been tracking her cystic acne in spreadsheets since her early teens - logging products, ingredients, routines, and photos across a decade of trying everything and understanding almost nothing about what was working - so that by Imperial College London she had a precise picture of exactly what a tracking tool needed to do that nothing on the market could. When she pivoted to Clear at Y Combinator with four days between the idea and the interview, her MVP was a web app where she personally made every recommendation - “powered by Ahana.AI” - because her domain expertise was, at that point, more valuable than any algorithm she could have built in the time available. Clear was featured as App of the Day four times; YC and ~$1M in pre-seed funding followed roughly three years in.

Clear is a skincare intelligence platform built for people who’ve run out of patience with guessing. Ahana nailed that products aren’t the problem - the inability to tell what’s working is.

Primary intent: Skin clarity (the understanding kind, not just the visual kind)

Job it does:

“I’ve spent years buying products and not knowing if any of them are actually doing anything. I want to track my skin the way I’d track a workout — and finally see what’s moving the needle.”

It translates to:
✔️ Selfie intake
✔️ Product logging
✔️ Routine tracking
✔️ AI-backed analysis
✔️ Community validation
✔️ Personalised recommendations

Clear fixes what dermatologists, influencers, and Reddit threads all leave behind: a personalised, longitudinal view of what’s actually working on your specific skin.

Churn watch: After the first improvement plateau. Users who see progress in the first 8–12 weeks have a strong retention signal, but if the skin stalls and the app can’t explain why or suggest what to change, the tracking starts to feel like just another chore. The moment the data stops producing insight, the habit starts to break.

6 retention mechanics these founders got right

1. Authentic voice as UX. None of these products sound like they were written by a content team, and that’s not a coincidence - it’s a direct consequence of founders who lived the problem so thoroughly that the copy couldn’t be generic even if they’d tried. When users feel genuinely understood at the language level — not just the feature level - they don’t go looking for alternatives, because there is nothing to compare it to.

2. Personal response as a feedback loop. All three founders treated every user review as a direct signal rather than a support ticket, with Ania personally responding to every one - investigating the issue, shipping a fix, encouraging the user to update their rating - which compressed the feedback cycle from months to days and built a visible public record of a founder still paying attention. Users who feel heard by the person who built the thing don’t just stay - they tell people.

3. Freemium structured around the highest-need moment. Rootd’s panic button was free for years — ungated, front and centre — not because Ania hadn’t thought about monetisation but because she had experienced a panic attack and knew that interrupting someone mid-crisis with a paywall would be a betrayal of everything the product stood for. The cost to revenue was real; the gain in trust was larger, and it meant users had an emotional relationship with the product before they were ever asked to pay.

4. Growing through a gap, not against a competitor. None of these products won users away from a competitor — they found people who had given up looking, which means discovery ran through organic search, word of mouth, and press rather than category switching. That growth is slower but produces users who are pre-qualified, high-intent, and far less likely to churn.

5. Niche precision as retention architecture. Rootd launched as a panic attack app before becoming an anxiety app — owning the narrower, higher-intent keyword first — because the most specific version of a problem is always the one the founder understands best, and users who find exactly what they searched for have no reason to look for alternatives.

6. The founder’s experience as the product’s implicit promise. Each of these products carries something no competitor can replicate: the legible presence of someone who has been exactly where the user is and built this because nothing else was good enough. When users can feel that presence — and they can, even if they can’t name it — the product becomes something closer to a relationship than a subscription, and relationships churn at a fundamentally different rate.

These apps split along lines of what each founder had personally lived through — one for surviving a crisis, one for preserving what matters, one for finally understanding what’s been happening to your body. What they share is not a product type but a founding condition: a problem personal enough, urgent enough, and poorly served enough that the founder couldn’t stop thinking about it until the product existed.

Churn emerges when a product forgets why it was built. Retention grows when users can feel — at every touchpoint — that the person who made this actually needed it too.

Key takeaways:

• The technical build is a permission slip, not a competitive advantage — execution buys you the right to be in the market, and what keeps you there is the depth of understanding that no amount of engineering can replicate.

• Recognition in a “replace nothing” category takes years, not months, and the founders who win are the ones who sustain the quiet period without mistaking slow growth for a broken product.

• Freemium structured around the user’s highest-need moment builds more durable trust than freemium structured around withholding value — the paywall works better after the product has already earned it.

• Personal response to user feedback is a retention mechanism with compounding returns, not a support cost — every review treated as a signal and every response treated as a relationship pays out over years.

• Niche precision at launch is a retention decision as much as a positioning one, because users who find exactly what they searched for have no reason to look for alternatives and arrive with the highest intent the market can produce.
  

  Thanks for reading Do Not Churn! Subscribe for free to receive new posts and support my work.

In the early subscription economy - gym memberships, cable bundles, early SaaS - the cancellation process was an intentional wall, and an obstacle that was stopping users from leaving. The logic was simple: if leaving is harder than staying, people tend to stay - the company keeps protecting ARR. Nobody asks what went wrong.

For a long time, it worked. Primarily on monopolistic markets - cable, telephony, gyms - there were no real alternatives, and people badly needed at least some product. You couldn’t leave Comcast for a better provider if Comcast were the only wire running to your building.

And you couldn't cancel Amazon Prime without navigating what engineers internally called the Iliad Flow - a four-page, six-click, fifteen-option sequence designed not to let you cancel, but to stop you.

That, after lasting long, thankfully, has changed - the shift happened when switching costs dropped to zero. Once competition became very diverse and more alternatives appeared on the market, users started having somewhere else better to go. The energy required to cancel was less demanding than the energy required to stay frustrated. Companies that spent decades building walls instead of reading signals found themselves with a retention strategy, but no retention intelligence.

That’s the structural problem the Retention Wall always contained. It was optimised to prevent departures. It was never designed to understand them.

The Retention Wall - optimised for what I’d call defensive capture, and it worked, for a while.

Amazon Prime’s cancellation process was so architecturally complex that it had an internal code name: The Iliad Flow.

The reference to a 10-year siege was not ironic - it was descriptive. Users who located the cancellation page were redirected through multiple screens offering discounts, free months, and the option to “pause” instead - each one a diversion, not a path forward. The FTC described it as a four-page, six-click, fifteen-option sequence. Internal Amazon data showed that the flow reduced cancellations by 14%. That number was not a warning sign. It was the point.

The FTC sued in 2023. Three days into trial in September 2025, Amazon settled for $2.5 billion - including $1 billion in civil penalties and $1.5 billion in consumer refunds. Three senior executives were named personally. Internal emails presented at trial showed that engineers had flagged the cancellation features as confusing, and that those simplifications were rejected by leadership because they adversely affected Amazon’s bottom line.

Why it doesn’t work

1. The critical structural weakness of the Retention Wall is that the only metric it generated was “cancel/stay.”

Whether the user stayed because they saw value or because they gave up trying to leave was invisible. Companies accumulated almost no behavioral intelligence about why users were leaving or what they needed differently. Aggregate churn rate was the only visible signal. The wall looked like it was working right up until it wasn’t.

2. ‘The Wall’ begins to fail the moment an alternative exists.

Once a user has somewhere better to go and enough frustration to get there, friction becomes aggravation - and aggravation accelerates departure rather than delaying it. The Retention Wall’s churn risk isn’t at the door. It’s in the years of product signals the company never received, and the retention infrastructure it never built, because the Wall seemed to be working.

3. Any dark pattern kills trust that you’ll never gain back

This is the most valuable currency that is reflected across a several metrics, yet it’s hard to measure it purely. But, despite of the, it’s very clear that if you lose users trust - they would hardly ever recommend you to any of their friends; they would never come back, even if they need will reoccur.

The Signal Reader - optimised for what I’d call intent-aligned departure, and this is the new standard.

Lovable is an AI-powered app builder that hit $75M ARR in seven months and $400M ARR by March 2026 — one of the fastest-growing products of the vibe-coding era. Its growth was extraordinary. Its retention challenge was structural: the product creates something you build, not something you keep building every month. Users who finished their main project had no obvious reason to keep paying full price.

The cancellation data surfaced the same response over and over. Too expensive. Standard Retention Wall instinct would have offered a discount. Lovable’s Head of Growth, Elena Verna, asked a different question: expensive relative to what the user is doing right now?

The answer: users who had finished building didn’t need credits. They needed to keep their project alive - the custom domain, the private project, the paid features that kept their work from expiring. They weren’t leaving because Lovable had failed them. They were leaving because Lovable only had one pricing tier for a user whose job had changed.

The response was a $5/month Lite Plan - access to paid features, no additional build credits. Not shown on the main pricing page. Surfaced only inside the cancellation flow, for the specific cohort of users whose behavior matched the pattern.

broke down the full test - including why $5 beat $10 and $15 - in her newsletter: We stopped forcing the subscription model on our users. Here's what happened.

Lovable nailed that cancellation data is not a signal to suppress - it’s a product research instrument. The Lite Plan wasn’t a retention hack. It was a product decision made possible by reading what “too expensive” actually meant in behavioral context. That’s the structural difference: The Retention Wall treats “too expensive” as an objection to overcome. The Signal Reader treats it as a question to answer.

The additional mechanics Lovable built from this same philosophy: credit rollovers, daily free credits, in-product payment failure alerts (50% improvement in recovery rate over email-only), and ad hoc top-ups that produced a 7% lift in overall retention. Each one came from the same starting position — what is this user’s behavior telling us, and what does the product not yet have a response for?

Churn watch: Lovable’s residual churn risk is the user who builds one thing and doesn’t know what to build next. The Lite Plan addresses the user who has finished building. It doesn’t yet address the user who lost momentum - that’s closer to an activation and re-engagement challenge than a pricing one.

The Re-entry Designer - optimised for what I’d call relationship preservation, because it understood something Amazon didn’t.

Netflix studied its re-subscription cohort data and found something that looked, on the surface, like a churn problem but was actually something else entirely. Roughly 50% of users who cancelled eventually returned. They weren’t leaving because Netflix had failed them. They were leaving because they’d finished watching their current series, had nothing they wanted to watch right now, and were paying $15+ a month for a service they weren’t actively using.

That’s not a dissatisfied customer. That’s a seasonal one.

The insight reframed the entire retention problem. If the user is going to come back anyway - and the data said they would - then the job isn’t to prevent the cancellation. It’s to make the return as frictionless as possible, and to ensure that the relationship survives the gap intact.

Netflix’s response was structural. Pause functionality: billing suspended for up to three months, full access preserved, one-click resume.

Account preservation after full cancellation: watchlist, viewing history, and preferences stored for ten months. When a user returns, they arrive back into their own account - not a blank slate that makes re-entry feel like starting over.
One of the churn problems consists of the fact that the customer doesn’t fit into any of the created categories with his usage patterns and with what he thinks is valuable for him.

According to Churnkey, the insufficient usage is second-highest cancellation reason

Subscription products didn’t converge on a single retention philosophy — they split along lines of belief.

Churn emerges when a product’s retention mechanics contradict its retention values - when a company says it cares about customers and then makes them navigate an Iliad Flow to leave.

In subscription retention, trust grows when departure is as easy as arrival — and the product learns something from every decision either way.

A subscription remains worth paying for only if the company knows what it’s worth to the specific person paying for it right now.

Key takeaways

• Subscription products compete on whether they can learn from departure - not on how hard they make it.

• The Retention Wall worked in low-competition, high-switching-cost markets. When alternatives appeared, it had nothing left.

• Amazon’s $2.5B FTC settlement is the invoice the Retention Wall eventually sends.

• Lovable’s Lite Plan is what happens when “too expensive” is treated as a behavioral question rather than a pricing objection.

• Retention follows intent alignment. Churn begins when the product stops having a response for where the user is in their journey — and the Retention Wall, by design, never had one.

  Thanks for reading Do Not Churn! Subscribe for free to receive new posts and support my work.

If you liked this article, read these next:

Search to Swipe: what three decades of dating apps reveal about customer experience and churn - How engagement-first design rewarded motion over outcomes, and accidentally trained users to disengage.

Why AI note-taking tools no longer compete on features - Instead, they compete on intent: preserving decisions, building understanding, or capturing ideas before they disappear.

