Town stays neutral on models; enterprises should hoard data and memory
With no large model guaranteed to dominate in three to five years, Town stays neutral, keeping data and memory ownership with enterprises so they can always switch providers.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Tax software has a fatal flaw: you can only sell it once a year
Automating tax returns looked promising at first—lots of text and structured financial data, exactly what LLMs excel at. The team spent a year turning ‘financial data plus documents into tax forms’ into a working product. But the business model fell apart: tax returns sell once yearly, customers don't evangelize the tax experience, and there's no daily active user loop to create word-of-mouth virality. Meanwhile, established players can apply the same AI techniques to their existing customer base faster than any startup can steal market share. Even projecting $2.5–3 million annual revenue by year two wouldn't sustain venture-scale ambitions, so the team moved on.
— Jean-Denis GrezOpus 4.5's release gave the product instant product-market fit
After abandoning taxes, the team explored four directions before focusing on a simple insight: why hasn't anyone built truly usable agents for email and calendar? Former tax customers kept complaining they'd forget context before meetings or field the same questions repeatedly from clients. Everyday knowledge work had never been seriously tackled as an AI product. Then Opus 4.5 launched, showing coding agents could execute with greater autonomy. The insight plus the model's leap in capability created near-instant PMF. From a closed pilot in December to April, the product grew 1–1.5% daily, customer requests flooded in, and by then the team felt ready to go public.
— Jean-Denis GrezCheap tokens make previously uneconomical work suddenly worth automating
Jean-Denis frames Town as a ‘token allocation engine’. As tokens become cheaper, long-tail work that was economically unjustifiable for humans and too expensive for AI alike suddenly becomes worth automating. Consider a restaurant owner: suppliers send price quotes quarterly, often with 20–30% hikes, but the owner is too busy to compare them manually—it's too sporadic to justify hiring someone to monitor. With Town, a manager sets up a monitoring agent that runs monthly and generates alerts in five minutes; it might find nothing for eleven months at $1 per month, but the one month it catches a price spike can save hundreds or thousands. The real product work isn't the technology—it's convincing users that this task is now economically viable.
— Jean-Denis GrezStaying neutral on models beats betting wrong on one vendor
When asked how Town differs from OpenAI or Anthropic's built-in connectors, Jean-Denis centers his answer on neutrality. No one can predict which model vendor will dominate in three to five years. If a company locks its product into one model's ecosystem, a shift in capabilities or pricing—or one vendor's stumble—leaves the customer defenseless. Some enterprises have already hedged by adopting multiple vendors: betting on OpenAI for language work but discovering Anthropic's strength in code, then forced to maintain two vendor relationships. Town positions itself as a ‘neutral app layer’ that picks whichever models win at each task—voice, video, slides—and commits never to train on customer data. Once a model vendor ingests your business know-how into their weights and resells it to your competitors, your competitive advantage has moved to your rival's balance sheet.
— Jean-Denis GrezAI can route corporate information at speed with proper permissions
Town is building an internal CRM mode: when you ask ‘does anyone know this prospect's finance director?’ your agent queries your colleagues' agents directly, rather than waiting for Slack messages and hoping people volunteer information they should share. Jean-Denis calls humans ‘slow routers’. AI intermediaries can make permission and privacy judgments and return answers in milliseconds. He grants this needs meticulous privacy design, but argues it's inevitable: enterprises must adapt to fewer humans in the loop, and the economic gains will dwarf today's visible risks.
— Jean-Denis GrezCoding is make-or-break for large models; enterprise automation is bigger, just slower
Contrast Anthropic and OpenAI's stakes in the coding market—where they can't afford to lose—against their relative caution in the SMB automation space. Jean-Denis argues the SMB market is larger overall, just on a different timeline. Coding is already a nine-figure revenue opportunity for large models; they have everything to lose. But ‘AI helping enterprises do more things better’ won't move their quarterly earnings yet. He predicts rapid acceleration within twelve months. Large model companies are likelier to concentrate bets on outsized verticals—medicine, law, coding—where a single point of failure is catastrophic, leaving fragmented long-tail SMB scenarios to companies like Town willing to grind on niche needs.
— Jean-Denis GrezAmazon's advertising business could become a liability in the agent era
As the conversation turns to how agents consume supply and how e-commerce platforms should respond, Jean-Denis flags Amazon's ad business as a potential liability. Amazon's enormous ad revenue depends on customers browsing pages to see ads. Once purchasing decisions move to neutral agents, that flywheel breaks. Walmart faces no such handicap—it has no advertising empire to defend. Jean-Denis cites Muse, a personal agent, as an example: it doesn't need to embed ads; it feeds preference data to Instagram and similar entertainment platforms for more precise ad targeting and closes the sale there, completely decoupling monetization from the agent experience. Pure aggregation plays devalue in the agent era. Amazon's real countermove isn't shutting down Muse; it's building a better one.
— Jean-Denis GrezPrice comparison sites predict the divide in agent economics
Compare this to Camel Camel Camel, a price-tracking site with eighteen years of history and 20 million US users. The world has always split two ways: people with more time than money will spend two hours hunting coupons and comparing prices; people with more money than time won't. But once agents handle all this errand-running free of charge, the second group will adopt them immediately. Jean-Denis calls these models—credit card rewards, membership perks—instances of ‘breakage models’ that rely on most people never bothering to redeem. That assumption is about to fail. His conclusion circles back to the thread linking all eight points: aggregation plays gain value because human attention is scarce. When agents automate attention itself, the negotiating power of pure supply plays erodes systematically.
— Jean-Denis GrezIn their own words · checked verbatim
What if OpenAI is not the best model provider for the next three to five years? And so the first thing is we're just like, look, we're neutral. We will use whatever supplier we think is the best for you in the moment.
Jean-Denis Grez0:00
I think of our product as a token allocation machine. We're trying to give options to the human about where they could apply tokens.
Jean-Denis Grez23:13
But the interesting thing is the one month out of a year where it finds something, it's going to find like hundreds, if not thousands of dollars for you.
Jean-Denis Grez25:14
I think the FDE approach, you know, if you're an enterprise that's willing to spend like $5 million a year paying open AI for FDs, like it looks really good right now. But I think over time, as the models get better, the FDE approach actually gets much worse.
Jean-Denis Grez33:36
Humans are slow routers.
Alex Rampell41:01
And people that have more money than time, they're not going to use camel, camel, camel. People that have more time than money, they are.
Jean-Denis Grez1:00:50
I just think supply aggregation is a lot less valuable than it used to be very valuable because human attention was limited. And now, supply aggregation is a lot less valuable because human attention can be automated away.
Jean-Denis Grez1:01:28
Figures
| Typical annual enterprise FDE spending | approximately 5 million USD per year | 33:36 |
| Camel Camel Camel US user base | 20 million | 59:49 |
| Restaurant supplier price increase trigger | 20%–30% | 25:14 |
Glossary
- FDE (Forward Deployed Engineer)
- An engineer deployed to a customer site to develop and maintain AI solutions customized to their operations.
- Jevons Paradox
- An economic phenomenon where efficiency gains increase total consumption rather than reducing it.
- PLG (Product-Led Growth)
- A strategy where product experience itself, rather than sales teams, drives user acquisition and retention.
- breakage model
- A business model that depends on most customers not redeeming available benefits like discounts, rewards, or memberships.
How to listen
AI application founders, enterprise IT and AI procurement decision-makers, and investors wanting to understand model-neutral strategy.
The opening anecdotes about past work at Plaid and small talk can be skipped without losing the core argument.