Why a $1M Startup Wasn’t Worth Building in the Age of AI
Jean-Denis Greze, Town's CEO and former Plaid CTO, explains why he shut down a nearly $1M-revenue AI tax business for SMBs and pivoted to Town, an AI assistant that works from email and calendar. The conversation with a16z's Alex Rampell covers founder decision-making under time pressure, product-market fit, the consumer-grade enterprise thesis, token allocation as a business model, neutrality versus frontier labs, horizontal versus vertical strategy, uneven AI diffusion, and how agents may reshape commerce and advertising.
Overview
Jean-Denis Greze, co-founder and CEO of Town and former CTO of Plaid, sits down with Alex Rampell of a16z to explain why he shut down a profitable AI tax product for small businesses and pivoted to building an AI assistant that works from email and calendar. The conversation covers the strategic logic behind the pivot, product-market fit, enterprise go-to-market, the competitive dynamics with frontier labs, and the broader implications of agents for distribution, advertising, and commerce.
Walking Away From a Working Business
Jean-Denis left Plaid after taking time off, then spent several months doing technical due diligence on automating small and medium business tax preparation. The team built the product over roughly a year. The product worked, but the go-to-market problem was harder than expected: reaching millions of businesses efficiently was full of open questions. After the first year, the business was just under $1M in revenue, with projections of $2.5M to $3M for year two. As Alex put it, “pre-AI you would have been like that makes a ton of sense, but post-AI… this was not going to be a venture scale company.”
The deeper issue was competitive. Incumbents with distribution advantages could use the same AI to automate tax prep for their existing customers faster than a startup could build network effects. Tax has structural properties that make hypergrowth difficult: it is sold once a year, it is rarely a topic of word-of-mouth, and people who are satisfied with their preparer rarely shop for a new one at the moment of satisfaction. Jean-Denis compared this to an edtech friend whose school sales are similarly constrained to two selling seasons per year, noting that he had chosen to forget this lesson when starting the tax business.
The decision to shut down was framed around founder motivation. Jean-Denis describes two reasons for building a company: impact, and working with a great team. Once he stopped believing the company could deliver a great outcome, it became hard to face employees, candidates, and the tax team whose jobs depended on the company’s direction. He emphasizes that the emotional weight of admitting failure was larger than the decision itself, and that the team was supported through severance and help finding next roles. The annual tax cycle also created a hard deadline: there was roughly a one-month window around October in which the team could live in both universes before commitments made the pivot impossible. He contrasts this with continuous businesses, where founders can always tell themselves to try one more idea for three more weeks.
How Town Was Born
Town emerged from the observation that customers of the tax product kept asking for help with adjacent work: preparing for meetings, drafting answers to recurring questions, and handling routine emails. No one had built strong AI around these common knowledge-worker workflows. The team explored roughly four ideas seriously before converging on the question of why no one had built a great agent operating from email and calendar.
The timing mattered. Agents built for coding were becoming capable of more autonomous work, and this made the simple insight tractable. Product-market fit arrived quickly. From December through April, the product grew by a percent or a percent and a half each day, with a clear roadmap driven by customer requests, until they moved from closed beta to a public launch. Jean-Denis notes that the team did not initially sit down to model TAM; they followed the pull from users, and only later realized that most external-facing roles, especially sales, business development, and recruiting, run primarily on email and calendar, making the market potentially global.
Consumer-Grade Enterprise and the Product-First Bet
A recurring theme is the difference between building for the user and building for the buyer. Jean-Denis draws on his Dropbox experience: it is far easier to take a world-class consumer product and make it enterprise-ready than to start from compliance and IT requirements and try to craft a great experience. Town therefore prioritized product quality, with data security and privacy foundations in place, and planned to shape the admin and enterprise surface later. The thesis is that consumer expectations for AI will set the bar, and CEOs will want the personal-assistant experience they enjoy at home inside their companies. The current challenge is scaling from a handful of enterprise deals to dozens, which Jean-Denis considers a more knowable problem now that the core experience is working.
Alex draws a parallel to the BlackBerry and iPhone transition and to the enterprise email client Good, arguing that an enterprise-ready wrapper alone is not a durable business. Jean-Denis agrees that the risk for Town is being “Good” rather than Gmail: it must be both convenient and enterprise-grade. He estimates that assistants are used by only a few million people worldwide today, and that most people will take perhaps six to eighteen months to experience the magic personally before bringing it into work.
Adoption Path and Enterprise Requirements
The adoption path described is incremental. Town begins by connecting a few people’s email and calendar, which is enough to produce visible value through meeting preparation and deck updates with the latest figures from attachments. Once users feel that value, the question becomes whether the product has earned the right to connect CRMs and systems of record. Deeper integration requires audit logs, permissioning that keeps privileged information away from the agent, and cost controls so that CFOs can see how tokens are being spent across teams. Those requirements slow enterprise adoption, while SMB and mid-market customers face fewer barriers and adopt faster. Town’s current strategy is to focus product experience and self-serve growth on SMB and mid-market, and to build compliance and finance features for enterprise over time.
Alex adds a Jevons paradox angle: most people cannot afford an executive assistant, so the market is small, but if the function can be produced through software, demand could expand dramatically. He gives a personal example where Town caught a missed dinner invitation that was in an email but not on his calendar. Jean-Denis frames the broader idea as a “token allocation machine” rather than an assistant. As tokens get cheaper, tasks that were never worth human time become viable, such as writing detailed prep notes for every meeting or researching attendees’ recent public statements.
