The best consumer founders know the game. 35K+ of them read Consumer Startups every week.
Stay ahead. Get the playbook behind today's breakout startups.
Read time: 4 mins 41 seconds
We’ve been promised an AI assistant that runs our day for over a decade. Siri, Alexa, Google Assistant. Each one arrived with a big keynote and then turned into a glorified kitchen timer.
This AI personal assistant wave feels different, as the agentic capability is improving at an exponential rate. One product I like lately is Town. It pre-drafts email replies in my voice and sends a short briefing before each meeting: who I’m talking to, what we last discussed, what to know walking in. Even though the scope of the work has been somewhat limited, for the first time I could clearly see the value an AI personal assistant could bring to my life.
What’s even more interesting is that that Town is in one of the most dangerous and competitive neighborhoods in tech. Microsoft and Google own the calendar and email interface. This AI + calendar + email idea also sits squarely in the blast radius of OpenAI and Anthropic.
As a fan of the product and a fellow AI founder battling the competition and moat question on the daily, I was excited to chat with Town’s founder, Jean-Denis Greze, and dig into the questions that are top of mind for a lot of AI founders.
The pivot before the pivot
Jean-Denis spent seven years at Plaid, eventually as CTO, and scaled its engineering team from around 20 people to 350. Before Plaid, he ran engineering at Dropbox.
His co-founder, Tony Vincent, was Director of Applied AI Product at Google and Head of Design at Dropbox, and has sold two startups.
They started Town in late 2024 with a $18M seed from First Round to automate taxes for small businesses. The technology worked and there was early traction, but the market didn't have a venture-scale path. It would have made a fine bootstrapped business, not a billion-dollar one.
They took a hard look at the numbers and made the decision to pivot, before they decided what to pivot into.
They built a few prototypes against different hypotheses. The breakthrough insight came from watching their own tax team work. These were sharp people, and most of their day wasn't tax prep. It was moving data into spreadsheets, booking meetings, digging facts out of email.
"Very smart people in non-tech industries still aren't using AI to do much of anything. The tools existed. The behavior didn't.”
AI + Email + Calendar
With this insight in mind, they started prototyping what AI plus email, then calendar, could look like towards the end of 2025.
They built it in two weeks. Within days, 20 people were using it daily, all by word of mouth.
They started to witness the real impact their two week old product was having on people:
- A CFO drafted 120 individualized investor emails in an afternoon.
- A CPA and working mom used Town to handle client research and emails, and her kids' school calendars.
- A founder who observes Shabbat mentioned his Friday offline pattern once, in passing. Town built him a Saturday-night briefing without being asked.
It looks like an assistant, but underneath it targets a handful of common jobs, which is why founders, salespeople, recruiters, chiefs of staff, and inbound businesses like plumbers all picked it up.
However, the big insight came when they realized the real value is that the Town can be a wedge to help more people adopt more AI workflows in a proactive, organic way.
"We're a recommendation engine, for normal people, of AI workflows."
Town ingests your inbox, calendar, and connected apps, watches what you do, and offers to automate it.
I could do this for you, want me to?
You click yes and get an automation built around how you actually work. If the model can't do something reliably, it doesn't suggest it.
Over the past 6 months, Town has grown to tens of thousands of signups and 5,000+ engaged users that rely on Town to automate their email and calendar workflows.
They have also raised a $55M Series A led by a16z to further expand on this vision.
Building an LLM product
Jean-Denis and I also jammed on product building and the unique opportunities and challenges that come with building an LLM product.
1/ The opportunities
Tool traces are your analytics. For privacy, Town doesn't watch user sessions. But the trajectory of an agent, every model call and tool call, shows you where it's struggling.
For example, in one calendar request, Jean-Denis counted 74 LLM calls and 28 tool calls, 17 of them to the calendar, with the model retrying the same call because of scheduling conflicts. That pointed straight at a fix: rebuild the calendar tool so the first call returns the conflicts and a few open slots.
"What's different about LLM products that I hadn't appreciated is that high-level tool traces point you to the hotspots."