From swipe fatigue to IRL: Inside the apps racing to get you off your screen - The products building genuine connection - and why their churn profile looks completely different.

Why buying a plane ticket feels harder than it did 30 years ago - How flight booking evolved from delegated trust to endless comparison, and what the best products do differently.

Is app-of-apps becoming the new interface for work and life? - The six retention mechanisms that make daily assistants impossible to leave.

Building a product used to be the hard part. It isn’t anymore.

Lovable, Bolt, Replit (you name it) and a weekend and a clear idea is now sufficient to ship something real. Which means the internet is being flooded with products: hundreds per day, each one reasonably well-built. Each one looking for users, each one trying to be found in the same saturated channels, by the same finite number of people who have the same finite amount of attention.

Now, AI is being asked to solve the problem it created. Agents that audit your site, track your brand inside ChatGPT, write your social posts, queue your outreach -- all running autonomously, all promising to replace the marketing team you never hired.

The pitch is compelling: outsource the entire distribution function to a fleet of agents and get back to building.

The honest version of that pitch has a question buried inside it. Whether handing your distribution to an AI agent actually works -- not in the demo, but on the morning when the output lands in your inbox and you have to decide whether to trust it -- is what this article is really about.

About this newsletter

I’m Daria Littlefield. I spent a decade running customer operations across 35+ dating and social apps - the kind people delete in a rage and reinstall a week later. That taught me more about why products lose users than any framework ever could.

Do Not Churn is where I put that to use. Every week I break down the retention mechanics behind the products we all use - what keeps people, what loses them, and why the difference is rarely what you’d expect.

If that’s useful to you, subscribe below. If you have a take on whether AI can actually replace a marketing team - drop it in the comments.

Leave a comment

In the early 2010s, distribution was the unsexy part of building a company. You had a product, you needed users, and the path between those two things was entirely manual. You had to cold-emailed journalists, post in niche forums.
If you had a budget, you ran Facebook ads and A/B tested subject lines.
If you didn’t, you did the same things with less coffee and more desperation.

Then growth hacking became a discipline. Sean Ellis coined the term in 2010. Airbnb reverse-engineered Craigslist. Dropbox built a referral loop.

What these companies understood -- and what most founders missed -- was that distribution wasn’t a single campaign, it was a carefully designed system - designed it once, it ran continuously, and it compounded. The founders who grasped this scaled fast. The ones who kept treating distribution as a one-time effort to revisit quarterly - didn’t.

The problem was that building a distribution system still required a team - a growth engineer, a content strategist, and an SEO specialist, also someone to run LinkedIn, someone to manage outreach sequences.

Most early-stage companies couldn’t afford all five and tried to split those jobs across two founders who were already running out of hours.

AI was supposed to fix that: automate the execution. Let a solo founder punch above their weight. And the tools that arrived between 2023 and 2025 mostly delivered on the automation half of that promise. Scheduling, drafting, sequencing -- all reduced to minutes.

The real shift happened with the collapse of the building barrier.

For years, distribution was hard because building was also hard. The two problems competed for founder attention, and most early-stage teams made a rational choice: ship first, figure out growth later. In reality, it reflects something more structural now: building is no longer the bottleneck. It hasn’t been for about eighteen months.

Lovable, Replit, Bolt, Cursor -- the vibe-coding era didn’t just lower the cost of building a product. It removed the floor entirely. A weekend and a clear idea are now sufficient to ship something real: what used to take a three-person engineering team three months now takes one person with a few good prompts and a free Sunday afternoon.

The consequence nobody fully anticipated: the internet is now being flooded with products: hundreds per day, thousands per week. Each one is reasonably well-built, functionally indistinguishable from the next at the feature level, and competing for the same finite amount of human attention.

Distribution became harder as the supply side exploded while the demand side stayed flat. More products are looking for users than there have ever been, and the channels that used to carry them -- Google search, Product Hunt, Hacker News, X virality -- are saturated in a way they weren’t three years ago.

Okara, Typefully, and Castmagic are not, fundamentally, marketing tools -- they are the first wave of infrastructure responding to a structural imbalance: the gap between how fast products can now be built and how fast they can actually be found.

AI distribution replaced scheduling, then sequencing, and now it’s racing to keep pace with a building wave that never stops shipping

Okara is positioned for what I’d call ambient organic authority building -- and it launched on March 16, 2026.

The company was founded in 2025 by Fatima Rizwan, a founder based in Singapore running a globally distributed team spread across India and Malaysia. Rizwan built Okara first as a privacy-first AI chat platform -- encrypted conversations, 20+ open-source models, no data training -- before pivoting the product surface toward autonomous marketing. That background matters. She didn’t come from growth or SEO. She came from a conviction that professionals needed powerful AI tools they could actually trust with sensitive information. The AI CMO is that thesis applied to distribution: an agent that works on your behalf, on infrastructure you don’t have to second-guess.

Primary intent: Organic visibility

Job it does:

“I want traffic and users to find me without spending my entire week on SEO. And… I don’t want an agency sending me a monthly PDF of vague recommendations I can never act on.”

It translates to:
✔️ Submit URL
✔️ Agents audit site and brand presence daily
✔️ Specific fixes land in your inbox each morning
✔️ GEO Score tracks visibility inside ChatGPT, Claude, Perplexity
✔️ Social and community agents distribute content autonomously

The GEO Score is the one genuinely novel idea here. As AI-powered search handles a growing share of discovery queries, appearing inside LLM-generated answers is becoming as important as a Google ranking. Okara is the first tool systematically measuring this signal.

Before you pay, the product shows you an agent working -- scanning your site, building context, loading company data. It feels like something significant is happening. What you actually get on the free tier is a business description and a list of pending tasks, all locked behind the paywall. The dopamine of watching an agent run is building up anticipation, but the output, at least publicly, is not yet commensurate with that experience.

What no one is saying loudly enough is what Okara -- and any AI CMO -- structurally cannot replace: strategy and taste. An agent can audit your site, fix your meta descriptions, and schedule your posts. It cannot tell you whether your positioning is wrong, whether you’re talking to the right audience, or whether the angle of your content is interesting enough for anyone to care. Those decisions don’t live in a prompt. They live in a founder’s judgment about their market, their product, and what makes their thing worth finding in the first place. The tools that survive long-term will be the ones that amplify good taste -- not the ones that promise to substitute for it.

Optimised for:
• Pre-revenue founders with a working product but no marketing function
• Bootstrapped solo operators who can’t justify $3k/month tools
• Indie builders post-launch, trying to get first organic traction
• Early-stage SaaS teams that need daily momentum without daily effort

Churn watch: Okara’s churn window is days 30–60. SEO and GEO are lagging signals -- you won’t see traffic movement in the first month, regardless of what the agents do. If the daily inbox briefs feel generic or the GEO Score doesn’t show directional improvement within six weeks, founders will cancel before the system has had time to work. The bigger risk is the paywall architecture: when the most impressive part of the experience is the loading animation and the most valuable outputs are locked, users who haven’t converted will never know if they missed something real or just a more polished version of what they already have.

Typefully is optimized for what I’d call social presence compounding -- and unlike the other two tools in this analysis, it has the receipts to prove it works.

The product was born in June 2020 as a weekend side project by Fabrizio Rinaldi, a designer, and Francesco Di Lorenzo, an engineer -- two Italians who met on Twitter in 2011 and had previously built and sold Mailbrew, an email digest tool. They made Typefully to scratch their own itch: the official Twitter thread experience was, in their words, simply broken. The side project kept growing. By 2022, Evan Williams -- one of Twitter’s original co-founders -- came on as an angel investor. Today it’s used by 220,000+ creators, bootstrapped, no VC, based in Portugal. Founders report growing X followings by 5,000+ in six months. Users who barely posted on LinkedIn started posting consistently the week they signed up.

Primary intent: Social reach

Job it does:

“I want to show up consistently on X and LinkedIn without managing two separate tools, losing drafts to platform bugs, or spending my evenings manually adding calls-to-action under posts that happened to go viral. And I don’t want to think about which platform to post to -- I just want to write once and be everywhere.”

It translates to:
✔️ Write once in distraction-free editor
✔️ Cross-post simultaneously to X, LinkedIn, Bluesky, Threads
✔️ AI rewrites and optimises per platform
✔️ Schedule at AI-recommended best time
✔️ Auto-Plug fires a CTA automatically when post crosses your engagement threshold
✔️ Analytics show what landed, by platform and post type

The Auto-Plug feature is structurally underrated. When a post crosses a founder-defined engagement threshold -- say, 200 likes --Typefully automatically adds a promotional reply without the founder needing to monitor their feed. In a world where viral moments are unpredictable and short-lived, this converts attention into action at the exact moment it’s most abundant. It’s not automation for the sake of it. It’s distribution timing solved.

What Typefully gets right is the friction removal. X’s native editor loses drafts. LinkedIn’s interface punishes frequent posters. Typefully gives both channels a single, clean writing surface -- Notion-like, auto-saving, with real-time pixel-perfect previews of how the post will actually render. Users don’t post more because Typefully tells them to. They post more because the experience of posting has finally stopped being painful.

Optimized for:
• Founders building personal brands alongside their products
• Creators who live on X and want LinkedIn reach without LinkedIn effort
• Solo operators whose distribution strategy is their own voice and following
• Small teams where the founder is the marketing department

Churn watch: Typefully’s churn is a posting consistency problem, not a product problem. The tool delivers exactly what it promises -- but only when the user keeps showing up to write. Founders who post fewer than three times a week see marginal follower growth, open their analytics dashboard, feel disappointed, and quietly downgrade to the free tier.

The streak feature -- a GitHub-style contribution graph for posting frequency -- is the right counter-lever. It makes consistency visible and psychologically costly to break. The metric to watch is weekly active posts per user: below three, retention risk climbs sharply. Above five, users almost never leave.

Castmagic is optimized for what I’d call content surface area multiplication -- and it’s the most validated product in this group by a significant margin. 75,000+ users. 10 million+ minutes processed. It has the receipts.

The company was founded in Miami by Blaine Bolus and Ramon Berrios, who were already running DTC Pod -- a podcast for founders in the commerce and creator space -- when they realised that producing the show was taking more time than recording it. They enlisted engineer Justin Tormey and built Castmagic to solve their own problem. Bolus is a serial founder: before Castmagic he co-founded Seated (backed by Greycroft, Craft Ventures, and Insight Partners) and OmniPanel (backed by CRV). He bootstrapped Castmagic from zero to $2M ARR in under a year.
That trajectory -- a podcaster building a tool for podcasters, then watching it expand into agencies, coaches, and enterprise teams -- is exactly the kind of founder-product fit that makes the churn mechanics worth studying.

Primary intent: Content amplification

Job it does:

“I want every podcast, client call, and recording I make to turn into a week of published content across every channel. And I don’t want to spend six hours reformatting the same ideas for LinkedIn, the newsletter, and the blog.”

It translates to:
✔️ Upload recording or pipe from RSS/Zoom/YouTube
✔️ AI transcribes with speaker ID and filler word removal
✔️ Blog post, show notes, social posts, newsletter, video clips generated simultaneously
✔️ Magic Chat lets you query the recording as a document
✔️ Content exported or auto-scheduled

What Castmagic gets right is the ratio. One hour of audio becomes eight to twelve publish-ready assets in minutes. That’s not just a productivity improvement -- it’s a distribution model change. The content already exists. Castmagic multiplies the number of places it can be found. In a world flooded with vibe-coded products competing for the same attention, showing up consistently across every channel is no longer a nice-to-have - it’s the baseline.

Best users:
• Podcasters and video creators publishing regularly
• Marketing agencies managing multiple client content programmes
• Coaches and consultants converting session recordings into distributable IP
• Revenue teams are turning sales calls into objection libraries and case studies

Churn watch: Castmagic’s churn is a consistency problem, not a quality problem. Users who process their first five episodes are genuinely delighted. The risk emerges when the library grows past 20 or 30 recordings and navigation becomes cognitively expensive. If past content doesn’t resurface semantically — if users can’t search across episodes, connect recurring themes, or see what performed — the library becomes a graveyard. Assets accumulate but value doesn’t compound visibly.

