为什么在 AI 时代，年营收百万美元的创业公司不值得做

Ben Horowitz writes to me, “I found this incredible tool. It’s called Town. It changed my life.” And I’m like, “Wait, this is Jean-Denis’s company, but he does taxes for small businesses? This makes no sense at all. It didn’t look like a venture-scale business.”

No one has created a great AI for these typical workflows, such as meetings, preparing for them, and setting people up for the workday. We’re trying to create a product that helps people find more and more opportunities to use AI to work better. It’ll cost you about a dollar a month. The interesting thing is that in that one month a year when it finds something, it’ll save you hundreds, if not thousands, of dollars.

The product experience can’t be limited to just a text box. You have to show people your value and do it for them.

We’ll use the provider we believe is best for you at the time.

People are slow routers.

That’s right. And I think in the future there will be significantly fewer people in the decision-making chain than there are today. I think our job is to—

1. Leaving Plaid for the SMB tax bet

Hello, everyone. I’m Alex Rampell, a partner at Andreessen Horowitz, and I’m here with Jean-Denis Greze, CEO and founder of Town.

Thank you for inviting me.

I think it’s cool that we worked together when you were CTO at Plaid and I led our investment in Plaid.

Yes.

I heard you were leaving to start a company. Then I remember you told me you were working on a small-business tax tool.

Yes, we had coffee.

We had coffee here, in Menlo Park or Palo Alto. I was a little upset. I was excited, and then I heard “small-business tax tool,” and thought, “Oh, this is difficult.”

Well, now you know. The cards are on the table, right?

You didn’t tell me about it then.

I know, I know, but I always have to keep a happy face.

You never know. That’s what I find so funny about this—I think you know the whole story. Ben Horowitz writes to me, “I found this incredible tool. It’s called Town. It changed my life.” And I’m like, “Okay, right? This is an app, so it falls under my fund. Wait, this is Jean-Denis’s company, but he does taxes for small and medium-sized businesses. This makes no sense at all. I’m so confused.” That’s what I wrote to you.

A week later, before I knew it, we’d become partners, which I’m extremely happy about. But maybe tell us what Town is about. I think everyone will be extremely interested to know how you got to this point since leaving Plaid. You left Plaid and started doing taxes for small and medium-sized businesses. Tell us about the founder’s journey from then to today.

I left Plaid and took a few months off to be with my family because I knew that if I became a founder, I wouldn’t have any free time anymore. Then I started researching, in a B2B style, what markets would be interesting and where AI could really change the product experience. For many reasons, the tax space turned out to be very logical. Taxes contain a lot of text, and large language models work very well with text. You’re dealing with a lot of structured input formats, like financial statements and leases, so this seemed like an environment where we could take the process of generating a business tax return—which is very different from a personal one—and automate it.

I did what engineers do: we did a technical check to prove to ourselves that it was possible. We felt that, within the next 2 years, the process for small and medium-sized businesses—from financial reports and piles of paperwork to a finished tax return—could be automated using AI. We spent a year building it. I think it made sense, and it worked from a product perspective, but not from a go-to-market perspective. I have no doubt that tax preparation will be automated, but how do we effectively reach millions of companies and sell them this product? That raised many more questions than we expected.

A year later, we had what I would describe as a good business before the AI era. If I showed you the revenue numbers and growth in our first year—just under a million—and what we thought we could do in the second year, which was $2.5 million to $3 million in revenue, I think that, before the AI era, you would say, “That makes perfect sense.” But after the advent of AI, it became clear that this wouldn’t be a venture-scale company.

More importantly, we realized that incumbents would be able to use this technology to automate tax return preparation for their clients faster than we could gain market share. That’s a key point to think about if you’re a founder. Your advantage as a new player is that you move faster, can create a different product, and can use different go-to-market strategies. But the clock is ticking: how quickly can you get real word-of-mouth and market penetration before the incumbents, who have other advantages, do? They have distribution advantages, right? For example, before they can catch up with you and copy what you do effectively enough.

In the case of taxes, it seemed we wouldn’t be able to reach hundreds of millions in revenue before our incumbent competitors, because taxes can only be sold once a year.

Yeah. And it’s not just that they sell it once a year; the interaction itself fundamentally can’t be a tool for everyday use.

That’s true. It doesn’t have much of a word-of-mouth effect, because few people brag about their experience preparing tax returns. Even when they do, they can only do it once a year, after they’ve filed their return. When you’re bragging about how good your experience was, that’s usually not the time when people are purchasing these services or looking for a new tax advisor or accountant. They’ll forget everything by next year.

That’s true. These gaps make hyperscaling extremely difficult. That’s why it wasn’t a great business idea. It’s funny: one of my best friends built a successful EdTech company, and one of the things he always complained about in EdTech is the same thing. You can only sell this technology twice a year. You can sell in the summer for the fall, and then there’s a short window at the end of the fall and winter season, right before the New Year. Those are the only 2 periods when you can realistically sell software to schools and school districts. He said it really limited his ability to hyperscale.

I decided to forget about it when we were dealing with taxes. So, yes, we created Tax. A year later, after the first tax season, my co-founder and I went out a lot, drank a lot of coffee for a week, and really thought about the meaning of life, entrepreneurship, and capitalism. We decided we should do something else. We needed to find a better, more fertile market if we wanted to impact millions of people. There are better ways to achieve that goal.

