Within 5 years, your agent will decide what data to share

1. Failed tax startup — ▶ 2:05

   After a year building AI business tax prep, Town's team found some product market fit but not enough, and reset for 3 months. They asked why nobody had built AI that operates out of email and calendar. The timing was Opus going fully agentic in November-December. A prototype built in a couple of weeks had product market fit almost immediately, without 100 customer interviews. OpenClaw was blowing up at the same moment, but it is open source and only tinkerers can use it. Town targets just about anyone.

2. Moats are a luxury — ▶ 4:20

   Denise is sure Town is a top three priority at Google and Apple within 12 months, not a top 10, and it wakes him at 3 a.m. Yet he calls talking about moats "a little bit of a luxury" before he has 100,000 or a million paying users in a market with a billion potential payers. None of today's products has deep product market fit, he says. Grockbot is a power-user product. Big players will not innovate their way to mainstream. They will copy someone who has already succeeded.

3. Agent-to-agent network effect — ▶ 5:50

   The winning product will have a network effect at the agent level. In Town's agent-to-agent feature, you ask your assistant, called a Towny, a question it cannot answer. It finds a co-worker's Towny that can, and asks. The feature sits in a sub-part of the product, but users who find it love it most. Once a whole team is on it, moving to another product is hard to imagine. Denise says no one has figured out multi-user, multiplayer AI, and that this will be the moat.

4. Rival moat theories — ▶ 7:09

   The market has competing theories: post-trained models per company or per person; context and connected data sources, since users resist re-connecting everything; pure distribution, where whoever has the users wins; and devices as the AI entry point. For personal use, Denise would worry most about Meta. A WhatsApp assistant is coming, in a day or 3 months, and everyone already messages there. Running through the list, his reaction is "oh my god what do I do", given all the competitive forces.

5. Agents decide what to share — ▶ 11:01

   Hot take: in 5 years users will trust agents to decide what data to share without intervening. Put an agent in a room with two friends planning a trip. When one jokingly asks for your medical history, it refuses, though no rule was set. Models do better the less siloed their information is. The old fix, classifying data and writing policies, is costly and leaves data stuck in inboxes. In his sales-intro example, one agent surfaces Liz's personal contact and Bob's meeting next week. Salary and medical records stay sacrosanct.

6. Humans set goal and budget — ▶ 16:03

   LLMs will make far fewer errors than humans, he says, and soon. His example is a colleague nicknamed "big reply", a top 1% performer who sent a private jab about an acquisition to the whole company. On goal-seeking agents, he has no answer on whether it is a feature or a bug. His position is that the human sets the goal, allocates the token budget and monitors the shape of the actions. Monitoring might eventually be another agent, but never the one pursuing the result.

7. Routing breaks personality — ▶ 19:30

   Town picks models by task and cost, for example Gemini or OpenAI for images and 11 Labs for voice. The user-facing layer is different, because people notice personality drift. They file tickets when their text agent writes 8 sentences instead of 4 or changes its capitalisation. Anthropic keeps its model families consistent, so swapping in Kimi for final output makes people feel their AI changed. Below that layer, on pure reasoning, routing is free. For coding it matters less, since the test is whether the code works.

8. Costs fall, frontier share unknown — ▶ 22:49

   Town subsidises in part today, betting the cost curve fixes it in 18-24 months. Email labelling needs neither Opus nor Sonnet, and prices halve every 9-12 months, so he can price now to reach 20-30% margins in 18 months. The unknown is whether 10%, 20% or 30% of tasks stay at the frontier, and "literally nobody" knows. Town is mostly frontier today. Engineer hours are better spent on network effect features and new data sources than on lowering COGS.

9. Tinkerers drive adoption — ▶ 26:44

   A single tinkerer who builds team skills, integrations and routines predicts faster adoption, because everyone else gets those building blocks free. Town wants single-player onboarding to work without one, but power users next to others make them more successful. Underserved functions adopt early. Executive assistants, chiefs of staff, HR, junior finance and recruiters get little AI and still run on email, while sales ops people get 10 AI pitches a day. Penetration then spreads through ops teams.

10. Three bets, Europe test — ▶ 30:12

    Town makes three bets. One is a single named, personalised AI, defensible the way Snapchat is, because of its strong opinion. The second is network effects. The third is heavy pre-processing, spending compute before any question to build a mental model of the user. For a European rival he sets a test: enough TAM and war chest to match capabilities. Codex has about 100 people on it, and users will ask why they should pay $50 a month when OpenAI costs $24.99.