The 6 retention mechanisms

1. Proof before paywall (the Okara problem) The most dangerous conversion architecture in AI distribution tools is one where the most impressive moment -- the agent working, the progress bar moving -- happens before the user has paid, and the actual output is locked behind the subscription. This creates a trust inversion: the product earns maximum engagement at zero revenue, then asks for commitment before delivering proof of value. Tools that show a real output sample before asking for payment will outperform those that use the working animation as the conversion mechanism.

Why it prevents churn: Users who converted on real output evidence cancel less than users who converted on aesthetic anticipation.

2. Streak mechanics as posting infrastructure (Typefully) Typefully’s posting streak -- a visible chain of consecutive days with at least one published post -- is a behavioural design choice that has nothing to do with features and everything to do with retention psychology. The streak makes the cost of stopping visible. Breaking a 34-day streak feels worse than not starting one. This is the same mechanism that keeps Duolingo users coming back at 11pm to avoid losing their streak -- not because they suddenly want to learn Spanish, but because the chain has acquired its own value.

Why it prevents churn: Distribution tools that make consistency visible and costly to break retain better than those that only reward output volume.

3. Asset library gravity (Castmagic) Every recording processed adds to a searchable, queryable library. Magic Chat means users can ask “what did I say about pricing in the last three months?” and get an answer across their entire back catalogue. The library becomes a second brain. Leaving means abandoning accumulated IP.

Why it prevents churn: The depth of the library makes Castmagic harder to leave with each recording added. Switching cost is informational, not contractual.

4. Metric ownership as category lock-in (Okara’s GEO Score) If the GEO Score becomes how founders measure AI search visibility -- the way Domain Authority became the SEO proxy metric for a decade -- then Okara owns both the measurement and the optimisation simultaneously. Users don’t leave tools that define the benchmark they’re optimising for. The risk is that Okara needs to establish this metric before a better-funded competitor builds the same measurement natively.

Why it prevents churn: The metric creates attachment before the outcome does. Users stay to watch the number move even when traffic hasn’t yet confirmed the theory.

5. Speed-to-first-result as the conversion event (all three) Across every AI distribution tool, the churn that happens in days one through fourteen is almost entirely caused by the same thing: the user hasn’t seen anything real yet - a follower gained, an article was published, a score moving. Tools that engineer a visible first result within 24–48 hours of signup retain dramatically better than tools that require setup, warm-up periods, or patience. In a category built for founders who can now ship a product in a weekend, patience is not a feature.

Why it prevents churn: The brain doesn’t cancel things that have already produced something. Even a small first win reframes the tool from “experiment” to “part of the workflow.”

6. The visible cost of cancellation (all three, but sharpest in Typefully) Most SaaS tools go quiet when you cancel them. Distribution tools do the opposite -- they go loud. The day a founder stops using Typefully, their posting streak breaks and their follower growth flatlines on the analytics dashboard they can still see. The day they stop feeding Castmagic recordings, the social queue empties. The day Okara's agents stop running, the inbox goes silent and the GEO Score freezes. The absence is immediate, visible, and personal. This is structurally different from cancelling, say, a project management tool, where the cost takes weeks to feel. Distribution tools create a daily present reminder of what you had -- not just a vague sense that something useful is gone.

Why it prevents churn: Cancellation in most SaaS is passive -- you stop paying and life continues. Cancellation in distribution is active -- you watch something you built start to decay in real time. Founders who have seen their posting streak sitting at zero for three days are far more likely to resubscribe than founders who simply forgot the tool existed. The loss is visible before it becomes permanent.

KEY TAKEAWAYS

• AI distribution tools are not a marketing trend -- they are infrastructure responding to a structural imbalance: building a product now takes days, but getting it found still takes months

• Organic visibility tools like Okara retain users by owning a new measurement metric (GEO Score), but face a structural paywall problem: the most impressive moment happens before the value is proven

• Social compounding tools like Typefully retain users through streak mechanics -- making consistency visible and costly to break turns posting from a chore into a chain nobody wants to end

• Content amplification tools like Castmagic retain users through asset library gravity -- the accumulated recording library creates a switching cost that compounds with every upload

• Retention follows output credibility; churn begins the moment a founder edits an AI-generated output and thinks, “I could have written this faster myself.”

So - can marketing be automated?

The execution, yes. The judgment, taste and strategy fit -- not yet -- and that gap is exactly where products go invisible.

If you tracked what you ate yesterday, you probably took a photo. Not typed a name into a search bar. Not weighed anything on a kitchen scale. Just point your phone at a plate and let the app figure it out. That felt normal. Maybe even obvious. Hard to imagine it was a feature that didn’t exist two years ago in any app that actually worked at scale.

That was then. This is now, and now just got more interesting.

That wasn’t always the case.

In the early 2010s, calorie tracking was a discipline. MyFitnessPal had already cracked the habit loop - build the streak, log the meals, watch the numbers compound. But logging required effort. You searched a food database. You estimated portion sizes. You picked the closest match from a list of 47 variations of “grilled chicken breast.” Memory was distributed and manual.

That design reflected the user it was built for: someone motivated enough to learn the system and disciplined enough to maintain it. That’s a real user. That user still exists. But it’s not everyone.

The broader population -- the one that downloaded a wellness app for January and deleted it by March -- was never going to sustain that workflow. Logging required a willingness to confront every calorie, in real time, with precision. For many people, that precision felt like punishment.

Then AI-native photo recognition arrived. OpenAI’s Vision API matured. Two high school students -- Zach Yadegari and Henry Langmack -- looked at what was newly possible and built Cal AI: point your camera, get a calorie estimate, move on. Just a photo and a number.

The product was not trying to be accurate to three decimal places. It was trying to remove the activation cost of logging entirely. That turned out to be exactly the right bet.

The real shift happened with the collapse of the effort barrier.

For two decades, calorie tracking had been evolving -- from paper food diaries to searchable databases to barcode scanners to photo logging. On the surface, this looks like an incremental tooling improvement. In reality, it reflects something more structural:

The moment nutrition tracking stopped requiring nutritional knowledge.

Market at a glance:

The diet and nutrition apps market reached $2.14 billion in 2024 and is forecast to hit $4.60 billion by 2030, advancing at an 13.14% CAGR. Growth drivers include rapid improvements in AI-based food recognition, expanding corporate wellness budgets, and wearable ecosystems that link continuous biometric data with dietary guidance. Zooming out, the broader digital fitness apps market stood at $13.59 billion in 2025 and is forecast to climb to $24.74 billion by 2030, at a 12.7% CAGR. The broader gravity includes digital therapeutics, GLP-1 companion apps, and corporate wellness platforms -- making this a consolidating category, not a fragmenting one. The players with data depth and distribution will absorb the players with UX breakthroughs.

Key market drivers:
1 - AI food recognition: AI-driven portion-sizing is reducing logging friction at a projected 13.89% CAGR to 2030
2 - Subscription monetization: Premium tiers are advancing at 14.5% CAGR as users pay for AI coaching and clinical analytics
3 - Corporate adoption: Enterprise B2B licensing is poised to rise at 15.37% CAGR through 2030 as employers bundle subscriptions to curb healthcare costs

Source: Grand View Research, 2025

Cal AI soared to over 15 million downloads and $30 million in annual revenue in under two years, with seven full-time employees, built in a classroom. That number deserves a second read, not just for what it says about Cal AI, but for what it says about the new economics of app building.

What made it work wasn’t the technology alone -- OpenAI’s Vision API was accessible to anyone willing to build. What made it work was the founder’s clarity about the job, and the distribution machine he built to scale it.

Cal AI is optimized for what I’d call frictionless entry -- and it executed that positioning nearly perfectly.

Primary intent: Speed-first calorie tracking

Job it does:

“I want to have some sense of what I’m eating without it becoming a part-time job. And… I don’t want to build a system, maintain a streak, or feel guilty every time I skip a day.”

It translates to:
Snap photo ✔️
AI estimates calories and macros ✔️
Log confirmed ✔️
Move on ✔️

Optimised for:
• Fitness-oriented users who already have dietary intuition but want tracking confirmation
• First-time calorie trackers are intimidated by database navigation, Gen Z users with high expectations for AI
• Performance athletes who want macro awareness without logging overhead

The distribution ladder. Yadegari has been unusually candid about Cal AI’s growth engine in public interviews, and the detail is worth unpacking. It was a deliberate progression that hit each ceiling in sequence. He started with fitness influencers -- which got Cal AI to roughly $2 million per month -- then stalled. The fitness influencer audience is finite and heavily overlapping; the marginal impact of the 40th fitness creator collapses quickly. So he expanded to non-fitness creators, then made one of the boldest bets in the story: a $500,000 Mr. Beast sponsorship. Front-end ROI was slightly negative on direct conversions. The brand authority it unlocked -- legitimising the product for tens of millions of viewers, making future creator deals easier to close -- made it profitable in hindsight. Then came paid Meta and TikTok ads, managed in-house after a failed agency stint, layered on top of a creator affiliate program run through Tribe that pays creators a revenue share rather than flat fees.

January 2026 revenue: $5.7 million. ARR at the time of sale: approximately $50 million.

The product detail no one talks about. Cal AI built two versions of its app -- one for regular users, and a parallel version designed exclusively for content creators. The creator version displays the Cal AI brand prominently on every screen and shows a perfect multi-week logging streak regardless of actual usage. It is built to be filmed. When an influencer records a morning routine and holds their phone for four seconds, the app on screen looks like a product people actually commit to -- long streaks, clean UI, prominent logo.

Yadegari described it plainly: it’s the difference between a real hamburger and the hamburger in a McDonald’s ad. The design was not an afterthought. It turned every creator video into a brand moment that appeared completely organic.

The attribution insight. Most founders running multi-channel acquisition have no clean way to separate paid ad revenue from influencer revenue. Yadegari’s team built a partial solution using App Store custom product pages -- one per paid campaign, each with a trackable conversion path isolated from the main listing. Users who see an ad and click through hit a dedicated page; the revenue that page generates is directly attributable to that campaign. Yadegari estimates about 30% of conversions still happen off-attribution -- users who see an ad but then go search the app name manually -- requiring a manual uplift correction.

But the floor is clean enough to make real capital allocation decisions: kill underperforming sets, double down on winners, and stop confusing influencer halo with paid performance.

The onboarding machine. I went through the full Cal AI onboarding, and it’s worth reading screen by screen -- because the conversion psychology is deliberate and layered. The flow opens with a trust-building claim before asking for anything: “80% of Cal AI users maintain their weight loss even 6 months later.” That’s not a feature. That’s an anxiety reducer placed at the exact moment a new user is deciding whether to bother.

Then the quiz begins -- goal, date of birth, desired weight, pace preference, barriers, aspirations, diet type, Apple Health connection. Ten questions. Each one increases the user’s felt investment before any paywall appears. By the time the “Congratulations, your custom plan is ready!” screen lands -- with a specific calorie target, a specific macro breakdown, and a specific goal date -- the user has mentally committed.

The plan feels personalized because the data they gave was personal, even if the underlying logic is a standard BMR formula. The “We’re setting everything up for you — Applying BMR formula...” loading animation at 44% completion makes a spreadsheet calculation feel like a personalized AI build. It’s not deceptive. It’s just good design.

The social proof is strategically embedded again right before the paywall: “90% of users say the change is obvious after using Cal AI.” At that point, the user has a custom plan, a goal date, and two data points telling them this works. Only then does the subscription screen appear -- Monthly €9.99/mo versus Yearly €2.91/mo, with the yearly pre-selected, a “No Payment Due Now” reassurance, and a transparent timeline showing exactly when the reminder and billing will hit. This is a textbook commitment-escalation funnel.

The paywall doesn’t feel like a gate. It feels like the last step in an already-started journey.

Rollover calories. Cal AI asks during setup whether you want to carry unused calories from one day to the next -- up to 200 -- as a credit toward the following day. This is a behaviorally sophisticated design. Traditional calorie apps punish under-eating with nothing: you had a great day, and the app simply resets to zero. Cal AI’s rollover makes discipline feel rewarding and prevents the catastrophic “I had a bad day, I’ve blown it, why bother” churn spiral. Good days compound. Bad days are buffered. The app is subtly teaching users that consistency beats perfection -- which happens to also be the mindset most associated with long-term subscription retention.

