Former Plaid CTO Shut Down a Profitable Business to Bet on an AI New Species in One Month

A Good Business Crushed by AI

The story opens with a touch of dark humor. Ben Horowitz, a partner at the well-known venture firm a16z, messaged his colleague Alex Rampell saying he'd discovered a life-changing tool called Town. Rampell was stunned—Town was a product built by Dennis Glez, and he clearly remembered that this person had been doing tax filing for small and medium businesses, which was a pretty good business. What surprised him even more was that Rampell had led the investment in fintech company Plaid, and Glez was Plaid's former CTO—yet he had been completely unaware of this entrepreneurial venture.

For his first startup, Glez chose SMB tax filing, feeding a company's financial statements and contracts into a large model to generate a return that could be submitted directly. The direction looked promising: tax filing is all text, and it comes with a lot of structured data like financial statements and lease documents, which large models happen to excel at. He spent a year building the product, with first-year revenue just under $1 million, and by extrapolation it could reach $2.5 to $3 million in the second year. Before the AI era, this was a beautiful upward curve.

But once large models arrived, the ceiling changed. Incumbents could use the same batch of models to automate filings for their existing customers, grabbing share faster than startups. Glez realized this business was destined never to reach venture-scale. For a founder this is deadly: newcomers have an advantage, move faster, and can build different products, but there's a ticking clock on the wall—you must quickly build a word-of-mouth network before existing companies copy what you're doing.

Tax filing happens to have several unavoidable hard problems. Tax can only be sold once a year; it's not a daily-active tool people open every day, and you almost never see anyone bragging about how well they filed their taxes. Once filing is done, it's not the season when people are scrambling to switch tax preparers, and by the next year the whole thing has likely been forgotten. These breakpoints make hyper-scale expansion very difficult.

After the first tax season ended, Glez and his co-founder spent a week taking walks and drinking coffee, seriously thinking about the meaning of life, startups, and capitalism, and finally settled on one thing: the tax team had no future at this company. They decided to switch to a market more worth pursuing. Rampell called it a rather brave decision—one most people might not be able to make, especially when at the time it didn't seem viable at all.

Jevons Paradox is the key economic concept for understanding the market logic in this article. In 1865, British economist William Stanley Jevons observed that when steam engine efficiency improved and coal consumption per unit dropped, total coal consumption across Britain rose rather than fell. The reason is that efficiency gains lowered the cost of use, which stimulated many new uses that were previously uneconomical, amplifying total demand. This paradox was later widely used to explain the relationship between technological progress and resource consumption: the more efficient something becomes, the more of that resource gets used. In the article, Glez uses it to explain that as the inference cost of large models (Token prices) declines, people will find more and more scenarios to apply AI to work—preparing notes for every meeting, comparing wholesaler prices every month, and so on. These things that were previously not worth doing due to high labor costs suddenly become worthwhile, and total Token consumption and total AI value expand accordingly.

A Difficult Pivot Within a One-Month Window

Glez's response was candid: at that moment it felt like failure, like he had really screwed up. He repeatedly asked himself why he was doing this, and the answer was that he wanted to have some impact on the world, to build something useful that helps others. Since he no longer believed it could have that much impact, the belief collapsed. Then there was the team—everyone has their own opportunity cost and could have gone to many excellent companies. If the founder himself no longer believes, it becomes very hard to walk into the office every day and look people in the eye. The hardest part was telling people we failed—but once you actually say it, you realize, who cares what others think of you? Employees can be helped to find their next job, and severance can be given decently. Once things were sorted out, he realized this wasn't the difficult decision he had imagined in the heat of the moment—it was the right one.

Another key factor was time pressure. Taxes are levied annually, and the decision made at that point was the only opportunity they could control. You can't wait until April 16th, after telling many people you'll help them file, and then turn around; nor wait until January, after signing half your partners, to pivot. The window was about one month, around October 15th. During that time they were living in two universes at once: one continuing to do taxes, and one that could no longer go back. Once the window passed, the door would close. Glez knew clearly that if the industry weren't subject to cyclical fluctuations, people would always tell themselves "just one more month, then three," but they didn't have that buffer. When the window disappeared, they had to decide immediately—otherwise a year later they'd still be in the same place.

The company name ultimately wasn't changed—it's still called Town.

From Tax Filing to Email and Calendar: Demand Hidden in Customer Complaints

What happened between tax and today's Town? They seriously explored about four ideas and finally chose one problem: why has no one built a truly good product that runs on email and calendars?

This idea didn't come from nowhere. While building the tax product, many customers would say how great it was that you automated taxes, but they still had to spend a lot of time scheduling meetings, often forgot what had been discussed before a meeting, and customers were always asking the same tax questions—if only you could pre-draft answers for them. These needs weren't complex to address; it's just that at the time no one had applied powerful AI to these utterly common workflows—meeting preparation, post-meeting follow-up, replying to simple emails.

