Town's JD on the AI Assistant Race and Economics

A founder who wakes at 3 a.m. usually fears failure. Jean-Denis, who goes by JD and built the AI assistant Town, fears something stranger: being noticed too fast. "I know what I'm building is a top-three priority at Google and Apple, like, in the next 12 months," he says. "Not a top-10 priority, like, a top-three priority" (00:00). That sentence captures the strange physics of building inside the AI assistant category right now. The market is enormous, the tools are new, and everyone with a cloud budget and a model API key can see exactly what works within weeks of it working.

Town is an AI assistant that lives inside a user's email and calendar. It watches how someone already organizes their day, then suggests automations that quietly do parts of that work for them (04:26). JD arrived here by accident. He and his team spent a year building an AI tax-prep company, got some traction, and concluded it would never be a great business (05:13). They spent three months in what he calls "the wilderness" before noticing that almost nobody had built a serious AI product that operates out of email. The timing mattered: Anthropic's Opus model had just become capable of real multi-step, agentic work, not just a few isolated steps (05:51). A two-week prototype found product-market fit almost immediately.

Moats Are a Luxury

Ask most founders about defensibility and they reach for a tidy answer. JD refuses the premise. "Talking about moats is a little bit of a luxury," he says, "and you have to be more successful than Town is today for it to matter" (07:53). His actual plan is blunter: get to product-market fit so deep that scale becomes its own kind of protection. Underneath that, though, he does have a thesis about what eventually wins. He calls it the agent-level network effect. Inside Town, a user's assistant, nicknamed a "townie," can notice it doesn't know an answer and go ask a coworker's townie instead.

"I think the product in this category that will win will have a network effect at the agent level." — JD [00:00]

Once a whole team is wired together this way, JD argues, switching products stops being a software decision and starts being an organizational one. He thinks no company, including his own competitors, has genuinely solved multiplayer AI yet (08:50). It's a bet, not a fact, but it explains why Town's roadmap looks less like feature-shipping and more like plumbing between people.

That plumbing depends on trust, and JD's boldest claim in the conversation is about where that trust is going. He imagines a near future where a person's agent decides, on its own, what information is safe to hand to someone else. He offers a small, specific scene: two friends ask your agent about your medical history as a joke, and the agent simply declines, without ever having been told that rule explicitly (13:02). He extends the logic past privacy into corporate life, arguing that large language models, the AI systems trained to predict and generate text, will eventually make far fewer data-sharing mistakes than the humans who currently do that filtering by hand, the coworker who replies to the whole company instead of a small group, the HR staffer who shares a salary spreadsheet with everyone by accident (18:09).

None of this works if the economics don't work, and Town's numbers are unusually specific for a three-month-old product. JD prices Town's subscription tiers at $14, $49, $99, and $199 a month, on the assumption that compute costs roughly halve every nine to twelve months, which is why he expects the business to reach 20 to 30 percent margins within eighteen months even though it is not there today (25:02). Simple tasks, like labeling an email, already run on models well below the frontier, the most capable and most expensive models currently available. Complex custom workflows still eat frontier compute, and that gap, not raw usage, is what determines a company's margin ceiling. Voice is the sharpest version of this problem: ElevenLabs, a voice-AI company, sounds better than open alternatives but charges accordingly, and JD says he doesn't yet know whether voice quality will plateau before the price does (47:05).

Learning at the Speed of Humans

The single most striking claim in the conversation isn't about moats or margins. It's about tempo. "You can build now at the speed of machines," JD says, "but you can only learn at the speed of humans" (00:19). In the old startup playbook, a company that shipped a good feature first got months, sometimes years, to learn from it before a competitor caught up. JD says that window has collapsed to two to four weeks. When Town launched, he counted roughly fifteen startups worth worrying about. Within two quarters, that list shrank to two or three, not because most of them failed, but because the pace of imitation made a crowded field consolidate almost overnight (34:10).

That pressure shapes who JD watches. He says Town and the well-funded consumer assistant Instinct, which has reached a $2.5 billion valuation without a clear monetization plan (00:40), aren't really competing for the same customer: Town generates revenue from companies using it for work with network effects across team members, while he sees Instinct's strategy as more about customer acquisition with a product that is currently fully subsidized (36:03). He pays closer attention to GrokBot, built by Elon Musk's AI company, because it aims at a similar mainstream audience to Town's, even if he thinks its association with the X platform will put off some professional users on brand grounds alone.

"I think brand for them, I think some people just won't wanna touch it because of brand." — JD [36:55]

Underneath all of this is a simple test JD applies to his own product: does someone keep paying every month? He is openly wary of judging success by token usage, the volume of AI computation a user consumes, because more usage doesn't always mean more value. Early in Town's beta, before pricing existed, one user racked up $26,000 in compute costs over five months, with no natural brake on the spending (60:34). That experience pushed Town toward what JD calls a hard paywall: new users must connect their email and calendar before they can use the product at all, because without that access Town can't see enough to be useful. The requirement costs Town about 30 percent of prospective users on the spot (52:12), but the users who stay convert to paying customers at a rate above 15 percent, which JD calls unusually high for a product-led growth motion (50:26).

The most unexpected market Town found wasn't enterprise at all. It was families, parents drowning in school portals, sports schedules, and haircut appointments, a group with real willingness to pay but a smaller budget than a mid-market company (52:19). JD is still arguing internally about how much attention that segment deserves. It's a small tension that says something larger about building in this category: the technology moves fast enough that product-market fit shows up in places nobody planned for, and the discipline is deciding what to chase.

"Success for me is you pay me and you trust that we're the right platform that helps you both use AI but do so efficiently." — JD [43:49]

That discipline extends to headcount. JD says Town's run rate is at least $75,000 per engineer per year on AI tools like Devon, Claude, Cursor, and Codex (63:55), a number that would have sounded absurd three years ago and now reads, in his telling, as an obvious trade: AI-augmented engineers generate enough extra revenue that the tooling pays for itself many times over. It's a small, concrete number in a conversation full of large, unresolved bets, agent-to-agent trust, privacy norms, model economics, and it may be the most durable one. Everything else in the AI assistant race is still being negotiated in real time, at a speed no one, including JD, claims to fully control.