2. Walking away from a working business

It was quite a bold act. Most people probably wouldn’t do that, especially since it’s not as if it wouldn’t work. At that moment, it felt like a failure, like we were failing. But I think one of the main motivations for building a business is this, right? As a founder, you have these moments—more often than you’d like—when you really ask yourself, “Why am I doing this?” For us, influencing the world was very important. We really wanted to create something important and help people. We stopped believing that we could have the kind of impact on the world that we should have, because we didn’t feel like we could scale the business fast enough.

The second reason we build a company is the real pleasure of working with our team. But everyone on our team has opportunity costs. They could work at Town or at many other great companies in the Valley. If you stop believing that you can achieve great results, it’s really hard to come into the office every day and look people in the eye when they say, “Hey, I’m having a hard time with this.” Or when you’re trying to convince a candidate to join your company and you know they have an offer from somewhere else that’s liquid, you think, “Who am I as a person? I have to believe in myself.”

I feel your pain. I’ve been there.

There are a few things we looked at and said, “We have to make this decision.” The other side of the coin is that maybe we would have found a way out. You never know for sure, right? We’re very inventive people, so you can always try something different, but we just didn’t feel like we had the idea to make this a truly breakthrough business.

The hardest part is admitting to other people that you’ve failed, and to some of your employees. For example, our tax team wouldn’t have a future if we weren’t going to do taxes in the company. You come to this and think, “First of all, who cares what other people think of you? My friends will still be my friends, whether I succeed or fail. As for the workers, we’ll help them find their next job. We can treat them well in terms of severance pay and other things, and do everything right.” Then you realize it’s not as difficult a decision as you might have initially reacted to emotionally, and it’s the right decision. You start to feel that confidence, and then you just do it.

One last point: since tax filing is an annual affair, I think making a decision at that point in time was the only period when you could have done it.

You can’t change course on April 16, after promising a bunch of people that you would do their tax returns. You can’t change course in January, after signing contracts with half of your partners. We had a window around October 15, about a month, where we could live in 2 worlds: a world where we continued to do taxes, and a world where we didn’t. After that, the doors were closing, so I think that time pressure helped us make decisions more efficiently than otherwise.

If you’re not in a cyclical business, you can always say to yourself, “Well, let’s try the next idea for a month. Let’s try the following thing for the next 3 weeks. Maybe this will work.” We had a short window when we could potentially change direction, and as soon as the window closed, we realized, “Hey, we have to decide right now. Otherwise, we’ll be here in a year, and after a year we’ll no longer have the capital to change anything.”

So from there, of course, the name didn’t change. It’s still Town. This is a good domain. Town.com is a very nice domain. But how did you get from that point to where you are now?

We considered all 4 ideas quite seriously. No need to list all 4. Before Town became a thing, we were led by the question: Why hasn’t anyone created a great AI that works with email and calendar yet? We started asking this because when we were building a tax product, a lot of customers would say, “Hey, it’s great that you automate taxes, but I also spend a lot of time scheduling appointments with clients. When I go to meetings, I forget what the meeting was about. Clients keep asking me the same tax questions. Wouldn’t it be great if you could prepare the answers for me in advance?”

We understood from these needs that these aren’t very complex things, but no one had created high-quality AI for common work processes like meetings, preparing for them, following up on their results, responding to fairly simple emails that come in constantly, and preparing people for the workday. We started exploring this idea right around the time Opus 4.5 came out and agents, at least in programming, started doing more and more autonomous work. I’d say that this simple epiphany, plus the fact that the models suddenly became more autonomous, allowed us to achieve product-market fit very quickly.

Product-market fit is a magical thing. When you have it, you want to grab it with a death grip and pull on that string forever. So we didn’t sit back and ask ourselves, “What could the market size be? Will everyone want to do this? What does it mean to win?” We just thought, “Hey, we’re getting a huge response from users. Let’s dedicate a few months of hard work to this, study the market, and see what opportunities open up before us.”

In the process, we started to realize that most businesses operate through email and calendar, especially any role in business that focuses on external communications, like sales, business development, or recruiting. Companies collaborate mostly through email. So we thought, “The TAM must be big.” And this isn’t just in the U.S.; it’s all over the world.

When our customers started telling us what they liked about the product, we realized there was depth in the requests they made of us. It’s not as simple as connecting an LLM to your email. It takes a lot more effort to make the experience better every time. I’m not saying we’ve seen the “moat” yet, but there was enough complexity for us to understand that we could dive deep into this and create a great product.

We did this from December of last year until April and just kept going every day. Literally every day, the product grew by 1% or 1.5%. Our customers kept asking for more, the development plan was very clear, and we felt it was time to end the closed beta. We were ready to release the product to the world.

One of the things I thought was very brave about what you did is that you have product-market fit, but we also talked a lot about the PLG model. Do you choose big companies that say, “Hey, give us your product. I want it”? You understand that if the customer isn’t the user, the customer writes the check, and the user has to work with the product. If the customer isn’t the user, you can create a bad product because you’re satisfying the CISO and not the real needs of the user.

My co-founder and I worked at Dropbox, and I think one of the important lessons for us is that you can take a world-class product and adapt it for the enterprise segment. Let me rephrase: It’s a difficult path, but it’s possible. It’s almost impossible to start with an enterprise’s compliance requirements and IT infrastructure and then try to make the product incredible from that starting point.

Until recently, maybe a few months ago, we believed that the greatest value was in improving the product, and that’s what we should focus on. Growth occurred naturally. We reasoned like this: if we got the product working and created a solid foundation from an enterprise, data security, and privacy perspective, we could later adapt the administrative part of the product to meet business needs. I think it was the right decision for us, especially given that consumer expectations regarding AI capabilities are growing. That will be the level of consumer services.