They began exploring this around the release of Opus 4.5, when agents on coding networks could already independently complete more and more work. They didn't sit down and seriously calculate the potential market size; they just had a lot of user feedback and decided to sprint full-force for the next few months. During the push they discovered that most companies' emails and calendars are overflowing, especially for outward-facing roles like sales, business development, and recruiting—collaboration between companies happens mainly through email. They judged the potential market must be large, and not just in the US—the whole world over.

Product-led growth or direct enterprise sales? Glez's judgment: the customer is the one who writes the check, the user is the one who deals with the product every day. If the customer isn't the user, you'll likely end up building a bad product that delights decision-makers but torments users. His co-founder came from Jobbox, and the most important lesson from Jobbox was that you can always deliver a world-class consumer product experience first, then think about what it means for the enterprise; the reverse—starting from enterprise compliance and IT requirements and then pushing toward a great experience—is nearly impossible. So their focus has always been on making the product good enough and letting growth come naturally, while enterprise beta ensures data and privacy security and reliability. They're betting on consumer expectations—the CEO's mental reference is the personal assistant they've experienced in their personal life, and they want the company to have one too.

Four Positions of a Neutral Agent

Rampell added his own observation. When the iPhone launched in 2007, Steve Ballmer said no one would buy an $800 phone, and BlackBerry was soon finished. His explanation is that many CEOs received an iPhone or similar gift that Christmas, were taking photos of their kids and were blown away, followed immediately by: "I'm the boss—go figure out how to make these things work in the company." Some people still remain skeptical, saying AI doesn't seem that useful, I use Watson-like software, and it feels like the iPhone has already appeared, people feel the magic of AI in their personal lives, but that feeling hasn't yet entered the workplace.

Glez used a company called Good Technology to explain this concern thoroughly. It was an early enterprise email client for the iPhone, and he described the software as absurdly ugly. He brought it up to say this isn't just the BlackBerry story—sometimes making a good product isn't necessarily a good business. He emphasized that the speed of change is the speed of people: right now the agent category already has millions of users worldwide, but it's still very early in adoption, like the iPhone, and it may take 6, 12, or 18 months for most people to feel the magic themselves. Once they feel it, the question becomes: how to deploy it inside a company? This is more complex than the iPhone because agents are only effective if they can access data and the APIs that run the business. Their approach is to talk about the journey first, get some people to connect their inbox and calendar and feel enough value; at the enterprise level, there are still audit logs, confidential information that must not be shared with agents, cost controls demanded by the CFO, and so on. Moving down-market to SMBs and mid-market companies, the requirements are much lighter and adoption much faster. His current bet is to build a great product experience, enter the low-to-mid market through self-serve growth, and then win enterprises from a security and compliance angle.

Talking about the assistant market, Rampell shared his own experience: he has an executive assistant, but even so he wouldn't feel comfortable letting them see all his emails. While in London he thought he was going to a dinner, but the calendar didn't show it at all—because he himself hadn't replied to the email. So he said the executive assistant market looks small because most people don't have one, but if you actually build it, the scale would be considerable. Here's a Jevons Paradox: after steam engines became more efficient, people thought coal demand would fall, but instead it rose 1,000-fold because more things could be done efficiently. Glez defines the market as building a product that interacts well with people and helps them find more opportunities to apply AI to work. In a sense Town is like a Token allocation machine, letting people decide where to spend their Tokens. As Token prices fall, there will be more and more places to use Tokens.

He gave an example: preparing meeting notes before every meeting. An executive assistant wouldn't do this because it wastes too much labor, but using a large model makes it cheap. So people who need it can suddenly have prep notes for every meeting, and can also check who else attended the meeting. These scenarios become valuable simply because they've become so much cheaper. Their job is to constantly find such opportunities and then ask users: do you want us to do this for you?

Another example is restaurant owners. If you run a restaurant in the US, your wholesaler probably sends an email every month, or a PDF every quarter covering prices for all inventory items—thousands of items whose prices keep changing. In theory your part-time bookkeeper should do one thing every quarter: check how prices have changed, whether you could order from another wholesaler, whether you're overpaying because something went up 20% to 30%. But in most cases you don't do it because it's too tedious and outside your operational scope. With Town, you just run an analysis once a month and get results in five minutes. Even if 11 out of 12 months turn up nothing, it's still worth doing, because running it costs about a dollar, and in the month when there is a finding, it can save you hundreds or even thousands of dollars.

FDE stands for Forward Deployed Engineer, a delivery model pioneered by enterprise software companies like Palantir: engineers are stationed inside the client's company to deeply understand its business processes, then customize software or AI solutions for the client. The advantage is proximity to real needs and highly tailored delivery; the disadvantage is high marginal cost and difficulty scaling—each customer needs an on-site team. In the article, Glez contrasts Town's automation approach with FDE, noting that if an enterprise pays high FDE fees annually, it looks cost-effective in the short term, but as model capabilities improve, fewer and fewer steps will require human intervention, and the cost-effectiveness of heavy on-site staffing will keep declining. Town's chosen direction is to let the product itself accumulate industry knowledge through aggregate metrics and user interviews, automatically identify user types and push polished automation workflows, thereby achieving a similar customization effect at much lower marginal cost.

