Jean-Denis Greze on Leaving Tax Software to Build an AI Assistant

Andreessen Horowitz partner Alex Rampell interviews Jean-Denis Greze, founder and CEO of Town and former CTO of Plaid. Rampell says his firm invested in Plaid and later became a Town investor; he is also a Town user.

Rampell recalls being surprised to hear Ben Horowitz praise Town, because he knew Greze had started a small-business tax company. Greze explains how the business changed direction.

After leaving Plaid and spending time with his family, Greze looked for a business-to-business market where AI could change the product experience. Tax appealed because language models could process text alongside financial statements and other documents.

His team concluded that preparing a small-business tax return from financial records and documents could be automated within two years. They spent a year building it and, he says, the product worked, but reaching millions of businesses economically was much harder than expected.

The business made just under $1 million in its first year and the team expected $2.5 million to $3 million the following year. Greze says those would have looked like good numbers before AI, but he no longer believed the company could reach venture scale quickly enough.

His concern was that incumbents could use the same technology to automate tax preparation for their existing customers before a startup gained enough market share. Startups move fast, he says, but established providers have distribution advantages.

Tax is generally sold once a year, limiting both purchasing opportunities and word of mouth. Customers rarely discuss their tax preparation, and when they do, the listener may not be looking for an accountant until the following year.

Greze compares the constraint with education technology, where a founder friend found only a few annual windows to sell to schools. The team concluded it needed a more fruitful market to reach millions of people.

Walking away felt like failure, Greze says, but founders must ask why they are building a company. He and his co-founder wanted impact and needed to believe that the opportunity justified the team’s work and opportunity cost.

He acknowledges that they could never know for certain whether another idea might have rescued the tax business. They could, however, treat affected employees well, provide severance and help them find new roles.

The tax calendar imposed a decision window. They could not responsibly pivot after committing to customers’ returns, and Greze says they had roughly a month around October 15 to choose whether to continue for another year or change direction.

The team explored several ideas. One question stood out: why had nobody built a strong AI assistant around email and calendars? Tax customers had asked for help scheduling meetings, preparing for them and drafting answers to recurring questions.

These were ordinary knowledge-work tasks, but the team saw room to improve the experience. Greze says more capable agentic models arrived at the right time, and Town found product-market fit quickly.

Instead of starting with a precise estimate of the total market, the team followed strong user demand. Email and calendars connect many externally facing roles, while customers’ requests revealed a deeper product problem than simply attaching a language model to an inbox.

Greze says the product grew roughly 1% to 1.5% daily during its early period, and customer requests made the roadmap clear. By April, the team felt ready to move beyond its closed beta.

Rampell asks about product-led growth versus selling directly to enterprises. If the person buying software differs from the person using it, he argues, the company can optimize for the buyer’s requirements while neglecting the user experience.

Greze draws on his experience at Dropbox: building enterprise capabilities around an excellent user experience is difficult but possible. In his view, starting with compliance and IT requirements and later trying to make an exceptional user experience is much harder.

Town initially prioritized product quality while establishing a foundation for security and privacy. Greze expects executives who enjoy consumer-grade AI assistants to want a comparable experience at work.

The company now has large customers, he says, and must make enterprise security, privacy and administrative features repeatable across many deals. He considers that a more manageable problem once the underlying product experience works.

Rampell compares enterprise AI adoption to the iPhone: employees and executives experienced a compelling consumer product and then pushed companies to support it. Greze agrees but warns that an enterprise-ready product must also match the quality people expect from consumer apps.

Greze thinks assistants are still early in adoption. Their usefulness at work depends on secure, privacy-preserving access to business data and APIs, so enterprise rollout can move at the speed of human trust and organizational processes.

Town’s proposed journey begins with email and calendar, where it might help prepare meetings or update a presentation. Once users see value, the company hopes they will connect additional systems such as a CRM.

To win enterprise adoption, Greze says, the product also needs audit logs, controls over privileged data and cost visibility for finance teams. He expects small and midsize businesses to adopt faster because they often have fewer such requirements.

Rampell describes asking his assistant about a dinner that appeared in an email but not his calendar. He sees a potential expansion of the assistant market: inexpensive automation can perform tasks that people would not assign to a human assistant.

Greze calls Town a “token allocation machine”: it suggests places where users could apply AI and lets them decide whether they want the result. As tokens become cheaper, tasks that were previously uneconomic can become worthwhile.

Meeting preparation is one example. A language model can assemble notes and research on attendees for far less than the time a human assistant would spend preparing for every routine meeting, Greze suggests.

He gives a restaurant example: suppliers send spreadsheets or PDFs containing thousands of changing prices. A small business may lack the staff or time to compare those prices regularly, even though doing so could reveal savings.