Churn watch: Cal AI’s most serious structural risk isn’t accuracy credibility. It’s the calorie targets it recommends during onboarding. At the “Recommended” pace setting, the plan generated for a user trying to lose 2.5 kg gives a daily calorie goal of 532 calories. The “Fast” setting -- which the app allows users to select -- generates targets as low as 201 calories per day.

These are clinically dangerous numbers. The medically accepted minimum for adults is typically 1,200 calories for women and 1,500 for men; below that, the body enters conservation mode, energy crashes, adherence collapses, and the weight typically rebounds. The fast-setting screen does include a small warning -- “Fast loss can cause fatigue or loose skin” -- but the warning sits below a prominently highlighted “You will reach your goal in 12 days” badge in orange. The incentive architecture points toward urgency, not safety. A user who follows a 532-calorie plan will not fail because Cal AI’s photo scanning was wrong.

They will fail because the target itself was unachievable. And when they fail, they will blame themselves, delete the app, and not renew.

Industry annual subscription retention for health and fitness apps sits at approximately 30% -- meaning roughly 70% of annual subscribers don’t renew. Cal AI’s onboarding arguably accelerates that number by setting expectations that its own recommendations cannot deliver. This is the churn risk that matters, and it’s one MFP’s 20 years of science-backed nutritional data infrastructure is uniquely positioned to fix -- if they choose to.

Who acquired Cal AI - and why it matters

MyFitnessPal is the #1 global nutrition and food tracking app, founded in 2005, with 280 million members across 120 countries and a food database spanning 20 million foods, 68,500 brands, and meals from 380+ restaurant chains. That database is a 20-year data moat that no startup can replicate with a Vision API and a weekend.

MFP monitors a competitive suite of some 70 rivals using tools like Sensor Tower -- and when Cal AI started climbing the charts neck-and-neck with MFP in their category in early 2025, outreach started almost immediately. The deal closed in December, marking MFP’s third significant move in just over a year -- following the acquisition of meal planning app Intent and an integration with ChatGPT Health. This is not reactive M&A., but a move made by a company with a competitive intelligence infrastructure running a deliberate portfolio strategy.

Why not merging is the whole strategy

MFP has no plans to integrate Cal AI into its core product or replace its existing meal-scan features. Both apps serve different segments: Cal AI for speed, MFP for nutritional precision.

That decision is the entire thesis.

Most acquirers absorb their targets -- migrate the user base, fold the features, retire the brand. Sometimes that’s right. Here, it would be wrong, and Fisher knows it. The Cal AI user is not an MFP user who hasn’t upgraded yet. They are a structurally different user who has opted against the MFP experience. Merging the products wouldn’t unlock a new audience. It would just make both products worse.

The immediate proof of concept: Cal AI users already have access to MFP’s database post-acquisition -- 20 million foods, 68,500 brands, 380+ restaurant chains -- without the apps merging at all. Cal AI gets more accurate. MFP’s database investment compounds across two surfaces. No brand dilution required.

As Fisher put it: “People come to nutrition with different goals and different levels of experience, and they need tools that meet them where they are.”

The moment Cal AI starts meeting users where MFP is, it stops meeting them where they came.

The six mechanics that built a $50M ARR app in 18 months

1. The distribution ladder, not a single channel

Cal AI didn’t pick a channel and scale it. It moved through channels sequentially as each one hit its ceiling - fitness influencers, then broad creators, then Mr. Beast, then in-house paid performance, then creator affiliates. Each stage built the brand awareness that made the next stage more efficient. Why it prevents churn: users acquired at the top of a warmed-up funnel - people who saw an influencer, then saw an ad, then searched the app name - arrive with higher intent and lower early dropout rates than cold paid installs.

2. Engineering the product to be filmed (the creator-mode app)

A parallel version of Cal AI with a prominent logo on every screen and a perfect fabricated streak turns influencer videos into credibility demonstrations without looking like ads. Product-led distribution design. Why it prevents churn: users who download after seeing the product in use arrive with a more realistic mental model of the experience. Expectation alignment is one of the most underrated early-churn levers in subscription apps.

3. Attribution architecture as a capital allocation tool (App Store custom product pages per campaign)

By isolating paid traffic to campaign-specific App Store pages, Cal AI could distinguish paid performance revenue from organic and influencer baseline - giving the media buyer clean enough signal to kill underperforming creatives and scale winners with confidence. Why it prevents churn: better attribution routes acquisition spend toward higher-intent users, not just higher-volume installs.

4. Database moat as accuracy backstop (MFP’s 20M food database now powering Cal AI)

Post-acquisition, Cal AI’s estimates are grounded in MFP’s two-decade food database. The switching cost for a user is now historical, not technical. Why it prevents churn: accuracy is Cal AI’s single biggest retention variable. Every marginal improvement in calorie estimate credibility extends the window before a user decides the numbers aren’t trustworthy enough to act on.

5. Rollover calories as a behavioral safety net (Cal AI’s carry-forward calorie mechanic)

Cal AI asks during onboarding whether users want to carry unused daily calories forward - up to 200 - as credit toward the following day. This design choice is easy to miss and worth studying closely. Most calorie apps treat a day’s end as a hard reset: under-eat today and the app gives you nothing for it tomorrow. Cal AI’s rollover makes discipline feel cumulative. More importantly, it softens the psychology of imperfect days - the “I had a bad day, I’ve blown it, might as well quit” spiral that drives a measurable percentage of early churn in every habit-based subscription app. Why it prevents churn: the product implicitly teaches that consistency beats perfection, which happens to be the mindset most correlated with long-term subscription retention. It’s a small feature with an outsized behavioral effect.

6. Speed as an operating system, not a value statement

In interviews, Yadegari described speed as Cal AI’s number-one company value - operationally, not aspirationally. His practice: identify bottlenecks in real time, get the blocking party on a call immediately, resolve on the call. No waiting. No process theater. For a seven-person team competing against a 20-year incumbent, cycle time is the only structural advantage that can’t be bought. Why it prevents churn: features that reduce friction get shipped faster. Bugs that erode trust get fixed faster. The product improves at the rate needed to keep early adopters engaged long enough to become habitual users.

Key takeaways

• Nutrition apps no longer compete on database size or feature depth -- they compete on how much effort they remove from the act of logging, and how well the commitment architecture in onboarding converts investment into subscription.

• Cal AI’s onboarding is a masterclass in commitment escalation: ten personalized questions, embedded social proof at two anxiety peaks, and a paywall that appears only after the user already has a custom plan, a macro target, and a goal date. By the time billing is introduced, the decision has already been made.

• The rollover calories mechanic is the most underappreciated retention design in the product -- it rewires the “I had a bad day, I’ve blown it” churn trigger into a compounding reward for consistency.

• Cal AI’s most serious structural risk isn’t photo scan accuracy -- it’s the calorie targets its own onboarding recommends. A 752 cal/day “recommended” plan isn’t sustainable; a 201 cal/day “fast” plan borders on dangerous. Users who follow the plan and fail won’t blame the plan. They’ll delete the app. MFP’s science-backed nutritional infrastructure is the most strategically valuable thing in this acquisition — not for the database, but as a guardrail for what Cal AI’s algorithm actually prescribes.

• Retention follows effort alignment; churn begins when the product demands more from the user than the user signed up to give — and accelerates when the plan the product designs is physically impossible to sustain.

Written right after an alfahores tasting in Buenos Aires, Argentina

If this one resonated, these are worth reading next:

✨ From swipe fatigue to IRL: Inside the apps racing to get you off your screen How Hinge, Timeleft, and Thursday built apps that optimize for outcomes rather than session length - the same shift Cal AI is attempting in nutrition.

🤍 Search to Swipe: what three decades of Dating Apps reveal about customer experience and churn How consumer apps built for engagement accidentally trained users to disengage - the longer version of the same argument this article makes.

📱Is ‘app-of-apps’ becoming the new interface for work and life? If the standalone brand strategy in this article caught your attention, this issue maps exactly how that portfolio model creates retention - and where it creates fragmentation risk.

You just finished your morning run, and before you’ve even caught your breath, you’ve already shared your route, collected three kudos from people you’ve trained alongside for two years, and quietly noted that someone edged your segment time by four seconds.

Hard to imagine doing any of this on a different app. Everything you’ve built lives inside Strava. Moving would mean starting over -- and starting over means losing all the progress the app stores.

In the early 2010s, building an app was genuinely hard - team, budget, months of iteration, App Store approval that takes forever, and feels like a lottery. The teams that shipped had a structural advantage just for existing.

Habit tracking was still largely analogue. When apps moved in, they brought real innovation: the digital streak, the gamified check-in, the progress ring. The first habit tracker you downloaded felt genuinely new.

Then the market matured, features got copied, and designs converged. By 2023, you could barely distinguish one habit tracker from another in the App Store carousel -- because anything one team could build, another team could replicate.

The shift happened when the cost of building apps didn’t just fall - it basically disappeared.

In February 2025, Andrej Karpathy named it: vibe coding. Describe what you want, and AI generates a working app in hours. By March 2025, Y Combinator reported 25% of its startups had built 95% of their codebases with AI. By the end of 2025, 557,000 new apps were submitted to the Apple App Store -- a 24% surge, portraying the fastest growth since 2016.

The conversation in founder communities flipped almost overnight. Building stopped being the hard part. Keeping users became the hard part -- and almost nobody had a good answer to it.

If anyone can ship your product by Tuesday, your features are a starting point. The only durable advantage left is what a competitor can’t copy: what the user has already built inside your product.

Feature advantage has collapsed as a retention mechanism in consumer apps.

Duolingo, Strava, and Daylio are three different answers to the same question: what would it cost a user to leave? Streaks -- the beautifully designed iOS habit tracker -- sits as the cautionary contrast: a product with no compelling answer to that question.

Duolingo is optimised for what I’d call commitment-compounding utility -- the clearest example in consumer apps of a product that has systematically engineered the cost of leaving. It’s also, as of 2025, the clearest example of what happens when an identity moat runs into the limits of the identity it’s selling.

Primary intent: Identity formation through daily ritual

Job it does:

“Help me become the kind of person who speaks another language -- and make sure I don’t quit when motivation drops.”

It translates to: Open app ✔ > Complete daily lesson ✔ > Protect streak ✔ > Ascend league ✔ > Accumulate XP history

Duolingo intentionally ignores linguistic depth. Its lessons are short, gamelike, and emotionally engineered rather than pedagogically rigorous. It treats showing up as the product. Once a user has maintained a 200-day streak, the streak itself has become the product -- regardless of whether the lessons are any good.

The mechanics behind Duolingo's retention are exceptionally well-documented. ’s piece in ’s newsletter -- 'How Duolingo reignited user growth'-- is the definitive breakdown of how the streak was rebuilt from near-zero into the engine that drove 4.5x DAU growth in four years.^[6] Over four years of streak optimisation alone, daily churn of best users fell by over 40%. A single copy change -- swapping 'continue' for 'commit to my goal' -- drove 10,000 incremental DAUs

The gamification engine

Around 2017–2019, Duolingo was bleeding users and sitting at $13M in revenue. The team asked a specific question: which app category had retention rates that education products couldn't touch? Mobile gaming. Angry Birds. Clash Royale. Candy Crush. Games built on the same daily habit loop Duolingo needed --mechanics refined over billions of player-hours to make quitting feel costly. Duolingo borrowed the playbook wholesale.

By 2020, revenue had reached $161M -- a 12x increase in three years, almost entirely from the gamification overhaul. The course content hadn’t changed. The psychological infrastructure had.

Social media virality

In 2020, a 23-year-old recent graduate named spotted a dusty owl costume in the Duolingo office and asked if she could make videos. "I don't need money. I have a crusty owl suit and an iPhone." The C-suite was skeptical about TikTok. She launched anyway. Five years later: 50,000 followers to 16 million, 8 billion impressions across social, and one of the most documented brand strategies in marketing history.

The core insight: Duo should behave like a creator, not a brand. On TikTok, Duolingo wasn't competing against other education apps -- it was competing against everything else in the scroll. Parvez treated the comment section as the content brief. Recurring storylines built Duo into a character with actual lore: unhinged lesson reminders that tip into threats, the Dua Lipa obsession running since 2021, the feud with Google Translate, the legal team as a straight-man foil. The account became satire of the creator economy.

