Why a $1M Startup Wasn’t Worth Building in the Age of AI

Summary

• Town abandoned a business with just under $1 million in first-year revenue because its distribution could not outrun incumbents. Greze thought year two could reach $2.5–3 million, but annual purchasing and weak word of mouth limited compounding while established providers adopted AI. It was “a good business before the AI era,” without the venture-scale trajectory the team wanted.
• Town’s second act starts with consumer-quality assistance and adds enterprise administration around it. Email and calendar supplied frequent, tangible problems; selling first to corporate gatekeepers risked optimizing for the check writer. Greze’s Dropbox lesson: a “world-class product” can move into enterprise, while starting with compliance rarely produces one.
• Cheap tokens expand the market into work nobody previously found economical to perform. Greze’s “token distribution machine” could monitor restaurant suppliers for about $1 a month, finding nothing useful in 11 months but saving hundreds or thousands in the remaining month. The pitch is a specific economic outcome rather than an abstract personal assistant.
• Town is betting that horizontal distribution can produce increasingly personalized vertical workflows. Interviews with power users feed reusable “routines,” while better models should reduce manual customization. Greze expects that approach to scale better over time than deploying people to build bespoke AI systems for individual enterprises.
• Model neutrality and customer ownership of context are Town’s proposed defenses against its suppliers. Greze wants customers able to move their memories, contacts and data elsewhere in roughly a day, accepting lower switching costs. Rampell sees a possible “Goldilocks zone” below labs’ immediate priorities; Greze thinks competition is a question of “when,” not “if.”
• The adoption bottleneck is turning existing intelligence into demonstrated, recurring outcomes. “You don’t need more intelligence,” Greze argues; users need products that identify useful work and perform it. Town also trusts agents with some internal information routing, replacing human coordination delays—the opportunity Rampell captures as “People are slow routers.”
• Agents threaten advertising and aggregation economics by automating the attention those businesses monetize. Greze argues for agents funded by users, while acknowledging free agents could monetize indirectly through better advertising elsewhere. Price monitoring, coupons and Chase or Amex benefits become easier to exploit; the pair expect pressure on businesses supported by customer inattention.

Deep dive

1. A working tax product still lost the venture race

• Greze’s original thesis was technically plausible: text-heavy inputs, financial statements and leases could become completed business tax returns, potentially within two years. After a year, revenue was just under $1 million; the team thought year two could reach $2.5–3 million. It was “a good business before the AI era,” but he no longer believed it would become venture scale.
• The binding constraint was distribution. Incumbents could automate preparation faster than Town could acquire millions of customers. Taxes sell once annually; recommendations arrive after filing, when recipients often are not shopping, then fade before next season. For a startup, “the clock is ticking” before established providers combine comparable technology with existing distribution.
• Leaving still felt like failure. Greze weighed the impact he wanted against employees’ opportunity costs and the difficulty of recruiting against liquid offers. He acknowledges that an inventive team might have found another route, but they lacked a convincing breakthrough idea. Treating departing tax specialists well made the decision emotionally clearer.
• Seasonality also forced a decision. Around October 15, Town had roughly a month when continuing and exiting remained possible. January partner commitments or outstanding customer promises would close that window. Waiting another year, Greze believed, would leave insufficient capital to change direction.

2. Consumer quality comes before enterprise administration

• Tax customers supplied the next problem: scheduling clients, remembering meetings and preparing answers to repeated questions. Exploration coincided with Opus 4.5 and increasingly autonomous coding agents. From December of the previous year through April, strong user response and a clear roadmap displaced abstract market-sizing exercises. Product-market fit made Greze want to “grab it with a death grip.”
• Rampell’s concern was who shaped the product: when the buyer differs from the user, satisfying the CISO can displace everyday usefulness. Greze’s Dropbox lesson was that a “world-class product” can be adapted for enterprises; beginning with enterprise compliance and infrastructure makes the reverse journey almost impossible.
• Rampell extends his iPhone analogy; Greze warns that Town could become not only like BlackBerry, but like Good Technology, the corporate iPhone email client he called “the ugliest app in the history of the world.” Users could instead choose the Gmail app, which combines convenience with corporate features.
• Greze expects adoption to move at human speed: assistants are still used by only a few million people, and it may take 6, 12 or 18 months for most people to experience the consumer magic and want it at work. The rollout starts with email and calendar, then earns permission to connect CRM and other tools. “They are only as smart as the data they have access to.” Larger deployments need audit logs, sensitive-information controls and CFO visibility into token spending. Town therefore targets SMBs in self-service experience while building security, compliance and financial controls for enterprises.

3. Cheap tokens make neglected work worth doing

• Greze calls Town a “token distribution machine”: discover places where computation creates value, then let users choose the outcomes. Preparing research for every 30-minute meeting would be an irrational assignment even for a human assistant; an LLM can do it for pennies, including checking what attendees recently said on X.
• His strongest example is restaurant procurement. Wholesalers periodically email spreadsheets or PDFs containing thousands or tens of thousands of prices. A part-time finance person should compare changes, including increases of 20–30%, but usually does not: the task is occasional, cumbersome and outside daily operations. Greze says a prompted monthly analysis can be created in five minutes.
• If nothing useful appears in 11 of 12 months, it might cost about $1 monthly; the useful month could save hundreds or thousands. The output is an actionable email about switching suppliers or reconsidering menu items. Rampell’s marketing translation: a restaurant owner may not want a “personal assistant,” but saving $500 annually for a $50 charge is understandable immediately.