Our thesis is that CIOs and CEOs will say to themselves, “Hey, this amazing personal assistant I have in my private life—I want to have the same one for my company.” So whoever offers a corporate tool that feels like a consumer one will have every chance of winning. We decided, “This is where we want to be as a business,” and I think we’ve mostly succeeded in achieving that.

Now the question is, “Okay, we already have big companies using the product. How do we go from one corporate deal to effectively doing dozens of such deals? How do we meet data privacy and security requirements not just for a few clients, but for any large company?” That task seems much clearer now, and we can fully devote ourselves to it because the core functionality of the product is already at its best.

3. The iPhone moment for AI at work

Yes. You know, I agree, and I know you and I have talked about this a little bit. I remember when I had a BlackBerry.

Yes.

Let’s say it’s 2007. The iPhone comes out, and Steve Ballmer says, “Nobody’s going to buy an $800 phone. It will be a failure. This is just ridiculous.” Microsoft eventually buys Nokia. But BlackBerry “died” pretty quickly because, at least in my theory, every CEO that year got an iPhone or something for Christmas, took pictures of their kids, and thought, “Wow, that’s really cool. This thing—not so much. Hey, I’m the boss here. Figure out how to make this thing work for us.”

Yes, true.

I talk to a lot of company leaders, and it’s interesting that those who have seen the future say, “I need this.” Some people say, “Oh, no, AI doesn’t work that well. I use Watson,” or something like that. There are still many skeptics, but it seems that the iPhone has become a kind of watershed: people have learned the wonders of AI in their personal lives, but not in their work lives. Why is that? This is an open question. Part of it is about permissions, but I’d like to hear your thoughts on whether you agree with this vision and what AI for work will look like in the next few years.

First, I agree with this vision. As a small digression, there was a company—I think it was called Good.

Yes, Good Technology.

Yes, it was a corporate email client for the iPhone. It was an app. It was the ugliest app in the history of the world.

Yes, I used it.

I think this isn’t just a BlackBerry story. I don’t want to be like Good. Good ultimately wasn’t a good business because people would say, “Yes, of course, you offer a corporate service for the iPhone, but in terms of quality you don’t reach the level of other applications.” I believe that’s our risk. We risk not only becoming like BlackBerry, but also becoming like Good, instead of people just using the Gmail app on their iPhone, because it combines convenience and corporate features that are truly important for business.

I often think about how everything happens at human speed. First of all, assistants as a category are only used by a few million people around the world today. We’re still at the very beginning of the adoption curve. I don’t think all CEOs have reached this point yet. Just like with the iPhone, it will probably take 6, 12, or 18 months for most people to experience this magic in their own lives. When that happens, the question will be how to implement it in the company.

I think it’s more difficult than the iPhone because agents work more effectively when they have access to the data and APIs that your business relies on. If they have access to these things in a secure, confidential way, they can do a lot for you. They’re only as smart as the data they have access to.

I think we see this path as follows: first, you need to come into a company and bring in a few people by simply connecting their email and calendar—not yet the systems of record—so they feel enough magic to say, “Oh, this will be very useful in my work.” As I said, for external roles, calendar and email are a real source of insights on how to automate some of the work. Then you can show up with an assistant and say, “I can help prepare for these meetings.”

I can adjust the presentations for the meeting to better reflect the corporate identity and update the latest figures from the PDF you were sent. From this, people will begin to feel the magic. Then you ask yourself, “Have I earned the right to ask you to connect your CRM? Have I earned the right to have you go to your IT department and ask to connect all the other tools?” I think this path will happen more at the speed of people.

You need the other side of the coin. You should have audit logs and settings that allow you not to share sensitive information with the agent. You need cost control for the CFO, so they understand that if the entire team is using AI, they can see what the money is being spent on and whether the tokens are being used effectively. This is the part that will take people more time to adapt to, because you’re still selling to a Fortune 100 company that has to be managed with a certain level of accountability.

As you move into smaller market segments, into small and medium-sized businesses, I think you have fewer of those requirements, and implementation will be much faster. So our current bet is to focus on product experience and self-service for the SMB market, and on security, compliance, and financial controls for the enterprise sector. We think we can cover this entire space.

Yes, I think there’s a Jevons paradox in the market for assistants, too. I have an executive assistant here. I trust her. I like her. I don’t want her to read all my emails.

In fact, this just happened. I was in London and thought I should go to dinner. I could have sworn I’d gotten an email about it. I wrote to Snap[?], my colleague, asking, “Don’t I have dinner tonight?” And he replied, “Yes, you do.” It wasn’t on my calendar. It was my fault, not my assistant’s. I hadn’t replied to the email, so they didn’t add it to the calendar. It’s not their fault; it’s mine.

4. How Town was born

But now it’s caught on, and it automatically scans everything for me. It reads my emails all the time. This is a situation where my firm is kind enough to provide me with an executive assistant. There are many people who don’t have one. You might say the market for executive assistants is pretty small because most people don’t have them. But if you can build one with silicon, the market becomes quite large. Then you find people saying, “I’m going to let someone read my email all the time,” something I would never do for privacy reasons. It’s a double Jevons paradox—a Jevons paradox of privacy.

For those who don’t know, the Jevons paradox is that when the steam engine became more efficient—when James Watt made a more efficient steam engine—the assumption was that demand for coal would drop. What used to take a pound of coal would now take an ounce. But demand actually increased a thousandfold, because now you had something that could do the job very efficiently. It seems like this probably applies to the assistant market as well.