11. Build fast, learn slow — ▶ 33:49

    Copying now takes 2-4 weeks, so a lead no longer compounds into learnings. "You can build now at the speed of machines, but you can only learn at the speed of humans." In the first half of the year he saw 15 startups that mattered. Now he sees 2 or 3, and refuses to name them. Anthropic, Cursor and Grockbot move at startup speed with cost and scale advantages. He never feels competitors are 6-12 months behind.

12. Apple's two disadvantages — ▶ 39:00

    Apple is not a cloud company, and agents improve with data that is not on the phone. It has also contorted itself around an on-device privacy story that keeps it far from the frontier, since local models are slower and dumber. The new Siri will be good, convenient and cloud-enabled, but he puts it about 9 months behind Town and Grockbot on capability. His advice is a hire with totally different DNA, willing to throw principles away or lose this generation.

13. Success means paying — ▶ 43:48

    A successful user is someone who pays every month, not someone who burns tokens. The same meeting-prep workflow costs identical tokens but is worth a lot to someone in back-to-back meetings and nothing to someone just sitting in. Town now emails users about rogue routines, which became one of its most popular features in the last month. In the unpriced private beta, one user spent $26,000 of compute in 5 months. The $15 plan has the worst unit economics and the $99 plan is the most profitable.

14. More users, paying less — ▶ 47:58

    He would take 100 million consumers at $20 a month over 1 million at $100. Personal use has a ceiling, but work spend grows as AI creates value. A recruiting firm pays Town $500-600 a month across its users and can take another client worth $3,000 a month because of the automation. He thinks he could sell it another $500 of usage on the same logic. He is also doubtful of ad-backed consumer AI, since tokens cost too much and ads pollute an assistant's recommendations.

15. No pricing power at frontier — ▶ 51:31

    Voice cost does not stress him. The share of work at the frontier does: "I have zero pricing power at the frontier." He points to Cursor, with big market share and loved by customers, but competing with its suppliers at 70% margin. He is competing with his suppliers "100%" today. At tens of millions of paying users he would still hand 20-30% of workloads' revenue to the labs, and he needs a solution by then. On ElevenLabs at $22 billion, he says it is a lot of money and voice may top out.

Silos are the tax on every AI deployment inside a company. Jean Denise puts it in information-theory terms. An LLM with access to all the world's information and unlimited time would be as effective as its intelligence allows, because it would always find the right context eventually. Real organisations look nothing like that. Data sits in inboxes, in unintegrated systems, behind policies that compliance teams spend months writing. The old fix is to classify data and decide who may see what. It is slow and costly, and the information the model needs often ends up stuck anyway.

Denise's fix is to stop making humans the filter. In 5 years, he predicts, people will trust their agent to decide what leaves their silo, without intervening and without a hard rule ever being set. Put your agent in a room with two friends planning a trip. They ask about your food preferences and when you can fly. One of them jokes and asks for your medical history. The agent refuses, and you never told it to. Salary and health records stay sacrosanct. Most other siloed information does not need to be.

The business version is concrete. Someone on a sales team wants an introduction to a person at a customer. Today they ask in Slack who knows them. In Denise's version, their agent asks every colleague's agent, each with access to an inbox. Options come back. Liz has a personal relationship, not a work one, so ask whether she is willing. Bob has a work relationship and a meeting next week, so ask whether you can join. That is the outcome the company wants. It requires employees to accept that part of their inbox is queryable by others. He concedes this sounds insane today.

This is also where he locates the moat. The product that wins will have a network effect at the agent level. Town's version is agent-to-agent. You ask your assistant, which the company calls a Towny, a question it cannot answer. It finds a co-worker's Towny that can, and asks. The feature sits in a sub-part of the product, yet users who find it love it most. Once a whole team is on it, moving elsewhere is hard to imagine. No one has figured out multiplayer AI, he says. The trust thesis and the moat are one bet. If agents are not trusted across silos, there is nothing to network.

He is candid that moats are premature. The market has rival theories. Post-trained models per company or per person. Context and connected data sources, which users resist re-connecting elsewhere. Pure distribution, where whoever already has the users wins. Devices, where whoever owns the entry point owns the assistant. For personal use he fears Meta most, with a WhatsApp assistant arriving in a day or in 3 months. But moats only matter after scale. Town needs 100,000 or a million paying users in a market with a billion potential payers. None of today's products has deep product market fit, he says. Grockbot is a power-user product. Big players will copy their way to mainstream, but only after someone proves the product. He still assumes he is a top three priority at Google and Apple within 12 months.