Then: Duo's death on February 11, 2025. Six days from concept to execution.
App icon tweak → Cybertruck murder mystery → collective XP goal to resurrect him.
Social mentions spiked 25,560% in a day.450 articles in 60% top-tier outlets.
Dua Lipa's response drove 667,000 engagement actions. The campaign out-mentioned every Super Bowl ad that week. Duo came back on February 24. The XP goal was hit. Over 5 billion lessons completed in two weeks -- because the resurrection mission was also a re-engagement campaign with an in-app mechanic built into it. The push notification, the streak system, and the viral moment all fired simultaneously. This is what it looks like when top-of-funnel virality is wired directly to the gamification loop.

AI Backlash

In May 2025 DUOL peaked at $544, ended the year down 46%, and was trading around $112 by February 2026 -- investors spooked by soft guidance, rising AI hosting costs, and one uncomfortable question the memo made impossible to ignore. Open ChatGPT and you get more speaking practice in ten minutes than a Duolingo lesson delivers all week. No streak, no subscription, no owl. The content moat Duolingo spent 12 years building was replicated in volume in under a year. What’s left is the behavioural moat -- the streak, the identity, the history. Reassuring because no LLM can clone a 900-day streak. Fragile because the AI-first pivot attacked that moat from the inside.

The research adds a second layer of pressure. On reading, vocabulary, and listening, Duolingo holds up -- learners hit comparable proficiency to one or two university semesters in roughly half the time. On speaking fluency, a 2021 systematic review found it not consistently effective. The structural problem is that Duolingo’s North Star metric is current user retention rate -- how many of yesterday’s users came back today. Excellent for retention engineering. A terrible proxy for whether anyone is actually learning. A user can hold a 900-day streak and still not hold a basic conversation. When that realisation lands -- and for motivated learners, it does -- the streak stops protecting against churn and starts looking like evidence of wasted time.

Churn watch:

Duolingo’s churn has two layers. Early churn catches the 52% who leave within 30 days before the streak builds real switching cost.
Post-realisation churn is more dangerous: users who’ve invested months or years and clock the gap between Duolingo’s retention metrics and their actual language competence. The AI-first backlash accelerated the second layer -- users who already had doubts about learning effectiveness then watched the content get handed to machines that are not always consistent with quality and make mistakes.

Closing that gap doesn’t require more features. The streak needs to remain a proxy for genuine progress, not a substitute for it: keep users coming back and give them something real to come back for.

Strava understood something most fitness apps missed: workouts aren’t the output. The permanent, social record of who you physically are is. Once you have five years of data, 300 followers, and a segment KOM you’ve defended through two winters -- leaving isn’t a product decision. It’s a simultaneous loss of your data, your community, and your athletic identity.

Primary intent: Athletic identity, community accountability, performance history

Job it does:

“Track my training, show my progress to people who care about the same things I do, and let me prove to myself - and occasionally others - that I’m the athlete I think I am.”

It translates to: Log activity ✔ > Auto-analyse performance ✔ > Share to followers ✔ > Receive kudos ✔ > Compete on segments ✔ > Accumulate years of personal athletic history

Strava has the deepest irreplaceability architecture in this category: a personal data archive that grows with every workout, a social graph with historical context, and a competitive layer -- segments, clubs, leaderboards -- that only carries meaning inside Strava. One data point tells the whole story: revenue grew from $110M in 2022 to $163M in 2024 even as quarterly downloads fell 16% in Q4 2024. Revenue up, acquisition down -- the signature of genuine lock-in. When Strava restricted its API in late 2024, the community reaction was visceral. Touching the perceived ownership of their data hit harder than any feature removal.

Users who feel they’ve built something inside the product that belongs to them don’t tend to leave.

Churn watch:

Strava’s churn risk is almost entirely self-inflicted. No competitor is close to replicating its social graph. The danger is internal: paywalling previously free features, restricting integrations, and compressing the free tier. When the moat depends on the community, the product team is the biggest churn risk on the board. The winning strategy is stewardship, not roadmap expansion.

Daylio is a minimalist micro-journal and mood tracker designed to build a longitudinal personal record of how your life actually feels. No coaching, no community, no wearable integrations.

It tries to be the most honest mirror of your inner life -- one tap at a time.

Primary intent: Self-knowledge through accumulated emotional data

Irreplaceability source: Months of mood patterns, habit correlations, and annotated life events -- nowhere else

Job it closes:

“Help me understand my own patterns. Why some weeks feel better than others, what actually affects my mood, and what I was feeling this time last year.”

It translates to:

• Log mood ✔ > Tag activities ✔ > Write micro-note ✔ > Surface patterns over time ✔ > Build a personal emotional dataset that grows more useful every month

Daylio intentionally ignores everything outside your own emotional data. No social layer, no community, no coaching. It treats your inner life as the only audience worth serving -- and makes the accumulated data beautiful enough that revisiting it becomes a ritual.

A solo founder could build a Daylio clone by Friday. Minimal design, simple mechanics, low technical complexity. By every measure that defined product defensibility before 2025, it should be trivially replaceable. And yet it isn’t — because for a user with 1,200 days of mood entries, Daylio isn’t an app, it’s a personal archive. Switching means losing a record of their inner life that exists nowhere else and can’t be reconstructed.

Most apps lose their novelty value as motivation fades. Daylio appreciates -- every logged day adds to an archive that becomes more meaningful precisely because it’s older. That’s a fundamentally different retention dynamic than anything a feature list creates.

Churn watch:

Daylio’s churn risk is concentrated in the first 90 days -- before data accumulation creates switching costs. The lever most underused: surface patterns at Day 30, not after months. “Your mood was highest on weeks you exercised three or more times.” “This time last month you rated Tuesday as your worst day.” The data already knows. Show it sooner.

Streaks scores on habit architecture and progress visibility -- genuinely good at making the daily ritual automatic and showing what you’d lose. But those are the two mechanics any AI tool replicates by Tuesday. A Lovable clone ships with a streak counter on day one.

Strava’s near-clean sweep is what genuine lock-in looks like. When you score high across five of six mechanics, users don’t leave even when acquisition stalls, as the product has too many roots.

Duolingo’s split score tells the 2025 story exactly -- strong on the behavioural layer, weaker on data accumulation and content irreplaceability. The moat is real and concentrated. That’s both its strength and its exposure to LLM substitution.

Daylio’s high scores on data accumulation and content irreplaceability, with a moderate score on habit architecture explains the churn profile: great long-term retention once the archive kicks in, vulnerable in the first 90 days before it does.

The moat isn’t what you built. It’s what your user built inside your product. And in 2026, that’s the only kind of moat that can’t be vibe-coded away by Tuesday.

In habit tracking, retention grows when the app holds something the user has invested in — not something the app built for them. A streak matters only if it represents a self worth protecting. A dataset matters only if it reflects a life worth understanding. A community matters only if it contains people the user would genuinely miss.

Key Takeaways

• Features are a starting point, not a moat. If anyone can clone your product by Tuesday, your differentiation lives somewhere else — or it doesn’t exist.

• Identity moats have a shelf life. Duolingo’s streak works until users realise the streak outlasted the actual learning. DUOL fell 46% in a year finding that out.

• Data accumulation is the slowest moat to build and the hardest to break. Daylio on Day 1 is a widget. On Day 1,000 it’s a personal document. The switching cost isn’t losing an app — it’s losing a record.

• Communities don’t churn alone. Strava’s revenue grew while downloads fell. Leaving requires convincing everyone else to leave, too.

Score your own product on the rubric. If you’re only hitting habit architecture and progress visibility, a well-prompted founder is your most dangerous competitor.

Written from a casita in Antigua, Guatemala, overlooking three volcano peaks - Volcan de Agua, Acatenango, and Fuego.

You probably have something on your phone right now that stops you from opening Instagram mid-task, or plays ambient sound that helps you concentrate, or shows you exactly how distracted you were yesterday. Hard to imagine we ever worked without that kind of environmental support. Today, it’s even harder to imagine trying to write, code, or think deeply without architecture that protects attention before the phone even lights up.

That wasn’t always the case.

In the 2000s, distraction was asynchronous -- email piled up in your inbox, instant messages waited for replies, but nothing grabbed your attention in the moment. You could walk away from the computer. Focus was a question of discipline, not design. The fight was internal: you versus your own impulse to check.

Later on, smartphones arrived. Push notifications normalised interruption. By 2018, the average person received 150 notifications per day and unlocked their phone 85 times daily. Responsibility for focus shifted from self-control to environmental management. Distraction became ambient, architectural, continuous. The device sat in your pocket, radiating possibility and provoking curiosity.

The velocity increased. Social media feeds replaced chronological timelines with algorithmic ones optimised for engagement, not closure. Infinite scroll became standard. TikTok shortened videos to under 15 seconds. By 2025, the average adult spend over 7 hours daily on screens. Attention spans stabilised at 40-47 seconds, and instead of fighting it, people accepted it as the new normal.

The real shift happened with the recognition that willpower had structurally failed.

For decades, focus tools evolved -- from simple timers to website blockers to gamified productivity. On the surface, this looks like a feature upgrade.

In reality, it reflects something more structural:

The collapse of sustained attention as the default state.

Market at a glance:

The ADHD and focus app market is scaling rapidly -- from $2.08B in 2024 to ~$7.7B by 2033, growing at a ~15% CAGR, driven by rising ADHD diagnoses (3-5% annual increase), remote work fragmentation, attention crisis among young adults (82% report concentration difficulties with social media), and smartphone ubiquity (5+ hours daily use). The broader gravity includes digital wellness, mental health apps, and workplace productivity tools, making this a structurally expanding category rather than a niche productivity hack.

User base: 11M+ Americans use ADHD apps regularly; 59% of adults use them specifically for work productivity.

Key market drivers:
- Rising ADHD prevalence: 3-5% annual diagnosis increase, adult recognition growing
- Attention crisis: 82% of young adults report concentration difficulties with social media
- Remote work: 46% of adults with ADHD use apps for work tasks, no physical boundaries

Focus apps replaced distraction, then temptation, and now they’re replacing the cognitive architecture that made deep work possible.

Endel, Minimalist Phone, and Opal are not variations of the same product. They are answers to different breakdowns in today’s attention.

Endel is optimised for what I’d call ambient cognitive support and does its job extremely well.

Primary intent: Environmental optimisation

Job it does:

“My environment is too chaotic to focus. I need something that creates a focus-conducive atmosphere without me having to think about it -- something that just works in the background while I’m trying to work.”

It translates to: Sense environment ✔️ > Generate adaptive soundscape ✔️ > Modulate in real-time ✔️ > Sustain over extended sessions ✔️

Optimised for: • Ambient focus support during work sessions • Passive integration into daily routines (runs like Spotify) • Habit formation through consistency • Extended deep work periods (2-4+ hours)

Endel intentionally leads with support, not restriction. While it includes an app blocker mode, this is positioned as supplementary infrastructure -- the soundscape is the core product. It treats attention as something that needs environmental optimisation first, with enforcement available when needed.

Best users: • Knowledge workers who work in open offices or noisy environments • People with ADHD who need sensory regulation • Creatives, writers, developers in deep work mode • Anyone who wants focus support that feels effortless

Endel nailed passive integration into attention architecture. Unlike blockers that require active decision-making (”Should I enable this now?”), Endel works like environmental design -- you turn it on once, and it becomes part of your focus routine. The AI adapts soundscapes based on time of day, heart rate (via Apple Watch), weather, and circadian rhythm, creating a personalised ambient cocoon that masks distracting noise while providing non-intrusive auditory stimulation.

The neuroscience is solid: 7x focus improvement demonstrated in peer-reviewed studies. The mechanism works through auditory masking (covering unpredictable environmental sounds) while maintaining consistent, non-linguistic audio that doesn’t hijack verbal working memory. For ADHD brains specifically, this addresses the challenge of sensory gating -- the prefrontal cortex’s reduced ability to filter irrelevant stimuli.

The app blocker feature (added later in product evolution) serves as complementary infrastructure -- it removes temptation while the soundscape supports focus. This is strategic: blocking alone creates friction, but blocking while experiencing focus-inducing audio feels like removing obstacles from an already-working system.

