JD Grèze on Building Town and Competing in AI Assistants

Harry Stebbings says he uses Town and welcomes its co-founder and CEO, Jean-Denis ‘JD’ Grèze. Grèze describes Town as an AI assistant connected to email and calendars that suggests background automations based on how a person organizes work.

The company first spent a year building AI tools for business tax preparation. Grèze says it found some product-market fit but not enough for the business he wanted to build, so the team reset and spent several months looking for another direction.

The new question was why no one had made a broadly accessible assistant that worked through email and calendars. Grèze says improvements in agentic models made more substantial workflows possible, and a prototype built in a few weeks received encouraging early feedback.

He contrasts products for technically inclined tinkerers with Town’s ambition to deliver value to mainstream users. His account of the prototype’s product-market fit is the founder’s assessment, not an independently verified measure of long-term retention.

Stebbings asks about competition from model providers and platform companies. Grèze says the prospect keeps him awake: he believes assistants are a high priority for major technology companies and that Town is building in the middle of a crowded market.

For an early startup, he says, defensibility matters less until the product has deep product-market fit. Town’s immediate challenge is to find an experience that many people will use and pay for; he says both his product and rivals still have work to reach mainstream users.

One longer-term bet is a network effect between agents. Grèze describes a Town feature in which one person’s assistant can ask a coworker’s assistant for relevant information. If a whole team uses the system and finds it valuable, he believes switching could become harder.

He lists other possible defenses: customized models, accumulated context and connections, existing distribution, or the device through which people access an assistant. These are competing theories, not established advantages for Town or any particular company.

Grèze sees messaging distribution as a potential advantage for a company such as Meta. He also notes the theory that an assistant could become a main entry point to digital services, giving device owners an advantage; neither outcome is assured.

He expects a person to have only a few user-facing entry points rather than choose a different interface for every task. A main assistant could call specialized tools behind the scenes, while work and personal information may need distinct boundaries.

Workplace ownership and privacy could require different data silos or providers even if a person uses one device. Grèze uses a legal assistant as an example: the user may talk to a general interface while specialized services handle privileged work in the background.

Asked for a surprising five-year prediction, Grèze says people may trust agents to decide what information can be shared without asking for approval each time. He imagines an assistant helping friends plan a trip while declining to reveal a person’s medical history.

He compares this with the discretion people already exercise when deciding what to tell others. Grèze expects models to become better at respecting sensitive boundaries, but this is a prediction, not evidence that current or future agents can reliably protect every private detail.

He argues that access to more relevant context can improve an assistant, while acknowledging the importance of privacy and people’s control of their data. In practice, more access also creates security and disclosure risks that a system must manage.

His workplace example is a salesperson looking for an introduction at a large company. Instead of asking a broad Slack channel, one agent might consult other employees’ agents and identify a colleague with a suitable relationship, while respecting limits on what may be disclosed.

Grèze predicts that users will become more comfortable allowing this kind of limited sharing as models improve. Stebbings asks how much room there is for error, especially when an assistant books the wrong thing or exposes sensitive information.

Grèze recalls a human accidentally replying to an entire company with a message intended for a small group. He argues that people also make disclosure mistakes and predicts models will eventually make fewer of some types; the anecdote does not establish a comparative error rate.

Stebbings raises goal-seeking behavior, citing an anecdote about an agent attempting unwanted purchases. Grèze says he does not have a settled answer for how long an agent should run or how much autonomy it should be given.

His preference is for humans to set goals and resource budgets and monitor whether actions fit their values. He allows that other agents may help estimate return or review actions, but does not assume the same goal-seeking agent should be the only monitor.

Town chooses models on users’ behalf, Grèze says, because most customers do not want to track which provider is best for each task. The team weighs the result and cost across text, image and voice tools, though open-ended tasks make success difficult to assess in advance.

Routing has a user-experience complication: switching model families can change an assistant’s voice or writing style. Grèze says users notice sudden changes in verbosity or phrasing. For customer-facing output, he therefore values consistency as well as raw model capability.

For internal reasoning or tasks where intermediate traces are less visible, he sees more freedom to choose the best model for the job. He contrasts that with coding, where the main question may be whether the code works, although implementation quality still matters.

On cost, Grèze says many startups currently subsidize some usage while hoping model prices and efficiency improve. Routine tasks such as labeling email do not require the most expensive model, in his view, and may eventually run on cheaper or open-weight systems.

The unresolved question is what share of a company’s workload remains near the expensive frontier. Grèze offers possible proportions as scenarios, not measurements; the answer affects whether lower costs on routine tasks are enough to support healthy margins.