But I’m curious: how do you define your market? Is it some kind of assistant? That seems like too narrow a definition.

5. Town as a token-allocation machine

I think of this market as an attempt to create a product that works with people to find more and more opportunities to use AI to do their jobs better. In some cases, I think of our product as a token distribution machine. We try to give people options for where they can apply the tokens, and then they can choose whether they want the outcome.

As tokens get cheaper, there are more and more places where you could apply them that you wouldn’t use them for with today’s expensive tokens, and certainly wouldn’t use people for. The simplest example is writing preparatory notes for every meeting you go to. Even if you have a personal assistant, they won’t take the time to write perfect preparation notes for every 30-minute meeting. That’s an irrational use of time. But an LLM can do it for pennies.

Suddenly, it’s possible for everyone to have prep notes for every meeting, including research on who else is in the meeting and whether they’ve said anything on X in the last week that might be relevant. All this niche research now makes sense because it’s much cheaper. Our job is to find more and more of these things and ask you, as the user, “Hey, do you want us to do this for you?”

I’ll give you another example I’ve been really into lately. If you’re a small business owner running a restaurant in America, your wholesalers—the people you buy ingredients and other things from—probably email you an Excel spreadsheet or PDF with prices for their entire range once a month, or maybe once a quarter. These are thousands, if not tens of thousands, of products, and prices are constantly changing. If you own a restaurant and have a part-time finance person, that person should check once a quarter how prices have changed, to see if you can order something from other wholesalers or if you’re being overcharged because the price of something has gone up 20% or 30%.

Mostly, you don’t do that. It’s a rare occurrence, and while it could save you money, it’s a hassle, it’s not part of your operating activities, and you don’t have a full-time employee to do it. With AI, you can set up a process to run once a month and perform an analysis. You can literally create a tool like this for yourself in five minutes with a prompt. If it doesn’t find anything worthwhile for 11 out of 12 months, it’ll cost you about $1 a month. But the interesting thing is that in the one month a year when it finds something, it’ll save you hundreds, if not thousands, of dollars.

Now what may or may not have been worth the effort is worth doing at least once a month. You should do this every time there’s a change. It might be worth doing every day if your wholesaler changed prices every day, but that’s not the case. Our job is to show this opportunity to the restaurant owner. It’s not that they’re not smart or don’t understand their business well; they do. They just won’t find this valuable unless we show them. So we’re saying, “Hey, listen. Do you want me to make sure your wholesalers don’t change prices in a way that hurts your business?” The answer is, “Of course, do it.”

You don’t really care how often it runs. You care that every now and then you get an email saying, “Hey, you should probably move these things to another wholesaler. They’re offering a discount, or this place is really raising prices on these items, and it’s getting a little too expensive. Maybe you should take those items off the menu.” It’s not high math. But you need a product that tells you this is possible and does it for you.

Well, the interesting thing about this example is that if you went to a restaurant owner and said, “Do you want a personal assistant?”—

He wouldn’t have any idea.

He wouldn’t have any idea. Yeah, yeah. Whereas if you said, “Do you want to save money?” It’s like GEICO, right? “Want to save 15% on your car insurance?” Or, “Want to save on iceberg lettuce? I can save you $500 a year for $50 a year.” The answer is yes. But how do you know? It’s a question or a challenge for product marketing.

I remember two things we talked about when we first discussed the future of business: PLG versus sales-driven growth, and horizontal versus vertical. If you want to build a product designed for the user, you have to build it for the user, not for the corporate bosses or all the lawyers in the company. Over time, you can move back to corporate sales and sales-driven growth.

The other thing is horizontal versus vertical, which is something a lot of companies struggle with. That brings us to another point: What do you think about the big players, like Anthropic or OpenAI and all the others building things for both enterprises and consumers? How have you thought about horizontal versus vertical?

What I hope I’m conveying correctly is that we’re going to start as horizontally as possible because we don’t want to be limited to something narrow. We don’t want to be siloed and have a brand as “the company that saves you money on iceberg lettuce.” We want to do that, and we’re going to learn. This is my pitch to your candidates right now, by the way, so I hope I’m getting this right. We’re going to learn from thousands, tens of thousands of companies and people using our product. Then eventually we can go vertical, because we’re going to let the market determine that, rather than assume we know about the iceberg lettuce problem.

We wouldn’t know about it until someone downloads Town and uses it, and we learn from their experience. Then we can understand what vertical it could be. How do you think about vertical versus horizontal?

6. Horizontal vs vertical

I think most AI products today will have very similar capabilities. Most of the tools in the work context will be able to do the same things. So the differentiator—the thing that has been the main thing in Silicon Valley for a long time, having product features that are ahead of the market—is going to be harder and harder to sustain as an advantage.

We can put that aside and have an interesting conversation about the fact that I think the unit of value in software has shifted from the app to the computer. But what that really means is that it’s going from the app to everything a human can do, and a human can do almost everything with a computer. So we’re completely abstracting away from the app level, and I think the value at the app level is going to go down.

If the capabilities are the same, how do you build a successful company? Our view is that, as AI gets better, it will be able to tailor the product to each user’s needs much more specifically. If you’re an electrician who signs up at Town, the things we recommend you automate during onboarding will be very specific to electricians. If you’re a venture capitalist, they’ll be very specific to venture capitalists. If you’re an executive assistant, they’ll be specific to EAs in that industry, and so on.

