Town Founder Interview: How Email and Calendar Agents Achieved PMF

From Tax Tool to Agent: A Decisive Pivot

Town founder Jean Denis was formerly the CTO of fintech company Plaid. After leaving Plaid, he chose a seemingly dull domain—tax automation for small and medium businesses. The team spent a year polishing the product, which was a success from a product perspective, but commercialization hit a ceiling.

"If you look at our first-year revenue, slightly under a million, and in the second year we thought we could do two and a half to three million. Before AI, you'd think that was reasonable. But after AI, that's not going to be a venture-scale company."

More critically, Jean realized that incumbents could use AI to automate tax processes faster, while a startup could only sell tax services once a year, and word-of-mouth spread was extremely limited. After the first tax season ended, he and his co-founder spent a week in deep deliberation and decided to abandon this direction.

This decision was not easy. Jean admitted it felt like failure at the time, but core motivations drove them to change: first, to have a sufficiently large impact on the world; second, to be responsible to the team—everyone staying at the company has an opportunity cost. "If you no longer believe you can achieve good results, it's hard to walk into the office every day and look your colleagues in the eye."

Why Has No One Built Great Email and Calendar Agents?

After the pivot, the team seriously explored four directions and ultimately locked onto one question: why has no one built an excellent agent around email and calendar?

The clue came from customers of the tax product. They told Jean that automated tax filing was great, but scheduling client meetings, preparing meeting materials, and rehearsing common questions still consumed a lot of time. These needs weren't complex, yet they had never been well addressed.

It happened to coincide with model capabilities beginning to support more autonomous agent work, and Town quickly achieved product-market fit (PMF). Jean describes PMF as a magical thing—once you grab it, you just want to keep pulling on that thread. The team didn't stop to assess market size, but went all in for several months; the product grew every day, and customers kept asking for more.

They gradually realized that most external communication for businesses—sales, business development, recruiting—runs on email, which means the market is not just the US but global. More importantly, connecting an LLM to email is only the starting point; there's a lot of complexity to truly deliver a great experience, which gave the team room to dig deep into the product.

What is an Agent?

In the AI context, an agent refers to a system that can perceive its environment, plan autonomously, and execute multi-step tasks to achieve goals, as distinct from a chatbot that merely responds to a single input. It typically has tool-calling capabilities (e.g., accessing email, calendar, databases), memory mechanisms, and task decomposition abilities. The "email and calendar agent" in this article refers to one that can automatically schedule meetings, prepare materials, rehearse questions, etc., rather than a passive assistant that only offers suggestions.

Agent implementation depends on the reasoning and planning capabilities of large language models (LLMs) and external tool interfaces. When early models were insufficient, agents were mostly demos; with advances in long context, function calling, and autonomous error correction, building reliable production-grade agents became possible. Town's pivot coincided with this window, which is why it could quickly achieve product-market fit.

Product Experience Can't Just Be a Text Box

Jean repeatedly emphasizes a core point: AI products can't just give users an input box. Ordinary people don't know what AI can do for them; the product needs to proactively demonstrate value and directly do things for the user.

He uses a restaurant owner as an example. Wholesalers regularly send Excel or PDF files containing prices for thousands of products, with prices constantly changing. In theory, finance staff should analyze price changes every quarter and look for better suppliers, but in reality almost no one does this because it's too tedious. Town can automatically run an analysis once a month—costing maybe just one dollar—and even if nothing is found in eleven out of twelve months, as long as one month uncovers hundreds or even thousands of dollars in savings, it's worth doing.

"Product experience can't just be a text box. You have to show someone the value proposition, and then do it for them."

This example also illustrates the challenge of product marketing: if you ask a restaurant owner whether they want an administrative assistant, they'll look blank; but if you say "I can help you save money on lettuce," the answer is yes.

Using AI to Customize Software for Each User

In Jean's view, in the AI era most agent frameworks' capabilities will converge, and traditional advantages based on feature leadership will become increasingly hard to maintain. Town's bet is that as AI gets stronger, it can customize the product to each user's specific needs—electricians, venture capitalists, administrative assistants, each receiving completely different recommendations.

The implementation is somewhat semi-manual. Because they can't look at user data, the team analyzes high-frequency users on the platform, interviews ten restaurant owners, distills five common automation needs, builds prototypes of "routines" internally, and then pushes them as templates to new restaurant owners. As the scale of interviews expands and models improve, this library will become richer and recommendations more precise.