Town still uses frontier-provider models for much of its work, he says, partly because users build complex workflows. It could devote engineering time to replacing some models with cheaper alternatives, but the team currently prioritizes product improvements and growth.

Grèze frames this as an allocation of scarce engineering hours: reducing cost competes with building integrations, improving workflows and increasing adoption. He says Town’s economics could be better but considers product growth the more immediate constraint.

One adoption signal is whether a team has a technically curious person who builds shared integrations, routines or skills. Grèze says these power users can help coworkers find useful workflows even though Town also wants a simple experience for people who never configure anything.

He sees opportunities in roles that have received less attention from AI vendors, including executive assistants, chiefs of staff, recruiters and some finance or HR staff. Many of their daily workflows still involve email and scheduling, he says.

Grèze emphasizes time to value for one person. Meeting preparation and follow-up can feel useful immediately without configuration, in his account; once users see a benefit, they may explore more workflows and invite colleagues.

Asked how investors should assess new assistants, he outlines three Town bets: a named assistant with a distinctive user relationship, collaboration across a team, and preprocessing information to create useful context before a user asks for something.

That preprocessing might help an assistant understand a person’s projects, organization and working relationships. Grèze treats it as a potential product advantage, while the benefit and privacy implications depend on how the data is accessed, stored and controlled.

For a local competitor, he says an investor should ask why it can win: distribution, regional privacy needs, or a market overlooked by global providers might be possibilities. He cautions that maintaining competitive capabilities is expensive.

A startup also has to keep up with fast-moving model providers and tools. Grèze says a compelling distribution story is not enough if the assistant cannot perform tasks users consider essential; the size of the required research and development budget is a serious risk.

The biggest unexpected difficulty, he says, is market speed. Features that once gave a company time to learn might now be copied in weeks. Teams can build at machine speed, but learning what people actually value still requires human feedback.

Grèze says that makes it harder to accumulate an uninterrupted lead from successive product insights. Established competitors can move quickly too, sometimes combining a startup-like pace with far greater financial and distribution resources.

He still sees a large market because many prospective customers have not tried any assistant beyond a general chat product. Competition creates pressure to execute, but companies can choose different customers, business models and product strategies.

He characterizes Town as focused on paid work use cases and collaboration across teams, while seeing Instinct as pursuing a different early approach. These are his outside observations, not confirmed descriptions of a competitor’s plans or economics.

Grèze pays close attention to GrokBot because he believes it may target customers closer to Town’s. He says a connection to X can help early distribution while brand associations might deter some prospective professional users; both effects are his assessment.

On Apple, Grèze argues that an assistant needs access to cloud information and that a strong on-device privacy orientation can involve capability trade-offs. These are a competitor’s views, not a verified description of Apple’s roadmap or the limits of its technology.

He acknowledges Apple’s device distribution and expects its assistant experience to improve, but speculates that it could initially lag other products on some capabilities. He says the outcome depends on how Apple responds to changing interfaces and company priorities.

Asked about security, Grèze says business assistants must handle access controls carefully. He expects humans to continue reviewing critical security-sensitive code while more code is generated and tested by machines; this is his view of the trajectory, not a substitute for security review.

He compares the emerging situation to industries that developed safety practices after harmful incidents. His point is that testing, controls and regulation may evolve as risks become clearer; he does not suggest that failures are harmless or inevitable in every deployment.

Grèze defines a successful customer primarily as someone who keeps paying because the product delivers value. He warns that maximizing token consumption can produce activity whose benefit is unclear and eventually cause users to leave.

A meeting briefing can have very different value to two people while using the same number of tokens. He says the assistant should help customers understand where their money goes rather than assume more computation is always a better outcome.

Grèze says Town began warning users about expensive recurring routines so they could reassess them. His goal is for customers to trust the product to help them use AI efficiently, although he says that trust is harder to express as a single metric.

He describes different pricing tiers and usage-based charges after plan limits. According to Grèze, the lowest-priced plan is the most subsidized, while a higher tier has better unit economics. These are company statements about a point in time, not independently audited margins.

Given a choice between a very large number of lower-paying users and fewer high-paying users, he favors the broader base. He expects work customers to discover additional valuable uses over time and potentially increase spending; this is his growth hypothesis.

He gives an example of a recruiting firm that he says can take on an additional client after automating work with Town. The illustrative revenue and subscription costs are his account of one use case, not proof that customers generally earn a particular return.

Grèze expects more natural, conversational voice interaction to be important. Asked about ElevenLabs at a reported valuation, he says Town uses the service and likes its quality but is uncertain how long a premium over cheaper alternatives can persist.