How do you do that? The models aren’t sophisticated enough yet that, if you just ask a model, “Hey, recommend things Alex, who’s a venture capitalist, could automate with Town,” you’ll get more than pretty good surface-level suggestions. You won’t get anything really deep. And we don’t look at user data, because part of being ready to work with enterprise customers is that you can’t look at user data. So we look at the overall metrics for who our most active users are on the platform, and then we go talk to them. When we talk to them, they tell us what they’ve automated.

We can talk to 10 power users who run restaurants. In those conversations, we might notice that there are 5 things they all do in common. Then we sit down as a team and prototype automations, which we call “routines,” for restaurant owners—things we think are really useful, like one that deals with wholesale suppliers. We use those as a foundation. When someone comes on board, we look at their email and inbox and say, “It looks like you’re a restaurant owner.” Then we offer them templates of problems and potential solutions for restaurant owners that we think are really good, based on what we’ve extracted from those interviews.

You sign up and start using the product. Then the AI gets better, and we talk to more restaurant owners. Now we don’t have 5 things we can do for restaurant owners; we have 15. It’s still a little manual, but it can be sped up. You can automatically schedule interviews, offer restaurant owners a free Amazon gift card, and interview them with AI to see what they’re doing with your software. So you’re not working at human speed anymore; you’re working at AI speed. That library gets better, and the models get better. The recommendations then rely less on the library, and we think that over time, with this approach, we’ll be able to offer 80/20, highly personalized software for people.

People in the future won’t just have one role. You might be a restaurant owner, but you also manage your restaurant’s finances because you don’t have a separate person for that. You might also have a vacation rental property. You want your Townie to be good at all 3 things. And you have 2 kids in school, and it can help you plan their summer classes. This is completely custom software, and we believe you can use AI to build it.

Another strategy you’re seeing in the market, and what I think the labs are doing, is the FDE approach. That’s a completely different thing. With FDE, a person comes into your business, learns how it works, and then builds custom software using AI from the labs to serve you better. If you’re a company willing to spend about $5 million a year paying OpenAI for FDE, that looks pretty good right now. But I think that over time, as the models get better, the FDE approach will actually get a lot worse, while our approach—using automation and AI to build custom software—will scale a lot better.

We’re betting on starting with PLG: small and medium-sized businesses that are very tech-savvy, because our approach scales really well there. We’re going to let others compete in the enterprise sector with this more hands-on software. Our bet is that the knowledge we gain from the flywheel of development will make us more successful in the long run than these more individual approaches.

7. Why Town over OpenAI and Anthropic

Maybe you could also explain why Town is better than OpenAI and Anthropic, which now have connectors too. For example, you can connect Slack and LinkedIn. Of course, I believe in Town; I’m both an investor and an active user. So I have my answer, but I’d like you to elaborate on why someone should use Town.

And as a business, a lot of people listening to this who might be entrepreneurs are probably wondering, “Oh, are these guys going to eat my business?” Not you, Town. There’s this weird situation where there’s an infrastructure layer and an application layer, and that’s always been the case. Cisco and Akamai were infrastructure companies in the ’90s, and Amazon and eBay were application-level companies. But Cisco never said, “You know what? We’re going to build a bookstore. We’re going to build an auction site. We’re going to build a catalog like Yahoo that competes with…” No, they were just happy to sell their switches, load balancers, and all that stuff.

I’m wondering if you can define how you view Town compared to the people you often buy tokens from, and how others should think about this strange time we’re living in, where infrastructure is becoming an application. I mean, it’s risky, always competing with your suppliers, right? You’re competing with your suppliers. So you have to ask yourself: Is this the natural state of things?

The short answer, although it’s not the best one, is that we have a very different vision for the user experience of email and calendar. For some reason, other players haven’t focused on that. There’s a lot of complexity in email and calendaring, which makes our product much, much better at it. But that’s the short answer, because someone at OpenAI, Anthropic, or Google could wake up and decide to do the same thing.

The better answer is to think about our customers and what they want. When we talk to midsize businesses—companies with 500 to 1,000 employees—we ask them, “Listen, do you know who’s going to be the best model provider in the next 3 to 5 years? What’s your opinion?” They don’t know. They really don’t know. So we say, “If you buy an OpenAI product, it will always have OpenAI models in it. What if OpenAI isn’t the best provider, or doesn’t have the best intelligence for you, in the next 3 to 5 years? As AI does more and more for you, you may find yourself at a competitive disadvantage.”

Will they raise the price? Yes, all of that. The first thing we say is, “Look, we’re a neutral app. We’ll use the provider we believe is best for you at this time.” This is a significant part of our strategy. It’s not just about general intelligence, but also about who’s better at working with voice, video, or slides. We consider it our duty to determine what’s right for your business among all the possibilities in AI: frontier models and open-source models. I think you can not only save money, but also avoid the risks associated with choosing the wrong partner.

There are companies that chose OpenAI about 2 years ago, and then Anthropic got better at coding over the course of 12 months. Those companies had to use both because they realized that, to be competitive, they needed Claude. Then they needed those exact models. If you’re a smart business, you really think about the neutrality of who’s providing you with the intelligence and who’s doing the routing. So that’s one reason.

The second reason, which is very important, is our position that, as a business, you truly own the data, memories, and contacts that are created in Town. In the old world, before computers, businesses had different assets: physical assets, and maybe human assets. Now, increasingly, people might call it context, but there’s a set of assets—everything AI knows about your business to help you run it effectively. I believe you should own that entire set. You should own it to the point where, if there were ever a better solution than Town, switching to another solution would take a day. That’s what I’m selling: the idea that you want to own this asset.