What makes Endel particularly strong is its habituation strategy. After 2-3 weeks, users associate the soundscape with focus states. It becomes an internal trigger: open Endel → brain enters work mode. This is retention through neural conditioning, not feature addiction.

The product also benefits from low perceived effort. There’s no setup, no configuration anxiety, no “am I doing this right?” Users who struggle with complex productivity systems find Endel refreshingly simple: press play, work happens.

Churn watch:

Endel does not need feature expansion to reduce churn. Instead, it should defend simplicity and deepen the habit loop.

When retention depends on ambient presence, the winning strategy is not adding tools -- it’s becoming invisible infrastructure.

The churn risk emerges when:

1. Habituation plateau: After 3-6 months, the soundscape becomes “wallpaper” -- users stop noticing it, which can make them question if they still need it. Enough algorithmic variation could make soundscapes not feel stale, but not so much that they break the habit trigger.

2. Value attribution problem: Because it’s passive, users may not realize how much Endel is helping until they stop using it. Usage insights surfacing value would help to communicate it to the user: “You focused for 47 hours this month with Endel -- 2.3x your average without it”.

3. Competitive displacement: Spotify/Apple Music playlists feel “good enough” and are already paid for

The strategic insight: Endel should behave like Spotify for focus -- something you never consciously choose because it’s already running.

A lot of you guys reading this article don’t know me.

Hi, I’m Daria Littlefield. I’ve spent the last decade leading customer operations across a $400M ARR portfolio - 35+ apps in dating and social space.

The kind of apps people open 50 times a day, then delete in a rage, then reluctantly reinstall a week later.

Do Not Churn is where I break down the products we all use and show you exactly what makes users stay or leave. Why do some apps hit $100M ARR in under a year while others bleed users before launch? The retention mechanisms you use every day but never notice. The difference between a feature and a dependency.

I answer every comment here.
So tell me: What’s one app that became essential without you realizing it?

Drop your answer below. Or just say hi -- I’m trying to actually know who reads the stuff I write.

Leave a comment

Minimalist Phone is a radical architectural intervention built for people who realize their attention is hijacked before they make a conscious choice. Minimalist Phone nailed that apps are not the problem, but the actual visual interface triggers are.

Job it does:

“My phone is designed to grab my attention constantly. Colorful icons, red notification badges, infinite feeds - I want to use my phone as a tool, not get trapped by it. I need the interface itself to stop triggering me.”

It translates to: Replace icon grid ✔️ > Enable text-based launcher ✔️ > Add grayscale mode ✔️ > Remove visual triggers ✔️ > Reduce cognitive pull ✔️

Best users: • People who compulsively unlock their phone without knowing why
• Those who’ve tried blockers but still get sucked in by the home screen • ADHD users sensitive to visual stimuli • Anyone who wants friction before they open an app, not after

While tools like Opal block apps after you try to open them, Minimalist Phone prevents you from wanting to open them in the first place.

Minimalist Phone has a permanent architectural effect on interface conditioning. That’s why its churn profile looks fundamentally different.

The core insight is neuroscientific: visual pop-out effects in app icons automatically capture attention through the ventral attention network, which operates pre-consciously. Your brain registers the red badge, the bright colours, the spatial memory of “Instagram lives here” before you’ve decided to open anything. By the time conscious choice arrives, you’re already halfway through the unlock-swipe-tap sequence.

Minimalist Phone attacks this at the root. The text-based launcher removes visual salience entirely. Instead of a grid of colourful icons, you see a list of text labels.

The grayscale mode compounds this. Without colour, dopamine prediction errors (the “ooh, something new!” response) are dampened. Your phone becomes a tool, not a stimulus.

Minimalist Phone has clearly nailed prevention-at-interface-level and friction-as-feature.

The retention mechanism is particularly interesting: after 3 days, 93% of users reduce screen time. But more importantly, after 2-3 weeks, returning to a normal launcher feels overwhelming. This creates neural rewiring as a retention moat. Churning from Minimalist Phone means going back to an interface that now feels wrong.

Churn watch:

Where churn can emerge is during the first 72 hours, when the interface still feels unfamiliar and inconvenient.

A key friction point is legitimate use cases for visual interfaces -- maps, photos, and design apps that need colour and visual information. Grayscale mode makes these harder to use, and users can feel penalised for non-problematic phone use.

The strategic insight: The first 72 hours should feel like an upgrade, not a downgrade. If users push through that window, habituation locks them in. Leveraging progressive adoption by starting with the text launcher only, and adding grayscale gradually as the user adapts. Positioning this as “becoming someone who uses their phone consciously” rather than “restricting yourself” will reinforce identity shift.

Opal is a hard enforcement tool designed to stop you when self-regulation has already failed. It doesn’t try to support you, it doesn’t try to prevent temptation -- it tries to be stronger than your impulses.

Job it does:

“I know exactly what I shouldn’t be doing -- scrolling TikTok during work, checking Instagram before bed, getting lost in YouTube. I’ve tried willpower, and it doesn’t work. I need something that physically stops me and shows me how bad the problem actually is.”

It translates to: Track all attempts ✔️ > Generate Focus Score ✔️ > Compare to peer benchmarks ✔️ > Create accountability ✔️

Best users: • People who’ve failed with softer interventions • Those who want data on their actual behaviour, not their intended behaviour • Anyone who needs external enforcement because internal discipline is exhausted

Opal nailed hard enforcement with data-driven accountability -- blocking apps through Screen Time API so there’s no easy override, tracking every unlock and app-opening attempt, and translating behaviour into a Focus Score that users can compare against peers.

This is a behaviourally aligned product.

The strength is in making invisible behaviour visible. Most people underestimate how often they reach for their phone. Opal shows you attempted unlocks, time spent per app, and Focus Score trends over weeks. The leaderboard feature adds social proof: “You’re less distracted than 76% of Opal users” becomes a metric worth protecting.

The 430K+ user base creates network effects -- the more people on Opal, the more meaningful peer benchmarking becomes. With a reported 94% reduction in distraction and an estimated $350-450K MRR, Opal has found product-market fit with people who need discipline-as-a-service.

Churn watch:

The churn mechanism is score ceiling and loss of novelty:

1. Months 1-3: Focus Score improves dramatically (45 → 78), users feel transformed, app feels essential

2. Months 4-6: Score plateaus (78 → 81), improvements get smaller, dopamine from “getting better” fades

3. Months 7+: Score stabilises, app becomes invisible infrastructure, users question if they still need it

The paradox: When Opal works perfectly, users feel like they’ve fixed the problem and no longer need help.

Over time, users may feel that it helped them learn discipline, but they’ve internalised it and no longer need external enforcement.

The strategic insight: Opal should behave like a gym membership for attention -- you might not go every day, but canceling it means admitting you’re okay sliding back. Long-term trend insights, “Your focus has improved 34% since January -- here’s what changed,” can surface value beyond daily scores.

The six retention mechanics

1. Baseline shift lock-in

After 2-3 weeks of use, the user’s perception of “normal” fundamentally changes. Minimalist Phone users find colourful launchers overwhelming. Endel users can’t work in silence anymore. The product doesn’t just solve the problem -- it rewires what feels acceptable. Churning means returning to a state that now feels broken.

The goal isn’t to fix the user temporarily -- it’s to make them unable to tolerate unfixed conditions.

2. Data archaeology value

Historical usage data becomes increasingly valuable over time, creating sunk costs. Opal’s 6-month Focus Score trend. Endel’s cumulative hours in flow state. The longer you use it, the more painful it is to lose that history. It’s not just abandoning a tool -- it’s abandoning proof of your progress.

3. Identity formation over behavior modification

Users transition from “I’m trying to use my phone less” to “I’m someone who has a minimalist phone.” From “I’m using Endel to focus” to “I’m an Endel person.” The product becomes part of identity, not a temporary fix. Churning feels like losing part of who you’ve become. Marketing and UX should reinforce the identity language. “Join 430K people taking control” (Opal) works better than “Try blocking apps.”

4. Platform entanglement

Cross-device sync, wearable integration, and ecosystem presence create compounding switching costs. Endel on iPhone + Apple Watch + Mac means churning requires finding three replacements. Opal’s VPN configuration across devices makes switching technically annoying. The more integration points, the higher the exit friction.
Race to occupy multiple surfaces in the user’s tech stack. Calendar integration, desktop apps, browser extensions, smart home integration -- each adds a removal barrier. The question shifts from “Should I keep this?” to “Is it worth the hassle of uninstalling everywhere?”

5. Comparative anxiety retention

Users fear regression more than they value improvement. “What if I go back to 7 hours daily?” becomes a retention moat. The product isn’t just providing value -- it’s preventing feared outcomes. Fear of baseline regression (losing gained ground) is stronger than the desire for marginal improvement. Occasionally remind users of their “before” state. “Remember when you averaged 142 phone unlocks daily? You’re now at 48.” Don’t make it guilt-inducing -- make it empowering proof that regression is possible and they’ve escaped something they don’t want to return to.

6. Gamified progress metrics that create their own goal

Opal’s Focus Score doesn’t measure an external outcome (productivity, deep work completed) - it measures itself. Users arrive wanting to “use their phone less,” but within weeks, the goal shifts to “improve my Focus Score.” The metric becomes the objective, not a proxy for the objective.

This creates a self-perpetuating retention loop: the score only exists within Opal, so maintaining it requires staying subscribed. Unlike external goals (”finish my novel,” “launch the product”), which can be achieved and completed, the Focus Score is infinite -- there’s always room to improve from 78 to 81, to introduce streaks, to beat last month’s average.

Create proprietary metrics that only exist within your product. External goals can be achieved elsewhere; internal scores lock users into your ecosystem. The score should be infinite (no ceiling), comparable (leaderboards), and dynamic (adaptive targets prevent plateau).

Key takeaways

• Focus apps compete on primary intervention timing, not feature presence. Products may offer blocking, soundscapes, and analytics -- but the question is “what do users come for?”

• Ambient support tools (Endel) succeed through habit formation and passive presence. Retention comes from neural conditioning, not feature depth. The product should become unconscious infrastructure.

• Architectural prevention tools (Minimalist Phone) retain users through interface adaptation. After 2-3 weeks, the brain rewires to prefer low-stimulation interfaces. Churning means returning to an environment that now feels overstimulating.

• Hard enforcement tools (Opal) face a success paradox: when they work perfectly, users feel they’ve internalised discipline and no longer need help. Retention requires shifting from “improvement score” to “maintained streak” and expanding the enforcement scope beyond the initial use case.

• Retention follows primary intent alignment; churn begins when users feel the product solved a problem they no longer have -- even if the problem would return without the product. Products survive by either (a) making their value attribution visible, or (b) becoming so integrated that removal feels like losing a limb.

  ---

  Written in my head during a 10-hour ascent of Fuego, an active volcano outside Antigua, Guatemala, and written down shortly after.

The math stopped working somewhere around 2022.

Users were swiping more - averaging 90 minutes per week on Hinge alone - but matching less. Conversations were starting (barely) but dying in under four messages. Apps kept adding features: video prompts, voice notes, AI-powered icebreakers. Engagement dashboards looked healthy. Revenue was up. But in focus groups, the same phrase kept appearing: “I’m exhausted.”

In my previous analysis of dating app evolution, I traced how we got here - how swipe-based mechanics accidentally trained users to treat dating as lightweight consumption rather than intentional connection. The core finding: when effort doesn’t visibly compound toward outcomes, users slowly disengage.

The disconnect driving this shift isn’t subtle. According to Forbes Health Survey, 76% of Gen Z dating app users reported feeling emotionally, mentally, or physically drained from the platforms designed to help them connect. Meanwhile, the WHO declared loneliness a public health crisis carrying the same mortality risk as smoking 15 cigarettes per day. The US Surgeon General called it an “epidemic.” In the UK, 67% of university students living in halls reported feeling isolated - despite being separated from peers by paper-thin walls.

As writes in her essay on the loneliness epidemic, “It’s never been easier to get what you want and harder to figure out if you actually wanted it in the first place.”

Dating apps had delivered on their original promise: access. But somewhere between infinite choice and meaningful connection, the product broke.

Now, a new category is emerging - one that flips the core assumption. Instead of maximizing time-on-app, these platforms measure success by how quickly they get you off the app and into real space with real people.