You can point the LLM at your own roadmap. Town has a bug button. Most people don't share their session, which is fine, but the written reports are useful, and a lot of them are really feature requests ("I wish it booked restaurants").
Jean-Denis runs an LLM over the support tickets to group them and estimate how hard each would be to build. By the time a PM or engineer picks one up, some of the work is already done. His framing is that great products still come from human judgment and taste, and the question is how much AI can make you faster at it.
2/ The challenges
Reliability is the product. The worst failure isn't a clumsy draft, since people forgive those. It's being confident and wrong.
"It tells you the other person is free, you approve it, they're not free, and you think, what the f***, your one job was to get that right."
The fix is in careful context window management and tool design for specific scenarios. However, all users would eventually encounter challenges with LLM products. The more people trust Town, the stranger the tasks they hand it, and the more it breaks.
Knowing what it can't do. Sometimes Town might offer to do things it can’t, like logging into a specific third-party tool. The key is teaching the agent to flag low-confidence or out-of-scope tasks, and giving people a quick way to check its work. For scheduling, that could be as simple as showing your calendar next to the time it suggests.
The interaction model. This is also a common challenge among all GenAI startups. For Town, they know roughly what people want, an automated to-do list that tracks your goals and nudges you forward, but the right abstraction is still a work in progress.
3/ Delight-driven growth
Town's growth has come almost entirely from word of mouth. Many users are quite obsessive and loud about their love for Town.
The framework he uses to think about growth is building significant value for users while making the users feel proud and joyful to share the product.
For example, One thing I love about Town is that every AI assistant gets a custom avatar wrapped in a share card.
"If people love a product and they're proud of it, they feel cool for having found it."
Let people show how they use it, make that look good, and the act of sharing makes them look like they have taste.
4/ Building in the blast radius of AI labs
How to survive in an era where OpenAI and Anthropic seem to be swallowing everything alive is probably top of mind for most AI founders. I know I talk to my cofounder and other AI founder friends about this at least once a week.
Jean-Denis is calmer about it than I am. His case comes down to two things: the market is too young to be cornered, and the moat was never the model.
Market is too early – vast majority of knowledge workers still don’t automate anything with AI. The market is over a billion people, and nobody is close to winning. We’re in the MySpace era, before Facebook shows up. Even when a Facebook does arrive, there’s still room for WhatsApp, Instagram, and TikTok. If you capture 5% of this market, you’ll be huge.
Moats – models are trending commoditization so they can’t be the real moat. You either build a real one, like a network effect or a data flywheel, or you stay ahead on product speed. Incumbents who have the right to win in the space like Microsoft and Google are not moving fast enough to have the best product.
🌶️ My take on Town
A few days ago, Fryd Wiatrowski, founder of Viktor (an AI agent that lives in Slack), published a viral article titled "Anthropic killed Viktor.com. A post-mortem." He wrote it after Anthropic launched Claude Tag, a direct competitor to Viktor.
The title was mostly clickbait. The Claude Tag launch actually turned into one of Viktor's best growth weeks. Instead of folding, they bid on the search keywords around Claude Tag and rode the wave of search intent the launch created, picking up a flood of new signups while losing only a handful of customers.
Fryd's argument is that an AI lab like Anthropic is married to its own model, while Viktor can route to whichever model is best and own the context layer, which is far harder to leave than the model itself. Plus, a lab copying you is a sign you are onto something valuable.
Jean-Denis and Fryd are building very different products, but they landed in the same place. The model is rented, and anyone can rent it. The value doesn’t just sit in the model capability. It sits in taste, in the context a product builds about you, in the network around it, and in who owns the moment you decide you like it.
Could Google ship something good enough inside Gmail tomorrow? Maybe. However, generic has lost to "feels like mine" more often than people expect.
However, in a world where a large AI lab like Anthropic, moving faster than most startups, can copy any single feature in a weekend, is being loved a real moat, or just a head start?
What do you think? Reply to this email, I read every one :)
Leo
Thoughts on today's email?
📥️ Want to advertise in Consumer Startups? Learn more.