If you look at people who hire developers to build software for them, they’re actually trying to make it harder for you to switch to someone else. So that’s the second reason. First, you want neutral intelligence, but you also want an application layer above the intelligence that allows you to switch easily. In the new world, it’s about having that level of context and having someone who makes it available to you.

The risk I’m taking with Town is that the cost of switching is lower for my customers. But that is my advantage. If I can’t have the best product on the market, that’s okay; it just means I lose. That’s my competitive position against the labs.

The third part for me is just learning. If you’re a big enough company, the labs won’t be training on you. But if the way you do business ever gets into the hands of someone who resells it to others, you’re effectively selling what makes your business successful to someone else.

And you definitely don’t want to do that. We don’t train on your data; it’s just not part of our business philosophy.

One last thing about competition: we have a different view of the world and of the relationship between humans and AI. We’re probably more forward-thinking than even the labs themselves. In Town’s Teams product, there’s a mode for creating an internal CRM where we trust agents to decide what corporate information is available to other employees.

Here’s an example: you’re in Town and ask, “Does anyone know anyone in the finance department of this prospect?” Town can reach out to other employees’ agents to find someone connected to the lead you need. It’s a very powerful tool, but from a privacy perspective, it needs to be designed with the utmost care.

I think about it this way: we’re focused on getting people to work at the speed of AI, not the speed of humans. The human version of my question would be to write in Slack or Microsoft Teams, “Does anyone know anyone on that team?” Then everyone in the company has to read the message and think about it, and you might get a notification in a day or so.

That’s right. People are slow routers.

That’s right. With us, you’ll get an answer instantly, because we’ve decided that for questions like that, we’ll trust AI to ensure security and privacy within the company. We believe this can significantly speed up how a company works, and I don’t see anyone else doing it.

When you build a startup, you always have to think about what the future will look like. I believe that in the future, human participation will be minimal compared with what we see today. Our task is to help companies embrace this new world, because the economic benefits far outweigh the risks they see today.

Those are some of our differences. I believe data ownership and neutrality are extremely important. Look at what’s happening with Rogo or Harvey. OpenAI is creating vertical companies that compete with other companies. From a competitive perspective, you start to think, “Oh my God, these labs are truly unstoppable.” But a neutral business choosing who to work with has to think about how far up the stack those companies are going to go. Strategically, I just don’t think it makes sense for most large companies to depend on any one lab.

There also seems to be a “Goldilocks zone” of relevance here. If you had a trillion dollars in revenue, I’d be happy as a shareholder, and you’d be happy as CEO and co-founder, but you’d probably become the number one target for OpenAI and Anthropic. If you had $0 in revenue and no one was using your product, they wouldn’t care.

But if you’re in coding, you may have ended up on the wrong side of that zone. If you were a coding agent right now, you could be adding $10 billion a week. Who knows what that means for Anthropic—it’s so important to them that they can’t ignore it.

Just one objection: what you do may not be the most important thing in the short run. I think it’s actually the most important thing in the long run, but in the short run, it’s not coding, right? It’s not like, “Oh my God, I have to do this or it’s the end for me.”

OpenAI seems to have learned from Anthropic’s example. OpenAI had 50 bets, and Anthropic went all-in on code, and it worked out really well. So there’s still an element of focus and relevance. They’re probably not going to ignore healthcare, law, or coding. Those are big bets they’re almost forced to make, while the more horizontal, scattered directions seem a little harder for them to pursue.

For small and medium-sized businesses, I’d agree with you. But the AI market—the market that does more for businesses and helps them become more successful—is bigger than the programming market.

Oh, definitely.

For me, it’s a question of when, not if.

Yes.

I think that under current conditions, this won’t change the situation at these companies, but in 12 months it will. Everything will move very, very quickly.

It’s impossible not to be in coding now.

Exactly. You can’t help but be in the coding market.

8. Why AI is underhyped outside the Valley

True. Is that right? It’s something like iPhone versus BlackBerry. There’s a distribution point here.

Actually, this is another question I’d like to ask you. I have a theory, and I’d like to know if you agree with it. Silicon Valley is almost overestimating AI, if that’s even possible. I don’t think AI can be overrated, because it gets better every week. But the rest of the world, I think, almost underestimates AI.

People talk about “bridging the gap,” but that’s not the best example, because the technology doesn’t only exist here. It’s all over the world.

No, people use it all over the world; it’s just unevenly distributed.

Where do you think we’re lacking in AI? Which sectors of the economy, geographically or professionally, have gaps?

This is the cool thing about a horizontal product: people find Town all over the world. I remember you telling me, “Wow, we’re popular in Japan.” That wasn’t intentional. There was no plan to become popular in Japan; it just happened.

Are there areas that surprised you? Any niches where you think there’s a lot more opportunity than people expected? I’m thinking of a specific example: a business that rents surfboards on an island and uses Town to inspect the boards and beach buggies for damage, then decide whether to refund someone’s deposit.

That’s pretty cool. It wasn’t on my list of expectations, though—not on my list of 128-by-128 possibilities. It just wasn’t there.

What’s really happening is that there’s an enthusiastic inventor at this small company who was annoyed by something. They figured out, with Town, how to automate comparing these images and deciding whether to refund the deposit. It’s this inventor who makes it all work.

So my answer is that we’re seeing applications in the most unexpected sectors of the economy. The key to success is whether there’s someone out there who understands AI enough to want to invent a solution to their problem. The problem is that there aren’t many such inventors in the world, or enough people who have the time to do this for the benefit of their business.