4. A horizontal product can discover its own verticals

• Greze expects AI products’ capabilities to converge, weakening durable differentiation through features. His broader framing is that “the unit of value in software has shifted from the app to the computer”: agents can operate across what a person does with a computer, reducing value concentrated at the individual application level.
• Town’s answer is role-specific recommendations, although Greze says current models produce suggestions that remain too superficial. Rather than inspect customers’ private data, the team identifies highly active users through aggregate metrics and interviews them. Ten restaurant operators might reveal five common automations, which Town prototypes as reusable “routines.”
• That library might grow from five useful tasks to 15 while improving models gradually require less scaffolding. Greze thinks this can deliver “80/20 very personalized software.” Personalization must also span roles: one restaurant owner may manage finances, operate a vacation rental and organize summer activities for two children.
• He contrasts this with forward-deployed engineers learning a company’s operations and building bespoke software. For a business willing to spend about $5 million annually on OpenAI FDE work, that “looks pretty good right now.” His bet is that improving models favor Town’s automated approach over time, with SMB adoption feeding a broader learning flywheel.

5. Neutrality is Town’s answer to competing with its suppliers

• Rampell asks why connectors from OpenAI or Anthropic will not absorb Town. Greze concedes that superior email and calendar execution is an incomplete defense: a lab could choose to copy it. For companies with 500–1,000 employees, his stronger pitch is model neutrality over three to five years, choosing providers across general intelligence, voice, video and slides. He points to companies that chose OpenAI, then needed Anthropic after its models improved at coding.
• Customers should own the context accumulated around their businesses—data, memories and contacts—so thoroughly that moving to a better product should take roughly a day. “If I can’t have the best product on the market, that’s okay. Then it just means I lose.” Greze accepts lower switching costs and says training on customer information is outside Town’s business philosophy.
• Town’s Teams product also lets agents locate relationships across colleagues: ask who knows someone at a prospect, and agents query other agents. Greze says Town trusts AI to enforce internal information boundaries, requiring careful privacy design. A Slack request might take a day; Town answers instantly. Rampell’s framing: “People are slow routers.” Greze expects much less human participation in these decisions.
• Rampell proposes a “Goldilocks zone”: scattered SMB workflows may sit below labs’ immediate priorities, unlike coding. Greze agrees for SMBs today but challenges the protection over time. Business automation is larger than programming, making lab competition a question of “when,” not “if.” He expects the landscape at those companies to change within 12 months, though not under current conditions.

6. Adoption depends on showing people what to automate

• Unexpected usage illustrates the opportunity: an island rental business uses Town to compare photographs of surfboards and beach buggies for damage and decide whether to return deposits. Greze credits an enthusiastic internal inventor. The constraint is that most businesses lack someone with sufficient AI familiarity or time to invent these workflows.
• Retention strengthens when customers connect email, calendar and an important accounting system, then build one custom integration that feels magical. Greze wants to remove that final invention requirement. “You don’t need more intelligence. The intelligence is already there.” The challenge is delivering equivalent value to people who would never create an automation themselves.
• Rampell observes that early adopters can become skeptics by freezing their view around a disappointing experience two years earlier. He also says computer-and-browser tasks that were almost impossible six months earlier now succeed in 90% of cases, compared with about 30% before, over the last two and a half months.
• Rampell makes the marketing example concrete: rather than touting a “computer feature,” offer to make event posters in Canva using an existing subscription for 25 cents in five minutes, then repeat the task for future events, adapt to branding changes and eventually make promotional video. Greze’s broader principle is that “a product cannot be just a text field”; it must demonstrate useful outcomes.

7. Agents separate useful discovery from unwanted friction

• Rampell frames the conflict through Yahoo Finance: free stock quotes were supported by the user seeing brokerage advertisements. Agents disrupt that exchange. He sees another possibility in welcoming agent traffic as something like “the new SEO,” rather than treating every agent as an adversary.
• Greze says Amazon’s loss could be Walmart’s gain because Walmart lacks an advertising business and therefore has different incentives. He worries about compromised recommendations: if United Airlines advertises heavily with Meta, would a Meta agent still choose flights neutrally? His preferred arrangement is categorical: the agent belongs to the user, and “the sole source of funding for that agent is your wallet as a user.” Marketing should enter when the user wants it.
• Efficiency is not always the objective. Greze enjoys walking between shops in SoHo and observing other shoppers, while replenishing familiar toilet paper should be quick and cheap. Likewise, buying LEGO for his children can make advertisements for unfamiliar toys valuable. He sees room for discovery businesses, provided the experience earns the attention people choose to give it.

8. Automated attention weakens aggregation and breakage

• Greze offers a possible alternative around the free agent he calls “Muse”: it might never need advertisements inside the agent. Information learned there could improve Instagram targeting instead; wallet information and product searches could support purchases. In that scenario, monetization moves elsewhere while the assistant experience remains free.
• Amazon could lose a substantial advertising business, which Greze loosely sizes at “$80 billion, or what, $70 billion” annually. He thinks a “headless Amazon” could become enormous. His proposed response is a better shopping agent, potentially bundled into Prime, alongside advertising experiences people actively want. He is unsure whether Amazon can execute that response.
• Rampell’s existing example is CamelCamelCamel: he says 20 million Americans use the 18-year-old Amazon price tracker. A user can request an alert when an $8 Dr Pepper purchase falls to $6; the service emails rather than buys. Rampell’s inference is that agents could make such monitoring attractive even to people who presently have more money than time.
• The pair extend that logic to coupons and Chase or Amex benefits: AI can reclaim unused perks, but Greze says widespread optimization may force providers to change them and turn the arrangement adversarial. “The unused benefits model is dead.” Rampell’s closing argument is that supply aggregation becomes less valuable when the scarce resource it organized—human attention—can itself be automated.