He says voice costs would matter more at much larger scale and that, once quality is sufficient, price could become decisive. He is not making a recommendation about ElevenLabs or establishing the value of the company.

His deeper margin concern is the fraction of tasks that continue to require frontier models. An assistant company can be both a customer and competitor of a model provider, potentially depending on an expensive supplier that also offers a competing product.

Grèze describes this as an end-state problem contingent on Town gaining substantial scale. The actual economics will depend on future model prices, workloads, competition and customer willingness to pay; he says he does not yet know how the mix will develop.

He acknowledges that Town competes with some of its model suppliers today. He argues that the near-term opportunity is making AI useful to people who do not know what to ask of a general chat interface; his company’s claimed advantage is a guided product experience.

Grèze says more than 15% of people who try Town become paying users. This is a company-reported conversion figure without details of cohort, measurement period or independent verification, so it cannot by itself establish durable product-market fit.

He attributes part of the experience to requiring new users to connect email and a calendar, allowing the product to suggest tasks and automations. Stebbings questions whether that insight is distinctive and whether people will be willing to grant access.

Grèze says some users leave at the connection step and estimates the share at about 30%. He considers that friction a trade-off: the integrations enable the onboarding experience he wants, but they also require users to decide whether they trust the service with sensitive data.

The team has discovered an unexpected personal use case among parents managing school emails, activities and schedules. Grèze says some families find the product valuable, but serving them raises a strategic question about how to allocate limited marketing and product resources.

Business customers may offer greater spending and expansion within an organization, while parents may recommend the product through their communities. Grèze says Town is weighing those trade-offs rather than assuming it can pursue every segment equally.

Asked for revenue, Grèze declines to disclose it. Stebbings also raises reported fundraising and valuation details later in the conversation; Grèze does not confirm or deny the reported valuation, so the discussion should not be read as confirmation of a financing.

Grèze argues that a large assistant company could emerge if a startup becomes a common entry point for digital work and reaches tens of millions of users. That is a conditional market scenario, not a demonstration that any particular valuation is justified.

The speakers disagree about press strategy. Stebbings favors separating fundraising and operating milestones into distinct moments; Grèze says he prefers attention to come from people who genuinely value the product and does not want to manufacture a fundraising story.

As an angel investor, Grèze criticizes announcing a headline valuation that applies to only a small part of a round without making the rest clear. He says this can mislead employees about dilution and offers; he speaks about practices he says he has seen, not a confirmed Town transaction.

In a hypothetical bullish case, he says ten million paying users with continued growth could support a very large Town business. He reports more than $700 in annual revenue per user at the time of the interview; the metric and the valuation scenario are not independently verified.

He would like future board expertise in finance and operations, particularly as the company buys compute and manages token costs at scale. He anticipates that later-stage investors will expect metrics that the company can defend.

Stebbings asks why Town does not simply subsidize usage much more aggressively. Grèze says he considers it, but believes charging business customers is important evidence that the product delivers value and gives customers clarity about future costs.

During an early beta without pricing, he says, some users consumed thousands of dollars of compute in a month. His point is that unrestricted use does not automatically show whether the activity is valuable to the customer; he says the team discussed usage with those users.

He allows that acquisition may be subsidized further, but wants a real paid business once teams adopt the product. He also questions an advertising-supported assistant model because ads could create incentives that conflict with a user’s interests; this is his business view.

On hiring, Grèze describes an unusual exception: if a trusted high-performing employee has worked extremely closely with an outstanding candidate, Town may focus on selling the opportunity rather than repeating a full technical interview. The candidate still gets time to evaluate the team.

He reports a high annualized spend on AI tools per engineer and names several products used for coding and routine fixes. The figure is his description of Town’s spending at that time and may not be representative of other companies.

He frames engineering and compute spending as an expected-return decision. In an illustrative business, an engineer who becomes more productive with AI could justify an additional hire if the value created exceeds the cost; the arithmetic is hypothetical, not an employment forecast.

At Town, he says customers are requesting integrations, audit logs and other features. If additional compute helps the team deliver valuable work faster, he is willing to spend, subject to funding, hiring and revenue constraints.

Asked what excites him over the next decade, Grèze first names time with his children. He also hopes AI will reduce routine work and give people more opportunities, while recognizing why rapid change can feel unsettling.

He gives examples ranging from drug development to physical-world construction and transport as hopes for the technology. These are optimistic possibilities rather than established outcomes. Stebbings thanks him for addressing the competitive and economic risks as well as the potential.