Town's CEO on the self-organizing company

This podcast touches on AI. My fiancé works at Anthropic. See my full ethics disclosure here.

Last week on the Platformer podcast, Replit's Amjad Masad predicted that apps are about to enter a long twilight. Instead of installing individual pieces of software, he said, we’re much likelier to outsource work to agents powered by artificial intelligence. And so for the fifth episode of our miniseries about productivity in the AI era, I wanted to talk to someone working to realize that vision.

My interest here is more than merely journalistic. My Hard fork co-host Kevin Roose and I are in the process of standing up a new company, and it has given me a fresh influx of tasks to coordinate: hiring employees, finding office space, organizing formation documents, and much more. For the most part, these tasks and documents are hiding somewhere in long email threads.

Many small businesses today organize work like this using a tool like Notion: a blank page that you can fill with boxes and tables and databases. And I did create a Notion workspace for us, outsourcing as much of the process as I could to the tool’s embedded agents. It worked well enough, though I can’t say the process sparked much joy. Staring at our new workspace, I had the sinking feeling that I had given myself a giant new project to maintain.

For these reasons, I was interested to read about Town. Unlike most of the tools we’ve featured so far, this one is brand-new: the company only came out of stealth in June, having raised $55 million from Andreessen Horowitz. Its core offering may sound familiar — it offers a digital assistant, called a Townie, that plugs into your calendar and email and can brief you on your day, your meetings, and take limited action on your behalf.

But I was struck by Town’s onboarding process, which quickly builds a dossier about who you are, the work you do, and the coworkers you interact with the most. From there, it builds a wiki that includes your personal profile, communication style, work patterns and other preferences. Within minutes, Town understood that I am launching a new company; that some of our top priorities include finding studio space, securing a network deal, and preparing for our launch episode.

Practically speaking, the wiki contains information that Town’s assistant can read to inform its briefings and other work. But browsing the wiki, I saw the spark of something more interesting — a self-organizing company. What if, instead of building and maintaining a Notion workspace, an agent built and maintained the equivalent for you based on data you share from your email, your meeting notes, and your Slack?

When I got on a call with Town’s co-founder and CEO, I learned that the company soon plans to take a run at that possibility. Jean-Denis Greze — his actual title at the company is “mayor” — started the company after about seven years serving as chief technology officer at Plaid. Before that he ran engineering at Dropbox, and before that he spent exactly one year practicing law. (He would later say that his worst day as an engineer would beat his best day as a lawyer.) 

Town's core bet — which, at least for the moment, I could take or leave — is on its assistant, which it calls a townie. When you create an account, your assistant gets a name, a personality, and an animal avatar; mine is a cheerful finch named Rufus.

Town isn’t cheap: most people pay $49 to $59 a month. When Greze worried that the many mothers using the product wouldn't pay, one told him it saved her more time than the $75 manicure she buys herself every month, and that it was "worth $600 a year for me to not have to do those things." Greze called that product-market fit. "I live downstream of capitalism," he told me. "I build the best product that I can, and then people pay or don't pay."

The most newsworthy thing Greze told me is that a team version of Town's self-writing wiki is "coming out very soon" — a company knowledge base that assembles itself from what everyone's townies already know. This feels quite challenging, from a privacy perspective: pool everyone's private files too carelessly and the AI might put "people's salaries in a spreadsheet by mistake," he said. "Then you've got big egg on face." Over the next five years, though, he predicts that we will come to trust specially trained models to enforce company privacy policies on their own — and that small, high-trust companies will embrace unified company knowledge bases long before the enterprise does.

Greze talks about all this in strikingly intimate terms. "I think AI is the closest thing to life that we've ever created," he told me, describing townies as beings you nurture rather than tools you configure. But he paired that with firm limits: he rejects AI agents as independent economic actors ("I don't need a universe filled with paper clips — no, thank you"), and when I asked whether Town would ever guilt a canceling customer by showing their townie crying, he called the idea a dark pattern and shut it down: "It's software. It doesn't have human rights."

The intimacy has policy implications, too. Echoing what Granola's Chris Pedregal told me about managers who want to read employees' meeting notes, Greze said Town's enterprise agreements explicitly forbid employers from reading their workers' townie conversations — and that the product deletes session data after 15 days.

A long excerpt of our conversation is below, edited for clarity and length. Listen to the entire conversation wherever you get your podcasts — just search for Platformer — or watch it on YouTube at youtube.com/caseynewton. And let us know what you think — we welcome your feedback at casey@platformer.news.

Casey Newton: I started using Town just a couple of weeks ago, partly because I'm about to start this new company with my friend Kevin Roose. When I logged into Town and it just started building a bunch of stuff for me, I felt like: oh, this is the thing I actually want. Do you think we're getting to a place where humans are not going to have to do as much of the organizing of that kind of basic company infrastructure?