Concrete Routines and the Restaurant Example
Jean-Denis gives a vivid example of a routine: a restaurant owner receives monthly or quarterly price spreadsheets from wholesalers listing thousands of items. Comparing prices and finding better suppliers would save money, but most owners never do it because it is tedious and not part of operations. A Town routine that runs monthly costs roughly a dollar a month. In eleven out of twelve months it finds nothing, but in the one month it finds a price spike or a better supplier, it might save hundreds or thousands of dollars. The key challenge is that owners will not think to ask for this; the product must surface the opportunity and make it concrete. Alex likens the framing to a Geico-style pitch: owners don’t want an executive assistant, but they do want to save money.
Horizontal Versus Vertical
The team began horizontally to avoid being pigeonholed, with the intention of learning from thousands of users and then verticalizing based on real usage patterns. Jean-Denis argues that model capabilities are converging, so differentiation based on features is harder to sustain. He suggests the unit of value in software is shifting from apps to anything a human can do with a computer, which abstracts away the app layer and may reduce its value.
Town’s approach to personalization works like this: because the company does not inspect individual user data, it analyzes aggregate usage metrics to identify power users, interviews them, and finds common automations. Prototype “routines” are then built and offered as templates to new users, identified by inspecting their inbox. Over time AI-assisted interviews and automated scheduling accelerate this process, producing a growing library of routines that increasingly customize the product. Jean-Denis contrasts this with the forward-deployed engineer model used by labs, where a human goes into a business and builds custom software. He argues that FDE approaches look strong today, with enterprises willing to spend millions per year, but that they degrade as models improve because the value of hand-built work depreciates faster than automated customization scales.
Why Town Over OpenAI and Anthropic
Alex raises the infrastructure-versus-application question: Town buys tokens from the very labs now building connectors and applications that overlap with it. Jean-Denis gives a short answer first: Town has a different view of the email and calendar experience, and that is hard for labs to replicate. But the more substantive answer is about customer needs.
The first argument is neutrality. Mid-market customers cannot predict which model provider will lead over the next three to five years. Model leadership shifts: some companies chose OpenAI and then needed Claude for coding for a period. Town positions itself as a neutral routing layer that uses whichever model is best for a given task, including for voice, video, slides, and open-weight models. Jean-Denis argues that businesses should think carefully about who routes their intelligence and reduce the risk of having picked the wrong partner.
The second argument is data ownership. Jean-Denis says the data, memories, and context accumulated in Town belong to the customer, and that customers should be able to leave easily. He believes that FDE-style approaches intentionally raise switching costs, while Town’s competitive stance is to make switching easy and win through product quality. The third argument concerns training: large customers are assured that Town does not train models on their data, because allowing operational knowledge to leak into model weights would effectively sell a company’s success to competitors.
Jean-Denis also describes a more forward-looking trust model: an internal mode in which agents mediate which information is available to employees. For example, someone could ask whether anyone in the company knows a contact at a prospect, and the agents of colleagues would be consulted to route the request securely. He argues that humans are slow routers and the future involves much less human-in-the-loop work, which requires companies to trust AI to mediate privacy and security within the organization. Alex adds that Rogo and Harvey are examples of vertical companies that OpenAI is now competing with, and that enterprises should avoid being beholden to any single lab.
Alex and Jean-Denis discuss the “Goldilocks zone” of relevance for labs. Coding is central to Anthropic’s revenue, so the labs cannot ignore it. Smaller horizontal markets are less central to labs’ focus, and labs have to make hard bets across healthcare, law, and coding. Jean-Denis expects that the SMB and mid-market AI-for-work market will become larger than coding, though the timing is uncertain. He believes the market will shift within twelve months.
Underhyped Outside the Valley
Alex proposes that Silicon Valley overhypes AI while the rest of the world underhypes it. Jean-Denis agrees that the diffusion is uneven and that tinkerers drive adoption. A small surfboard rental business used Town to compare photos of rented equipment and decide whether to release deposits, a use case no one would have predicted. Jean-Denis argues that most businesses lack the tinkerers, time, and engineering culture that exist in Silicon Valley. Even though Claude and ChatGPT have offered features for twelve months that could save small businesses significant time, users treat them as better search engines. The problem is that a text box is not enough; products must show the value and do the work on the user’s behalf. Jean-Denis sees the real unlock as integration and product marketing rather than intelligence: the models already have the capability, but the deployment at scale is the challenge.
Both note that people judge AI by its past performance. Someone who tried an early model two years ago may dismiss it, while the current generation is substantially more capable. Alex notes that computer use and browser use improved dramatically over a few months, moving from around thirty percent reliability to ninety percent, and that the challenge is communicating concrete capabilities. Jean-Denis gives an example: rather than announcing computer use, Town can offer to make event posters using a customer’s existing Canva subscription, then proactively suggest updating them with a new brand design.
Neutral Agents, Adversarial Commerce, and the Broken Ad Covenant
The final section considers how agents change distribution. Agent traffic has reportedly surpassed human traffic, and companies are choosing between adversarial and synergistic stances. Alex describes the old internet covenant in which users gave attention to ads in exchange for free content. That covenant is breaking as agents mediate browsing. Jean-Denis argues that neutrality is essential: a personal agent should be paid only by the user, and its incentives should be to give the best information without marketing distortion. Users retain taste and can ask their agents what is trending or advertised, but the agent itself must remain neutral.
Jean-Denis distinguishes between friction people enjoy and friction they do not. He likes browsing SoHo shops but wants toilet paper reordered automatically. Amazon’s advertising business faces pressure, but it can still win by creating experiences where users want ads, such as discovering new toys for kids. Alex suggests that a headless Amazon could be a large business, and that Walmart may benefit because it lacks a large advertising incentive. He also discusses Meta’s Muse as a free personal agent whose monetization may come indirectly through targeting on Instagram rather than ads inside Muse itself. The conclusion is that supply aggregation loses value when human attention can be automated, and that breakage models such as coupons and card benefits may erode as agents optimize them for everyone.