Jean believes that another popular model—the "forward deployed engineer" approach of sending real people into enterprises to customize software—will actually get worse as models get stronger, whereas Town's path of using AI to build customized software scales better.

Forward Deployed Engineer (FDE) Model

The FDE model refers to software companies dispatching engineers to reside at customer sites, deeply understanding business pain points and customizing solutions, with Palantir as a typical example. Its advantage is the ability to quickly deliver highly tailored products, but it is costly and scales poorly, making it hard to serve large numbers of small and medium businesses.

Jean believes that as model capabilities strengthen, the FDE model will actually get worse, because AI can achieve customization at lower cost. Town's approach is semi-automated: by interviewing high-frequency users to distill common needs, building reusable "routine" templates, and then using AI to recommend different templates for users in different industries, achieving FDE-like customization effects at scale.

Competing with Model Providers: Neutrality as a Moat

Town directly uses models from providers like OpenAI and Anthropic, which in a sense means competing with its suppliers. Jean doesn't shy away from this, but makes neutrality a core selling point.

His question to mid-sized enterprise customers is: do you know which model provider will be best in the next three to five years? If you buy one company's product, you may be locked into that company's model—what if it's no longer leading, or raises prices? As a neutral application, Town selects the most suitable provider for each scenario—general intelligence, voice, video, slides each have their strengths.

The second differentiator is data ownership. Jean believes enterprises are accumulating a new type of asset: all the context about the business that AI knows. Town's position is that customers truly own this data and memory, and if a better solution emerges in the future, migration should take only a day. This lowers customers' switching costs, but it's Town's competitive strategy against model providers.

"If I can't have the best product on the market, I concede. But this is my competitive stance against model providers."

Additionally, Town doesn't train models, and bets that future human-AI relationships will be more forward-thinking than model providers imagine—for example, trusting agents to mediate information permissions within a company, so that questions like "who knows someone on that client's finance team" are answered in seconds, rather than waiting for slow human routing in Slack.

Model Neutrality and Data Ownership

Model neutrality means the application-layer product is not tied to a single foundation model provider, but dynamically selects the most suitable model for the task scenario (e.g., different providers for general reasoning, voice, and video generation). This prevents customers from being hurt by a model falling behind or raising prices, and also reduces the technology lock-in risk for application-layer companies.

Data ownership concerns the "AI context" that enterprises accumulate—all the knowledge about the business that the agent possesses. Town promises that customers own this data and memory, and migration takes only a day. This is actually a competitive strategy: by lowering customers' migration costs, it exchanges for customers' long-term trust in Town, thereby countering the threat of model providers extending upstream into the application layer. Jean's stance is that if the product can't be the best on the market, he concedes, but neutrality and data sovereignty are Town's differentiation bottom line.

In the Agent Era, the Value of Supply Aggregation Is Diluting

In the latter part of the interview, investor Alex Rampel offered an observation: agent traffic has far exceeded human traffic, and platforms like Amazon take an adversarial stance toward agents, stemming from the impact on the advertising business model.

Jean's response is more philosophical. He believes human friction comes in two types: friction you enjoy and friction you don't. Strolling in Soho is enjoyable friction, while ordering toilet paper at home is not. Agents shouldn't blindly eliminate all friction, but understand that humans aren't always optimizing for efficiency.

He further points out a structural change: in the past, supply aggregation was valuable because human attention was limited; when attention can be automated, the value of supply aggregation declines. This poses long-term pressure on giants that rely on advertising and aggregation models.

As for Town itself, Jean defines the market as "helping people find more opportunities to deploy AI"—a machine that allocates tokens on behalf of users. As tokens get cheaper, more and more things that weren't worth doing with human labor become worth doing, which is exactly the space for agent products to continue expanding.

One-line Summary

Town pivoted from tax tools to email and calendar agents, countering giants with model neutrality and data ownership.

Paragraph Summary

Town founder Jean Denis, formerly CTO of Plaid, decisively pivoted to email and calendar agents after commercialization of a tax automation product hit obstacles. The team found that external business communication relies heavily on email, while existing agent experiences were insufficient, and quickly achieved product-market fit. Jean emphasizes that AI products can't just offer a text box—they must proactively demonstrate value and do things for users, such as automatically analyzing wholesaler price changes monthly. Town uses model neutrality and data ownership as its moat against the risk of suppliers like OpenAI extending into the application layer. As agent capabilities converge, competition shifts to using AI to customize software for each user, building an automation template library by interviewing high-frequency users. Jean also notes that when human attention can be automated, the value of traditional supply aggregation and advertising business models will be diluted.