Jean-Denis Greze: I'm very AI-optimistic in the long term. I think we're still pretty far from that happening. Most people don't know what AI can or can't do. So onboarding people onto AI means understanding what you do and then suggesting: "Hey, you're a real estate agent. I could check new properties in your market every morning. Do you want me to do that? Here's a button, and if you say yes, then every morning I will do that task for you." 

That's important today because people don't have this mental model. Using AI to suggest what AI can do goes some of the way toward automating the toil out of what you do. But I think it touches today probably 10 or 20 percent of the toil of a knowledge worker. The other 80 percent of the work — I don't think we're there yet. I just think in practice, if you go to Cleveland, Ohio, and you talk to 100 knowledge workers at accounting firms and local businesses — how much are you doing with AI? How much time is it saving you? — we're still in the 10-to-20-percent bucket, and not in the "it's doing everything for me" bucket.

Now, the bet we're making as a company is that over time it can be better and better at suggesting things it can automate for you. But we're human. Take an executive assistant who spends a lot of hours every week scheduling meetings and labeling travel expenses, and say that person is using Town to spend much less time on those things. It's not like they're on vacation for the 20 percent of their time that they've saved that week. There are plenty of things they could do that would make them a better executive assistant. So now they're doing more of that kind of work — and that work is by definition not something the AI can do. Over time, the automation just gives us more opportunity to do the things that are higher value in the human economy.

Newton: Let me ask about one particular way you're proactive that is apparently expensive for you. When someone signs up, you ask folks to plug in their email, their calendar, their docs, and you build this personal wiki about them. I believe you said this costs you about $100 per user. Is that right?

Greze: Total over the first two weeks of the user, yeah, it can be about that. We're not going to have it be that expensive for everyone, but that's roughly the cost upfront.

Newton: I feel like it paid off, because I got this dossier back about who I am, what I work on, and who the people important to me are. And this happened within, I want to say, 90 seconds. You just knew right away. What role does this part of the product play in the experience you're building?

Greze: For you, my guess is because you're online — I'm not going to say famous, but there's a decent amount of content about you online — it probably was able to do that at a higher fidelity, faster than it does for most people. After you connect your account, it does a little bit of sampling of recent emails and some online searching about you to generate the first pass of the dossier — the one you see within about 30 seconds to a minute. And then in the background, for the first week, we build a much more complex version, which is the one that starts to cost double, maybe triple digit dollars.

The dossier is the first “a-ha” moment for the user. You sign up and you name your assistant — we call them townies. And then we say: hey, my name's Ivy, I'm your townie, here's what I know about you, and here are a few things I can do to help you. And most people are like: what? You know me? You understand me? And that's when it clicks. People feel like it's not a product — it's actually a relationship. This feels more like someone I work with than a piece of software. The reason we try to do it really quickly, even if it's not as accurate, is because it's the first thing that makes you feel different — a different emotional feeling toward this product.

Newton: Every once in a while, when I'm using productivity software, I have one of these moments where I think: oh, everyone is going to copy this. And I had that when I saw this wiki that you built for your customers. Tell us a little about it.

Greze: We cannot claim that the wiki is our original idea — Karpathy and others have talked about how you can use AI to take a set of documents and build a knowledge base out of them. Our job is to take this incredible technology and make it accessible to everyone. We had the wiki in the background for a long time as something that powers the experience — and it turns out people love reading about themselves, so at some point it just became natural to expose it.

The single-person version of this is interesting. The multi-person version is even more interesting. You're talking about building this new company — it's not just what it knows about each person, which is private information for each person. The company has things about the business as a whole that everyone could use knowing. Can you do that automatically? Most companies aren't even on Notion — I love Notion, it's a fantastic product, but in the world of productivity, it's like 1 percent of companies that are on something like Notion. How could you give that kind of power to 99 percent of companies, where it can have a really good understanding of what's happening across the business, so that everyone can be more effective in their role?

Newton: All my company formation documents, information from my lawyer — everything is essentially just buried in my email, and I really do want to give an agent the job of: you go find all the attachments, you build the document library, you create the wiki. If you could figure that out, you'd have solved a big need I'm having in this new company.

Greze: It's coming out very soon, so we've got that figured out. But there are some interesting questions once something single-user becomes multi-user. In the single-user product, we have this thing called “People” — basically short files on the people that you work with, interact with, email with. They say: this is Alice, you've known Alice for seven years, you mostly have a professional relationship, and these are the things you're working on right now. But three years ago, you also helped Alice get a job. And when you write emails to Alice, you start with "Hi," and she says "What's up?" It knows your writing style just with Alice.