Zaman calls it JoyTech: technology designed not to capture attention, but to catalyze presence.

Market at a glance: The JoyTech/IRL Social Connection market is an emerging category without formal market research from established firms. The 2025 baseline ($420M) aggregates revenue from Timeleft (€18M ARR, founder-stated), Thursday ($8-12M estimated from event attendance), and 15+ emerging platforms.
Growth scenarios derive from: (1) Timeleft's demonstrated 20-month trajectory (0→€18M), (2) comparable category growth patterns (dating apps, food delivery, rideshare during early phases), and (3) TAM validation against 300M people globally with zero close friends (CU Independent study). Base case mirrors early dating app adoption; bullish scenario assumes multiple accelerants (corporate wellness, institutional funding, Asia-Pacific expansion).

If anything, it’s not nostalgia for pre-digital dating, but a recognition that the hard problem was never “how do I see more profiles?”, but “how do I spend my limited time with people who matter?”- a coordination problem at scale that digital matching alone cannot solve.

The three platforms profiled here represent different answers to that question - and early proof that the post-swipe era is already generating millions in revenue.

Hinge launched in 2012 as a Facebook-connected dating app where you could list which friends you had crushes on. Cool idea, but idn’t scale. By 2015, it looked like every other swipe app - and Nancy Jo Sales’ Vanity Fair piece “Tinder and the Dawn of the Dating Apocalypse” put a target on the entire category.

Founder Justin McLeod took Hinge off the app store entirely to rebuild. When it relaunched in 2016, the tagline was a dare: “The dating app designed to be deleted.”

It wasn't just a marketing pivot. The product architecture changed. Limited daily likes forced intentionality. Profile prompts ("Two truths and a lie," "A life goal of mine") gave conversation starters that went deeper than "Hey." The "We Met" feature explicitly tracked whether matches led to actual dates - then used that feedback to improve future recommendations.

Today: 30 million users, $550 million in 2024 revenue (up 38% year-over-year), #3 in the US (18.75% market share, behind Tinder and Bumble but fastest-growing of the big three), and #1 in the UK, Ireland, and the Nordics (Business of Apps).

Primary intent: Serious relationships

Job it does:

"I'm tired of conversations that go nowhere. I want to know if we're actually compatible before investing an hour getting ready for a date that crashes in the first five minutes. And if this works... I want to be done with apps entirely."

It translates to:

✔️ Signal over noise (profile prompts reveal personality, not just photos)
✔️ Friction as filter (limited likes = users take each match seriously)
✔️ Outcome tracking (We Met feature closes the feedback loop)
✔️ Proof of success (35% of couples married via dating apps in 2023 met on Hinge - beating Tinder and Bumble)

Best user: •Deliberate romantics who read profiles, not just photos • Efficiency maximizers tired of texting-for-weeks-then-ghosting • Users who want algorithms to do the filtering so conversations can start deeper

In May 2023, and his co-founders were nearly out of money and ideas. They’d pivoted three times - social bucket lists, event discovery, networking for remote workers - and nothing stuck.

“One last try,” Barbier told his team. What if they stripped away everything - no profiles, no photos, no DMs - and just got people to show up for dinner?

They built a Typeform quiz, a WhatsApp group, and a Stripe payment link. No app. No code. If five strangers paid €10 each to have dinner together on Wednesday night, Timeleft would book a restaurant table and send the address that morning.

The first dinners sold out in hours.

Today, Timeleft operates in 200+ cities across 50 countries, connects 150,000 people per month, and generates €18 million in annual recurring revenue.

(check out this outstanding article breaking down Timeleft growth in more detail by )

Primary Intent: Make friends

"I want to meet people who might become real friends, not just matches. I want a low-pressure conversation with a built-in structure so it's not awkward. I want to try this once without committing to weekly therapy sessions or expensive hobby classes. And if it works - I want it to be a standing Wednesday thing I don't have to coordinate myself."

As observed in her loneliness essay: “The hardest part is rarely meeting someone brilliant. It’s spending enough time near them to notice that they are. To see the quiet ways their presence could change your life.”

Timeleft’s insight was that a three-hour dinner forces that time investment upfront, in a group setting where the stakes are lower but the exposure is deeper.

It translates to:

✔️ Group dynamics over 1-on-1 pressure (6 strangers = diffused attention, easier to be yourself)
✔️ Personality matching without photos (algorithm uses Big Five traits, age, language, budget - optimizing for conversation chemistry)
✔️ Structured spontaneity (you book, they handle logistics, you just show up)
✔️ Repeatability (weekly cadence builds habit; €10-15/week sits below "night out" threshold but above "commodity subscription")

Best user:
• Deliberate romantics who read profiles, not just photos
• Efficiency maximizers tired of texting-for-weeks-then-ghosting
• Users who want algorithms to do the filtering so conversations can start deeper

Churn watch: The risk metric isn’t first-dinner no-shows (those are low at ~8%). It’s second-dinner attendance.

If users don’t return within 3 weeks, they likely never will - the experience was fine but not habit-forming. Timeleft needs 50%+ of first-timers to book again within 30 days to sustain growth.

Another watch-out: restaurant dependency. If quality control slips - slow service, bad food, noisy venue - it worsens the entire experience. Timeleft can’t afford more than a 5-10% “bad dinner” rate or word-of-mouth flips negative.

If retention holds, lifetime value is massive. A user who attends 2 dinners/month for 18 months = €360 revenue per person at near-zero marginal cost.

Thursday launched in 2021 with a provocative constraint: the app only worked on Thursdays.

Your matches disappeared at midnight. Conversation threads vanished. The idea was to create urgency - stop endless chatting, start an actual meeting. Investors loved it. Early adopters thought it was clever. Within 24 months, Thursday hit 2 million users and raised £2.5 million in seed funding.

Then in January 2025, co-founders George Rawlings and Matthew McNeill Love did something unthinkable. They shut down the app entirely.

They realized the app was the path, not the destination.

Thursday now operates as a pure IRL events platform across 75+ cities globally (expanding to 150+ by mid-2026). No swiping. No matching algorithm. Just curated bar nights, run clubs, dinner parties, and international trips where everyone attending is verified single and looking to meet people.

Last year’s numbers: Thousands of weekly attendees. Hundreds of brand partnerships (Femfresh, Bolt, Extra Gum - campaigns reaching 500K+ impressions each). And according to their own count: 100+ marriages and 29 babies from Thursday connections.

Primary Intent: Get on an IRL date

Job it does:

"I want to walk into a bar knowing everyone there is single and open to meeting people - so I'm not guessing whether someone's interested or just being polite. I want the 'first date' to be a fun event, not a coffee interview. And I don't want to coordinate logistics - just tell me where and when, I'll show up."

It Translates To:

✔️ Pre-screened environments
✔️ Event variety (bar nights, running clubs, pottery classes, ski trips - different contexts for different personalities)
✔️ Built-in icebreakers
✔️ Community over matching

Best user: •“Better in person” personalities who treat dating apps like homework
• Selectively social types (structured events yes; random bar, cold-approach no)
• Efficiency maximizers tired of catfishing and texting-that-goes-nowhere
• Users who want real-time vetting over weeks of digital back-and-forth

Churn watch: Thursday’s risk is quality dilution at scale.

The magic happens when an event hits critical mass (50+ attendees, balanced gender ratio, good venue vibe). But scale too fast, and you get: 15 people in a massive bar feeling awkward, or 200 people where it’s just a nightclub with extra steps.

Repeat attendance is the key metric. If <30% of first-timers return for a second event within 60 days, Thursday is just “something I tried once” rather than a community. If >50% return, network effects kick in - regulars recruit friends, the scene develops, and word-of-mouth becomes the primary acquisition channel.

Another watch: Copycat risk. Unlike Timeleft’s personality matching or Hinge’s algorithm IP, Thursday’s model is embarrassingly replicable. Any nightlife promoter can host “singles night” events. Thursday’s moat is brand trust and operational execution- hard to defend if competitors undercut on price or over-deliver on experience.

The six retention mechanisms

that keep users coming back (or deleting successfully)

These platforms aren’t guessing. They’ve identified specific levers that prevent the silent churn—where users don’t officially quit but stop showing up. Here’s what’s working:

1. Outcome Visibility

(Hinge’s “We Met” Feature)

Traditional dating apps are black boxes. You swipe, match, message, maybe meet - but the app never knows if you actually went on a date.

Hinge closes this loop by asking both parties: “Did you meet? How’d it go?”

This data feeds back into matching recommendations. Over time, the algorithm learns which profile combinations lead to real dates, not just matches. Users who see improvement (better conversation quality, higher meet-up rates) develop trust in the system.

Why it prevents churn: Visible progress creates the belief that effort compounds. Without it, dating feels like spinning wheels.

2. Scheduled Cadence

(Timeleft’s Weekly Dinners)

Humans are terrible at coordinating spontaneous plans. “Let’s get dinner sometime” dies in calendar Tetris. Timeleft removes that friction by making Wednesday night standing plans.

Subscribe once, show up weekly, meet different people each time. The habit forms fast - within 4 weeks, “Wednesday dinner” becomes part of your routine.

Why it prevents churn: Behavioral science shows habits need consistency + trigger + reward. Timeleft provides all three. Breaking the streak feels like missing yoga class—you notice the absence.

3. Curiosity Loops

(Timeleft's Mystery Box Model)

Timeleft withholds who you’ll meet until you show up. You know the restaurant, the time - but not their names, faces, or profiles until you’re physically at the table.

Instead of endless browsing creating decision paralysis, mystery creates anticipation.

Why it prevents churn: First-timers show up because curiosity is a stronger motivator than confidence (”Who are these 5 people I’m 85% compatible with?”)

4. Intentional Friction

(Hinge’s Limited Likes)

Hinge caps daily likes at 8-10 for free users. This sounds anti-growth, but it’s retention insurance.

Unlimited swiping creates decision fatigue and devalues each match. Limited likes force users to actually read profiles before committing. The result: higher match quality, better conversation rates, and less burnout.

Why it prevents churn: Paradoxically, constraint increases satisfaction. Users feel like they’re making real choices, not passively consuming an infinite feed.

5. Group Buffer

(Timeleft’s 6-Person Dinners)

One-on-one first dates are high-pressure. If there’s no chemistry, you’re trapped for 90 minutes faking interest. Timeleft’s group format diffuses this anxiety.

If one person is boring, you pivot to someone else. If conversation lags, the group carries it. There’s no “failed date” feeling - just a fun dinner where you happened to click more with some people than others.

Why it prevents churn: Lower stakes = higher willingness to try again. Users who have one mediocre dinner will still book another because the format itself is enjoyable, not dependent on perfect matching.

6. Anti-Vanity Metrics

(All Three Platforms)

None of these platforms show follower counts, likes, or “hotness” scores. Hinge hides mutual likes until both parties engage. Timeleft doesn’t show photos until after the dinner is booked. Thursday doesn’t rank attendees.

This removes the “am I hot enough for this?” anxiety that plagues Instagram and swipe apps.

Why it prevents churn: When success isn’t defined by external validation, users stay engaged longer. You’re not chasing likes - you’re showing up for experiences.

Key Takeaways

For Product Teams:

1. Optimize for outcome, not engagement. Hinge’s “designed to be deleted” works because it aligns business goals with user goals. If your product succeeds when users leave, build the churn loop into the moat - users will pay more and refer more.

2. Structure beats choice. Infinite options create paralysis. Timeleft’s “we pick the restaurant, you pick the week” removes 90% of coordination friction. Users aren’t asking for more control - they’re asking to stop making decisions.

3. The app is not the product. Thursday learned this the hard way. Their matching algorithm was fine - but nobody cared. What users wanted was a bar full of singles on Friday night. Evaluate whether you’re building a product or the process

4. Social proof must be real-time. Post-event photos, user testimonials from last week, live counters of “43 people attending tomorrow”- these work because they’re immediate. “Trusted by 1M users” does not tell you anything. “Sarah met her boyfriend at last Thursday’s event” is a signal.

For Investors:

5. Look for inverted metrics. Healthy churn can be bullish if users are graduating to outcomes. Track: (a) DAU drop + NPS spike = users succeeding, (b) Repeat rate >50% = habit formation, (c) Referrals from churned users = proof it works.