I think that’s why AI isn’t spreading so quickly. Over the last 12 months, Claude or ChatGPT have had features that many small businesses could use to save a lot of time. But instead, people use them as a search engine replacement—as a better Google—and wonder why nothing changes.

A product can’t just be a text field. What we need to do is something you can’t teach someone: you have to show them the value of the product and do everything for them. I think that’s where the real breakthrough lies.

When you ask what makes someone stay with Town, it’s that they connected their email and calendar and, ideally, one important accounting system for their business. Then they created one of their own integrations. They did something that wasn’t in the standard set, and it was magical.

I’m trying to get rid of the need for individual integrations. I want to give everyone this opportunity: once you connect your accounting system, can I provide value in a way that’s right for you and really helps? That’s the opportunity, and it’s why you don’t see an even spread of AI.

In Silicon Valley, we work in companies side by side with engineers, so we’re much closer to this innate ability to solve problems and invent. We have inventors everywhere in our companies. Our companies are using AI effectively, while most of the world simply doesn’t have that opportunity. That’s the real problem.

If you can fix that—if you can make it so a person who isn’t an inventor gets the same value from AI as an inventor—you could create a huge business today, because AI can solve so many business problems. You don’t need more intelligence. The intelligence is already there. The only question is how to implement it effectively at scale.

And it seems like this is an integration problem. It’s a product marketing problem.

Yes.

I think this is also really interesting. I met a professor at Harvard, a very, very intelligent woman. She’s the second-most senior person there, and I was talking to her about AI. She said, “Yeah, I tried that a few years ago. It was good, but not great.”

I said, “No, no, no, it’s really good now.” So I asked ChatGPT to rewrite her thesis, which she wrote over 20 years ago. I took everything as it is today, and she said, “This is incredible. What kind of AI is this?” I said, “This is ChatGPT.” She said, “No, it can’t be.” She asked, “How is this possible?” I said, “Because everything is moving very fast.”

Many people judge things by how they are today. I think that’s the thing: Silicon Valley tends to overestimate what things can lead to. Sometimes you overestimate reality, create a distorted reality field, and so on. But you see where this could lead, while many other people see things as they are today.

The progress is such that, if you follow this curve, it seems people aren’t taking advantage of it—partly because, paradoxically, they’ve become early adopters.

They tried it 2 years ago. They say it hallucinated and didn’t work very well, and they cemented that version of reality. If they try it now—I’ve done this experiment 10 times—they ask, “Where did you get such an AI?” They say, “This isn’t the AI I have.” And I say, “Yes, that’s the one. You just need to try again.”

The product space is changing so quickly, and it’s going to continue to do so. The platform that continues to adapt its business to the best AI will win. And that’s a problem, because AI will just keep getting better and better and better forever. People’s mental models of what’s possible and what isn’t, and what’s difficult and what isn’t, will be outdated.

With a computer and a browser, assistants can do things today that were almost impossible 6 months ago. In the last 2 and a half months, they can handle those tasks in 90% of cases, compared with about 30% before. Most people haven’t felt it yet. When you do a demo where it does some shopping for you or uses your software to take an action, you think, “Wait, doesn’t that need an interface?” People are surprised: “Wow, I had no idea it could do that.”

I don’t even know how to sell it to people, because if you say, “Oh, now Town has a computer feature,” it doesn’t mean anything. It should be much more specific: “Hey, I can make posters for this event that look exactly like the ones you made in Canva. I can do this with your Canva subscription. You’re paying for it anyway. Do you want me to do it?” And the person will answer, “Yes, let’s try.”

Five minutes later, for just 25 cents, it shows you something extremely valuable. Then you ask, “Hey, do you want me to do this before every event like this that I see on your calendar?” When you change the design for your small or mid-sized business, you won’t even have to ask. The next time it creates flyers, they’ll already have the new design, because it should have noticed the change for you.

The next time it can effectively create a video, it’ll say, “Hey, do you want me to make some video footage so you can promote it on Instagram?” It should offer things that seem increasingly impossible with today’s technology, but it’s really just following the curve of what models can do for you.

9. Neutral agents and the broken ad covenant

Well, you’ve probably seen this: agent traffic—and not mostly assistant traffic, but agent traffic in the broad sense—has significantly exceeded human traffic. That’s the first point. There’s also a lot of talk in the news about an adversarial approach to agents versus a synergistic one. This is a new generation of agents—not the kind like ChatGPT that just search the internet, but ones that actively act on your behalf.

Amazon has obviously taken a somewhat hostile approach, which makes sense since they’re an advertising company. There’s been a certain agreement on the internet for a long time: I go to Yahoo Finance to look at stock quotes. Yahoo pays for a feed of quotes from Nasdaq or similar sources, but I see ads for E*TRADE. The deal goes like this: “I, Yahoo, give you, Alex, free quotes if you watch my ads.” And this agreement has been broken.

Those who’ve encountered this broken agreement are hostile, but there seems to be a great opportunity for synergy. It’s like, “Wow, this is the new SEO.” If I could accept traffic from agents and not block it—not think about how to stop it, but how to welcome it—that looks like a huge opportunity from the other side as well.

Yeah, I mean, Amazon’s loss should be Walmart’s gain, right? Walmart doesn’t have an advertising business, so its incentives will be different. And I think, to your point, distribution has changed. The most interesting thing about agents is how distribution changes if you no longer trust your agent because it might have an interest in something. That’s an interesting question, isn’t it?