So one user said: these files are amazing — what if you put all of them into a CRM for my company? That's a genius idea. Thank you, customer, for telling us what we should do. But the immediate problem is that you can't take the union of every user's files, because there's private information. Maybe in one of the files it says: hey, you're talking to Bob about this opportunity at his company. You don't want that in your company CRM. 

As soon as you start to build team versions of these things, there's this privacy question. And if you tell the LLM, "When you're creating the team library, make sure not to put any HR information in there" — well, what if the LLM evaluates that wrong and puts people's salaries in a spreadsheet by mistake? Then you've got big egg on face. You have to be thoughtful when you build the team and company versions of these things, and that's the real challenge.

Newton: Give us a thumbnail sketch of how you're going to approach this, because I am going to have this exact problem.

Greze: The non-AI solution is that in a team version of the product, you're just really clear with people: whatever data you connect at the team level will be used to create the team wiki, and you don't pull data out of individuals' emails. You just don't do that — you let people decide what goes through. That's the traditional SaaS answer.

I think the five-years-from-now answer is that we will trust LLMs that will have company policies about what to let through. In that universe, you'll collect much more data, and you'll have an AI with a set of rules about what can or can't go into the company, and it will read documents — legal contracts and HR files and finance documents — and decide what can be put into the library. That seems scary now, because we're in a world of: what if it hallucinates, or what if someone prompt-injected it so it lets the wrong thing through? But I think over time we will post-train AI models that are really good at enforcing privacy policies on behalf of companies — because the shared knowledge bases are just so, so damn powerful.

The final lens to this is: you're a small company. If you're a large corporation, and someone comes to you and says, "I would like to build a compounded library across all the knowledge that you have about the business," you're just like: that's too scary. But if you're a three-person company with high trust, maybe you just talk as a team: “it would be so valuable if we had this common library — is everyone okay with it trying to grab from their emails?” You've accepted the quote-unquote risk, but you're small enough that the risk for you is negligible. So I actually think the SMB side of the world will benefit from some of these unified data approaches before the enterprise, because SMBs will be willing to assume the risk and will see the benefit faster than really large companies, which will be stunted by their historical privacy practices.

Newton: Let me zoom out. There have been countless efforts to redesign work around AI, but there hasn't really been a runaway winner just yet. You have this line that the 80th percentile user opens ChatGPT or Claude at most three times a day. That's not very much. Why is the prompt box the wrong way to get humans to work with AI?

Greze: Maybe it is the right way. Our grand theory is that it's a real habit change, and most people are very good at their jobs and very busy, and don't have the time to learn new technology and change their habits around it. Also, ChatGPT is a fantastic product, but you use it on day one without connecting it to any of your data. It doesn't ask you to connect to your data universe. You can go into connections and connect things, but it's not the default experience — they decided that. And that's why they have a billion users today. It's also a free product — it's very expensive to deal with all these data sources, so it doesn't align with their business model.

I think only since last November have LLMs been good enough to work over all of this data. It's just now that you're starting to see the first few products that assume: I can draft emails, I can access your text messages, your Google Docs, your company CRM. As soon as you have those capabilities as an assumption — your product only works if those things are happening; you can only use Town if you connect your calendar and your email, and if you don't, we don't let you use the product — it brings the product in a different direction. 

But still, lots of humans have to change behavior, and that takes a long time. The iPhone was incredible, but it took a decade for everyone to be on a smartphone. So for a technology that's only really worked well enough to do this kind of stuff since November — it's not even a year — we shouldn't be surprised that it might take multiple years.

And today, it is weird to me that in the national discourse and the global discourse about AI, we talk about it as this transformative technology that's scary at times and going to totally change society. But in fact, on the ground, outside of support and coding and a few creative areas, it hasn't yet changed that much how we operate. It just hasn't. It will, because it's incredible. But it's going to take some real time to get there.

Newton: Let me ask about your townies — what you call your digital assistants. You told me that you don't like the word assistant. For those who haven't tried Town: once you create your account, you get an assistant that has a name and a face and a personality and is a cartoon animal. Why this approach?

Greze: We think you're building a relationship with a being. I don't think we're building software — I think we're building a relationship between a human and a being. And I say “being” because I think AI is the closest thing to life that we've ever created. It's going to evolve over time and understand you better. You're going to tell it: "My kids' soccer games are always on the weekend — that takes priority over anything else when you're trying to schedule something." And if you're building a relationship with something, it needs a name. And we're visual beings, so it needs a visual representation. As you tell it what you like and what you don't like, it remembers those things, so its personality changes. You are nurturing it. You're growing it to be the being that you want to support you.

The reason we call it a “townie” and not an assistant is that if you call something an assistant, people think it's a personal assistant, or an executive assistant. And I do think AI serves us as humans — I feel very s