6. The TAM is loneliness, not dating. 300M people have zero close friends. Dating apps address a subset. JoyTech platforms address the root problem. Timeleft users aren’t all looking for romance - they’re looking for connection. That’s a 10x larger market.

7. Watch for unit economics, not growth hacks. Timeleft grew without ambassadors or influencers - just ads that hit break-even on the first dinner. Thursday scales through venue partnerships that cover venue costs. Hinge users pay 2x Tinder pricing. If LTV doesn’t cover CAC in 60 days, the model is broken regardless of user count.

8. Retention is the real moat. Network effects in social products flip fast. Clubhouse had 10M users and died in 18 months. The question isn’t “how fast can you grow?” It’s “what happens in month 3 when novelty fades?” Platforms that build weekly habits (Timeleft) or community identity (Thursday) survive. Those relying on algorithms alone - don’t.

If you’re finding this newsletter valuable, share it with a friend, and consider subscribing if you haven’t already.

Subscribe now

If you opened your phone this morning, you probably checked at least three different apps before even getting out of bed: your calendar to see what meetings you have, your to-do list to remember what’s urgent, maybe your weather app to decide what to wear.

Then you checked Slack to see what blew up overnight. Then email. Then your notes app to find that thing you wrote down yesterday. By the time you’ve actually started your day, you’ve already switched contexts seven times.

Hard to imagine we thought this was normal.

That wasn’t always the case.

In 2016, the productivity software landscape celebrated specialization.

You had a calendar app (Fantastical), a separate task manager (Wunderlist), another app for notes (Evernote), different tools for habits (Coach.me), and distinct platforms for communication.

The philosophy was ‘best-of-breed’, - to pick the absolute best tool for each job.

Users accepted the coordination tax. We manually copied information between apps. We got good at remembering which system held what information. Context switching was just the cost of having the right tool for everything. We have to be doing this constant ‘manual context stitching’.

The real shift happened when we stopped being able to hold it all together.

The average knowledge worker now juggles 9-10 applications daily, switching between them over 1,200 times. Twelve hundred context switches. Every single day.

The cost: 40% of productive time lost. 23 minutes to regain focus after each interruption. An estimated $450 billion annually in the U.S. alone.

What started as ‘best-of-breed’; had become ‘death by a thousand apps’.

Market at a glance:
AI productivity tools are growing from $8.9 billion in 2025 to a projected $35 billion by 2034; 16% CAGR. Process orchestration tools are scaling even faster: $8.2 billion to $43.7 billion by 2034.

But the real driver isn’t market size. It’s user intent evolution.

2016: I need a tool that does X.

2019-2021: I need tools that work together.

2023-2026: I need one system that handles everything and learns from the context.

This mirrors what happened with smartphones. You don’t want a phone, a camera, a GPS, and a music player anymore. You want an iPhone.

Daily assistants are trying to become the iOS of productivity.

App-of-apps Era Is Here

Apple publicly acknowledged it needs Siri to become more personal + context-aware + capable across apps, and delayed some upgrades until 2026 - basically confirming that “cross-app action” is now the bar.

While we are all waiting for that, we have other players approaching this emerging next frontier.

Littlebird, Hero and Tiimo are not variations of the same product. They are answers to different cognitive failures in modern life: temporal chaos, overwhelmed psyches, and memory fragmentation are not variations of the same product. They are answers to different cognitive failures in modern life: temporal chaos, psychological overwhelm, and memory fragmentation.

You can’t remember what you decided in yesterday’s meeting. The context is somewhere - Slack? Email? That doc someone shared? -- but finding it means opening six apps and trying to piece together fragments of a conversation you were barely paying attention to because you were simultaneously responding to teams’ messages.

Littlebird watched the whole thing. It transcribed it. It knows what you decided. It even knows why.

It’s like ChatGPT that doesn’t need to be instructed about the context. It just captures it in the background. It totally feels like it’s your personal iOS. Or your Chief of Staff, who is present with you throughout all working hours.

Primary intent: Context continuity

Job it does:

“I want to pick up where I left off without reloading everything in my head.
And… I don’t want important threads to disappear between sessions.”

It translates to: Capture context ✔️ Preserve threads ✔️ Recall state ✔️ Continue thinking ✔️

Best user: • Deep knowledge workers • Builders, researchers, writers, engineers
• People working on long-running, non-linear projects • Users who think in threads rather than tasks

Churn watch:

Littlebird promises you'll never search again -- it watched everything. But if finding that vendor decision from three weeks ago still requires multiple queries and scrolling, users think: "I could've just CMD+F'd Slack."

When passive observation doesn't deliver mind-reading retrieval, expensive storage loses to active search.

A lot of you guys reading this article don’t know me.

Hi, I’m Daria Littlefield. I’ve spent the last decade leading customer operations across a $400M ARR portfolio — 35+ apps in dating and social space.

The kind of apps people open 50 times a day, then delete in a rage, then reluctantly reinstall a week later.

That means I’ve seen how products die. Not from lack of features, but misunderstanding why users signed up in the first place.

Do Not Churn is where I break down the products we all use and show you exactly what makes users stay or leave. Why do some apps hit $100M ARR in under a year while others bleed users before launch? The retention mechanisms you use every day but never notice. The difference between a feature and a dependency.

I answer every comment here.
So tell me: What’s one app that became essential without you realizing it?

Drop your answer below. Or just say hi -- I’m trying to actually know who reads this stuff.

Leave a comment

Built by former Meta engineers who worked on Facebook Stories, Hero positions itself as your ‘daily operating system’. It consolidates calendars, to-dos, habits, grocery ordering, notes, weather, and AI search into a single interface.

Primary intent: Daily orchestration

Job it does:

“I want my tools to stop competing for my attention.
And… I don’t want to decide what matters every time I open an app.”

It translates to: Aggregate signals ✔️ Prioritize context ✔️ Surface what matters ✔️ Reduce clicks ✔️

Hero’s core innovation isn’t any single feature: it’s temporal orchestration.

The morning briefing (accompanied by chirping bird sound the moment you open it to dive in).

SuperAlerts that break through Do Not Disturb.

Calendar events overlaid with weather so you remember to bring an umbrella.

Voice input so fast you can create tasks in 3 seconds.

Churn watch:

Hero’s retention depends on being the first app you open -- the morning briefing, the daily ritual. But if users realize they can check their native calendar, to-dos in Reminders, and weather on their lock screen in three taps, Hero’s consolidation value evaporates. The moment separate apps feel less taxing than Hero’s lock-in, users leave.

Winner of Apple’s 2025 iPhone App of the Year, Tiimo approaches daily assistance differently. It’s a visual planner co-designed with ADHD and autism experts, built around one insight: planning systems fail not because users are lazy, but because life is unpredictable and traditional tools are punishing. The app has probably one of the most clear and visually appealing onboarding flows I’ve ever seen.

Primary intent: Planning clarity

Job it does:

“I want my day to feel understandable and predictable.
And… I don’t want time to surprise me or overwhelm me.”

It translates to:

Visualize time ✔️ > Structure routines ✔️ > Reduce anxiety ✔️ > Support execution ✔️

Best user: • Neurodivergent users (ADHD, autism, anxiety) • Anyone struggling with time perception • Users who want structure but full control • People early in their productivity journey

Where Hero optimizes for consolidation and Littlebird for intelligence, Tiimo optimizes for emotional safety. Every design choice reduces cognitive load. Color-coded visual blocks replace text lists. Emojis make tasks scannable at a glance. The focus timer makes time ‘visible’ for people with time blindness.

AI voice chat that listens and turns your asks into time blocks in your schedule.

Churn watch

Tiimo promises planning without punishment, but life doesn't follow visual timelines. When a user's color-coded, emoji-filled day shows what they meant to do instead of what's actually happening, the app transforms from gentle companion to visual reminder of failure. If the timeline feels rigid instead of forgiving, users return to chaotic but non-judgmental todo lists that don't look disappointed.
Unlike the apps above, Tiimo doesn’t aim to own operational or informational context, while it targets to own experiential context (how the day is perceived).

That doesn’t sound like a problem; however, not being able to sync data with your personal/work calendar creates a necessity to ‘check in two places’ - exactly what other apps are trying to avoid.

Different products. Different user intents. Same retention architecture underneath. Below are the six mechanisms that make daily assistants impossible to leave -- steal whichever ones fit what you're building.

Share Do Not Churn

The six retention mechanisms

1. Data aggregation moat

These apps don’t just store data; they become your external brain. Hero consolidates calendars, tasks, habits, notes, and grocery lists. Littlebird captures meeting transcripts, daily journals, and cross-app activity. Tiimo holds your routines, mood logs, and visual timeline history.

The switching cost isn’t technical; it’s historical. Your entire personaland professional history lives in one system. Switching means losing that context and the interconnections between events, tasks, and notes that can’t be exported to another platform.

2. Habit formation through temporal anchoring

Hero’s morning briefing creates a ‘must-check’ ritual. SuperAlerts trains users that Hero is their attention authority.

Tiimo’s visual timeline becomes the way you see time itself. Habit streaks create progress you don’t want to lose. Research shows SaaS habits form in 30-45 days. Daily rituals accelerate this timeline. Once you’ve maintained a 90-day routine streak in Tiimo, abandoning the app means losing that visual representation of consistency.

3. Social dependency networks

This is the most powerful lock-in: making your productivity dependent on others’ usage.

Hero’s My People’ feature means your partner or family must also use Hero for full value.

Tiimo allows sharing up to 5 family profiles. If you’ve eliminated coordination texts via Hero, switching back to texting feels like a regression. The alternative app must also solve the social coordination problem and convince your entire network to switch with you.

4. Behavioral learning investment

Littlebird’s value proposition is that it already understands the work behind your requests. The first 30 days involve training the AI on your patterns. Tiimo’s AI Co-Planner learns how long your tasks actually take and suggests realistic time estimates. Switching means retraining a new system from zero.

Hero’s AI briefing gets better over time as it learns what news and quotes you prefer. Once the system predicts your needs, reactive tools feel primitive. This is the iPhone effect -- you don’t realize how much you value prediction… until you lose it.

5. The ‘first-open’ strategy

Daily assistants position themselves as the first app you open each day.

Hero’s morning briefing trains users that productivity starts here, not email or calendar apps. Tiimo’s widgets and Live Activities mean your next task is always visible on your lock screen -- you never truly close the app.

Littlebird runs perpetually in the background. It’s never truly closed; switching off requires conscious opt-out.

6. Strategic pricing and emotional commitment

Hero is 100% free. This isn’t altruism - it’s a land-and-expand strategy. Give away the core product to build critical mass and network effects. Once users and their social networks are locked in, introduce premium features that command high prices because switching costs are too high.

Tiimo uses a different approach: freemium with emotional positioning. At $54/year ($4.50/month), it’s priced accessibly for neurodivergent users, many on fixed incomes. But here’s the retention insight: Tiimo explicitly removed lifetime subscriptions despite user demand. Because subscriptions create ongoing relationships, not one-time purchases. Users who pay monthly stay engaged monthly.

The pricing psychology is sophisticated. Tiimo positions itself as supporting neurodivergent people, not exploiting them (100% ad-free forever, co-designed with ADHD experts, transparent about costs). This builds trust and community identity. Canceling Tiimo feels like abandoning a mission you believe in, not just dropping a tool.

Subscribe now

Key takeaways

• Daily assistants emerge from a $450B productivity crisis caused by tool proliferation and context switching.

• They don’t compete on features; they compete on becoming your temporal operating system, emotional support system, or memory augmentation layer.

• Three distinct approaches: Littlebird (ambient intelligence), Hero (consolidation), Tiimo (psychological safety)

• Retention is architected through six mechanisms: data aggregation, habit formation, social dependencies, behavioral learning, temporal anchoring, and strategic pricing.

• User language signals dependency -- “saved,” “impossible,” “everything” -- rather than feature satisfaction.

• Traditional productivity tools must evolve into platforms or risk commoditization as features.

• The fundamental question: Should AI assistants be personal agents or corporate platforms? Daily assistants claim the former but are designed as the latter.

This piece was born (and written) on the terrace of Café Valentina in Medellín, Colombia