What is working so far? Town gives up a lot to get the data. It requires users to connect email and calendar before anything works, and 30% leave at that gate. The reward is that Town can see the work someone already does and recommend automations for it. More than 15% of users who try the product pay, which he calls extremely high for product-led growth. ChatGPT asks users to connect email but does not force it, and he thinks its suggestions are plain bad. The network effect shows up in a pattern too. One tinkerer building team skills, integrations and routines predicts faster adoption, because everyone else gets them free. And the hungriest teams are the unglamorous ones: executive assistants, chiefs of staff, HR, junior finance, recruiters. Sales ops people get 10 AI pitches a day. These people get none.

Revenue is the test, not token consumption. Denise defines a successful user as someone who pays every month. Tokens are a dangerous proxy because the same meeting-prep workflow costs identical tokens and is worth a lot to someone in back-to-back meetings and nothing to someone who only sits in. Town now emails users when it spots rogue routines burning tokens, and that became one of its most popular features in the last month. The contrast is the private beta, which had no pricing. One user spent $26,000 of compute in 5 months, with no pushback. He wants the market to push back. He also will not give the product away to businesses, because a 500-person company will not run on something that may be switched off. The plans are $15, $49 and $99. The $15 plan has the worst unit economics. The $99 plan is the most profitable.

Given a choice, he takes 100 million consumers at $20 a month over 1 million at $100. Personal use has a ceiling, because there are only so many restaurant bookings and trips. Work use compounds, because AI that creates value makes companies want more of it. A recruiting firm pays Town $500-600 a month across its users. The automation lets it take another client worth $3,000 a month. Families are the awkward counterexample. Parents buried in school emails, sports leagues and camp scheduling show strong fit and talk about the product in WhatsApp groups. But they pay less and do not expand across an org, so the team argues over whether to market to them.

The weak point is cost. Town runs mostly on frontier models today and subsidises in part. His bet is the cost curve. Labelling email needs neither Opus nor Sonnet. Prices halve every 9-12 months, so he can price today to reach 20-30% margins in 18 months. He has not moved work to open-weight models because an engineer hour spent on cost does not make the product better. It is better spent on network effect features and new data sources.

The unknown is how much work stays at the frontier. It could be 10%, 20% or 30% of tasks, and nobody knows. That share is the problem because, in his words, "I have zero pricing power at the frontier." He points to Cursor, with big market share and loved by customers, yet paying its suppliers while competing with them at 70% margin. He competes with his suppliers today, "100%". Reaching tens of millions of paying users still leaves him handing the revenue from 20-30% of workloads to OpenAI and Anthropic. He needs a solution by then. He does not have one now.

Speed offers no relief. A lead in product market fit used to produce learnings that fed the next feature, and a copier stayed three learnings behind. Now copying takes 2-4 weeks. "You can build now at the speed of machines, but you can only learn at the speed of humans." In the first half of this year he saw 15 startups that mattered. Now he sees 2 or 3, and declines to name them. Anthropic, Cursor and Grockbot move at startup speed with scale advantages, and he never feels competitors are 6-12 months behind. Apple is the opposite case. It is not a cloud company, and its on-device privacy stance keeps it far from the frontier. Agents get better with data that is not on the phone. The new Siri will be good and convenient, but he puts it about 9 months behind.

The obvious objection is that none of this works if the agent errs. His answer is that LLMs will make far fewer mistakes than humans, and soon. He tells of a colleague nicknamed Big Reply, a top 1% performer who sent a private jab about an acquisition to the whole company. Humans leak data all the time, he says. That is fair, but an anecdote is not an error rate, and he offers no number. He also hedges on a related issue. In an example the host raises, a goal-seeking agent nearly made an unwanted purchase and was stopped only because the item needed engraving. Whether that tendency is a feature or a bug, he has no answer. His position is that the human sets the goal and the token budget and monitors the shape of the actions, perhaps with a second agent, but never the one chasing the result. That is a long way from an agent deciding alone what to share about you.

He also accepts that bad events will come. Humans will never again read every line of code, and he expects the industry to learn the way it learned with chemicals: dump, harm, then regulate. His thesis needs privacy boundaries to hold before that cycle plays out. Two questions stay open. One is how much work stays at the frontier, because that decides the margin. The other is whether people hand over the keys to their silos before the first big breach teaches them not to.