Let’s say you use a Meta agent, and Meta has attracted advertisers. You ask it, and United Airlines is advertised on Meta much more than some unknown airline. You ask it to pick a flight. What happens if you feel like it no longer offers you a neutral option? There are always human incentives to promote products and engage in marketing. What does this look like in a world where people no longer look at things themselves? That’s a very open question.

I firmly believe you need a neutral agent that belongs to you, whose sole source of funding is your wallet as a user. Whoever creates that agent should be interested in providing you, as a human being, with the best information, without distortions caused by marketing forces. You have your own taste and express it by making certain choices. You can ask your agent, “Hey, what are people talking about on Instagram?” Or, “What are they advertising there?” You might ask because, in this case, you want to experience the impact of advertising and marketing. But you want your agent to remain neutral about it.

On the question of competition or the lack of it, I find it interesting to distinguish between the friction points you enjoy and those you don’t. Walking around SoHo in New York City to shop—that’s friction I like. I like having to walk between stores. I like other people standing in line, because I can adopt their style. I like that friction. That’s the enjoyable part of shopping.

And there’s friction I don’t like. I want to order toilet paper to take home. I just want the same toilet paper I’ve been ordering for a long time, delivered as soon as possible and as cheaply as possible. Sometimes, when we think too much about agents, we focus too much on reducing friction and effort and making everything as efficient as possible. But people don’t always optimize everything for efficiency.

The Amazon example, in terms of advertising, is that sometimes I’m on Amazon buying LEGO for my kids, and I really want to see ads for other building blocks or toys for my son. I might see something that isn’t from LEGO that I didn’t know about, and it’s interesting. Then I want to ask my agent to find reviews. I can say, “This is an interesting thing. Is it popular with children? Maybe I’ll want to buy it.” In that experience, I think Amazon can make a lot of money, and there will still be value for platforms that aggregate the offer with advertising and marketing. But there are many times when you optimize friction and don’t need those things.

If I were Amazon and had $80 billion—or, what, $70 billion—in annual advertising business, the question I would ask is, “Hey, I don’t think you want to block agents.” I think there could be a company bigger than Amazon, like a headless Amazon, that could be a huge, just huge business. But I think it puts more pressure on them to create a separate experience where you actually want to see the ads, and it’s really good, and it’s worth going there for that. I think I would still go to Amazon for these things eventually.

AI is evolving so fast that it’s going to change distribution and human habits at such a rate that I don’t think anyone can predict which platforms will win in these areas. There are going to be huge losers, right? Amazon could lose a huge chunk of its advertising business trying to grapple with the future, but they can’t predict what it will look like without it.

Another example that just occurred to me is Muse. I don’t know if you’ve seen the discussions around Muse. It’s like a free, quote-unquote, personal agent. People say, “Well, one day they’ll have to advertise in Muse.” But they might never have to advertise on Muse. Maybe Muse is just giving them data that allows them to target you better on Instagram.

When you go on Instagram, you’re there because you don’t need efficiency; you want to have fun. If they can target better on Instagram thanks to the information from Muse, and Instagram can let you pay for a product you like using the wallet information they collected through Muse, or Muse can search the web for different variations of that product, it’s a really great flywheel. It means you’ve separated monetization from the agent experience for this personal use case.

And this is where Amazon, in my opinion, loses. Maybe because of agents like this, simply being an aggregator of goods may not be the best business model in the world of agents anymore. There’s really not much they can do about it. They’ll lose.

Yes. They must have their own agent—that’s their countermeasure. Don’t block agents; have your own. Have a better one, make it exclusive to shopping, hope that’s enough of an incentive, and add it to Prime. But I’m not sure they’re capable of that.

Interestingly, there’s a site called CamelCamelCamel associated with Amazon. It’s a price-tracking service for Amazon. You’re literally saying, “I want to know when Dr Pepper, which is currently $8, drops to $6,” and then buy it as soon as it costs $6.

No, it doesn’t make the purchase for you.

It just sends an alert, and you have to buy it yourself.

Aha. This site is used by 20 million people.

20 million people?

Yes, 20 million people in America. And it’s simple. This is an affiliate site that’s been around for 18 years.

Justin and I wrote an article about how the best way to predict the future is to observe the present. Again, going back to the fact that the future is unevenly distributed, this is the future. When I think about agents, I divide people into those who have more time than money.

Yes, that’s true.

And those who have more money than time.

Aha.

Those who have more money than time won’t use CamelCamelCamel, and those who have more time than money will take advantage of it. But wait a minute: What if you could give everyone who has more money than time all these tools, so an agent could do it for them? Of course they would use it.

This is an IQ test.

These things are definitely going to become much more common. Sometimes you don’t even have to imagine what could happen. Just look at coupons.

Yes.

You shouldn’t use a coupon if you’re rich, because it takes too much time. You’ll spend about 2 hours searching for the right coupon. But this is another arrangement that’s falling apart, because it was a very convenient way to divide the world.

Yes. I mean, the Amex premium membership is the same thing. Now it just optimizes all your advantages.

Yes, all your advantages. The unused benefits model is dead.

Yes, and so we’ll all be abusing all our Chase and Amex benefits, and they’ll have to change that. This is unprofitable for both, because then it will become a hostile confrontation. It’s a kind of confrontation between hostility and synergy.

But the bottom line is that I think supply aggregation has become much less valuable than it used to be. It used to be very important because human attention was limited, and now supply aggregation is much less valuable because human attention can be automated.

Yes.

It was a great conversation. Thank you again for coming today.

Yes, thank you very much, Alex. It’s always nice to chat. Thank you.