Greze estimates that a monthly AI routine might cost roughly $1 per month and occasionally uncover hundreds or thousands of dollars in savings. These are illustrative estimates, and he says the product must first show the owner that the task can be automated.

Rampell observes that a restaurant owner may not want an “executive assistant” but may want lower ingredient costs. This raises a product marketing challenge: communicate a concrete outcome rather than an abstract category.

They discuss horizontal and vertical products. Rampell describes Town’s approach as starting broadly, learning from many users and then applying those lessons to specific industries rather than committing to one narrow use case upfront.

Greze expects many work-oriented AI harnesses eventually to have similar capabilities. His bet is that better models can customize an experience for each user, whether that person is an electrician, venture capitalist or executive assistant.

He says the model alone currently produces suggestions that can be too superficial. Town looks at aggregate usage metrics, talks with power users and turns recurring needs into prototype routines; he says the company does not inspect user data for this research.

For example, interviews with restaurant owners can reveal common tasks, including supplier comparisons. Town can offer templates to a new restaurant user and expand the library as it learns from more customers.

Greze imagines an assistant serving multiple roles for the same person: restaurant owner, finance manager, vacation-rental host and parent. The goal is a more customized set of tools, assembled with AI rather than a separate application for every role.

He contrasts that with forward-deployed engineers who enter an enterprise, understand its workflow and build bespoke software. That approach can work for a large customer today, he says, but he bets that increasingly capable models will make a more automated approach scale better.

Rampell asks why customers should choose Town over OpenAI or Anthropic products with their own connectors. He discloses that he is both an investor and a power user; the labs are also suppliers of the models Town may use.

Greze acknowledges the risk of competing with suppliers that have lower model costs. He says Town’s current email and calendar experience is one advantage, but a major lab could choose to invest heavily in the same area.

His broader argument is model neutrality. A company cannot know which provider will be best over the next several years, he says, or which will lead in voice, video or slides. Town aims to choose among suppliers on the customer’s behalf.

Greze also argues that businesses should own the data, memories and context created through an assistant, and be able to move them to another product. He presents easier switching as a customer benefit even though it creates competitive risk for Town.

He says Town does not train models on customer data and argues that businesses should consider how suppliers use the information that makes them effective. These are the founder’s descriptions of his company’s approach.

For teams, Greze describes a feature where agents can help find a colleague connected to someone at a prospective customer. He says privacy and permission design are essential when agents mediate access to information across a company.

In his view, asking agents to route an appropriate request can be faster than posting to a company chat and waiting for people to respond. He predicts less human involvement in routine decisions, while recognizing the need to earn organizational trust.

Rampell notes that model providers are moving into applications, especially large markets such as coding, healthcare and law. Greze agrees that smaller business workflows may receive less immediate attention, but expects competition to intensify as the market grows.

They turn to adoption outside Silicon Valley. Greze describes an unexpected customer: an island rental business uses Town to compare pictures of surfboards and beach buggies for damage when deciding whether to return deposits.

He credits a technically curious employee for discovering that workflow. Many businesses lack someone with the time or inclination to experiment, he says, so a text box alone will not show most people what the technology can do.

Greze says users often connect email, calendar and an important business system, then create one valuable custom integration. His goal is to make that value visible without requiring each customer to invent the integration themselves.

Rampell says people may judge AI by an older experience and miss rapid improvements. He recalls showing a professor a current AI-generated update to her old dissertation; he uses the story to illustrate how much perceptions can lag technology.

Greze says a product should turn new capabilities into concrete offers rather than announce technical features. He imagines an assistant making event posters using a business’s existing Canva subscription, then offering to repeat and adapt the work for future events.

Rampell asks how agent traffic changes the advertising-supported web and shopping. A website that once exchanged free information for human attention may see that bargain challenged when an agent retrieves the answer instead.

Greze argues that users need an agent whose incentives are aligned with them. He worries that an agent funded by advertisers could steer choices, although users might still ask it to surface advertising or popular products when they want discovery.

He distinguishes enjoyable shopping friction, such as walking through stores and discovering styles, from chores like repeatedly ordering toilet paper. Some experiences benefit from discovery; others are valuable precisely because they save time.

Greze suggests retailers could offer both an efficient agent-friendly purchase path and a separate discovery experience with advertising. He also speculates that an agent might be monetized indirectly through another service; he presents these as possible business models, not settled outcomes.

Rampell cites CamelCamelCamel, which alerts shoppers when an Amazon price reaches a target. He argues that agents could make price tracking and coupon optimization available to people who would not spend their own time doing it.

Greze closes by arguing that aggregating supply may become less valuable if agents can automate the search that once demanded human attention. Both speakers expect changing incentives and distribution as agents become more widely used.