Start Horizontally, Leave Customization to Automation

Moving from product and growth to horizontal versus vertical, Glez is clear: he doesn't want Town to be pigeonholed as "the company that saves you money on lettuce." They prefer to start horizontally, learn from tens of thousands of users and tens of thousands of businesses, and let the market decide the vertical direction. In his view, most AI products will increasingly look alike; the unit of value has shifted from the application to the computer itself, applications are being abstracted away, value will decline, and it's increasingly hard to lead the market with a single feature.

So how do you build a successful company? Their answer is that as AI advances, the product can better customize itself for each user. Whether you're an electrician, a venture capitalist, or an executive assistant, the things that can be automated differ—but this is still done poorly today. If you just ask the model directly, you only get superficial suggestions. So instead of looking at user data, they look at aggregate metrics, find active users on the platform and talk to them. Talk to ten veteran restaurant people and you'll find they all do five common things, then build a prototype automation workflow and call it "the restaurant owner's daily routine." When someone joins, the system judges whether you look like a restaurant owner and presents the polished solution to you. AI is still advancing; as the number of things it can do grows from five to fifteen, dependence on external libraries decreases, and gradually it can deliver 80/20 customized software.

He brings up the lab's FDE approach for comparison: send a person into your company to understand the business, then use AI to develop custom software. If you pay OpenAI $5 million a year in FDE fees, it looks cost-effective now, but as models improve, this approach will get worse and worse. They've chosen a more automation-oriented path with much better scalability.

Why Better Than OpenAI and Anthropic

Next comes the comparison: why is Town better than OpenAI and Anthropic, especially since they now have connectors to link Slack, LinkedIn, and other apps. Rampell's analogy is the relationship between infrastructure and the application layer: Cisco and Akamai are infrastructure companies; in the 1990s application companies like Amazon and eBay emerged, and Cisco never thought about opening a bookstore or building an auction site itself. He says we're in a strange era where old infrastructure is turning into applications, and competing with your own supplier is risky.

A brief but not-so-good answer: the user experience around email and calendar is completely different, and no one else has really focused on it. But that's the simple answer, because someone at OpenAI or Anthropic could wake up any day and decide to do this. The better answer goes back to customer needs: talk to a mid-market company with 500 to 1,000 people and ask which model provider will be best in the next three to five years—they probably can't answer. So if you buy OpenAI's product, do you have to keep using OpenAI's models forever, even if it's not the best? So they position themselves as a neutral application, using whichever provider is best for you right now—not just for general intelligence, but also for voice, video, slides, and other capabilities, where the decision stays in their own hands.

The second position is that ownership belongs to you: data, memory, and context. Enterprises used to own assets like physical assets and human resources; now there's another category, perhaps called context—everything AI knows about your business. He believes you should own that entire bucket, and if there's a better solution than Town, migrating should take just one day. Those who deploy FDEs for you are actually trying to make it harder for you to switch. The third part is training: a company large enough that the lab won't experiment on you—but if you get rolled into the weights, you're essentially selling the key to your company's success to someone else. So Town doesn't do training.

The last position is about competition—they may see the human-AI relationship as more forward-looking than the labs do. Town has a product called Teams, centered on building an internal CRM, where they trust agents to decide what information the company holds or whether other employees can use it. For example, in Town you ask if anyone knows someone on the finance team, who is this prospect—the Town assistant can see other employees' agents and help you find the person relevant to your lead. Although this feature is powerful, from a privacy perspective the design must be excellent. Their idea is to let people operate at AI speed: in the past, if you asked in Slack or Teams whether anyone knew someone on that team, you'd wait about a day for a response; with Town you get an answer immediately. They've decided to rely on AI to handle security and internal privacy in such matters; the future will involve far less human intervention than today, and their job is to help companies embrace that world, because the economic benefits far outweigh the risks they may face.

For most large companies, truly depending on any single lab makes no sense. There's a sweet spot of relevance: if your revenue is a trillion dollars, you're naturally the lab's number one target; but if your revenue is zero and no one uses your product, they won't care at all. Coding has already crossed that line—Anthropic has invested heavily in code and clearly can't ignore it—but healthcare, law, and coding are all enormous fields, and labs are almost forced to make these big bets, making it hard to go deep in small markets through horizontal diversification.

"Crossing the Chasm" originates from Geoffrey Moore's 1991 book of the same name, a business extension of Everett Rogers' diffusion of innovations theory. Moore pointed out that there is a chasm between early adopters (technology enthusiasts, visionaries) and the early majority (pragmatists) of high-tech products, and the two groups have completely different purchasing motivations: the former are willing to tolerate flaws for new technology, while the latter demand complete solutions, reference cases, and a mature ecosystem. Many products stall after becoming popular among early adopters precisely because they fail to complete this leap. In the article, Rampell borrows this concept to point out that AI diffusion is not an "all or nothing" issue but is extremely unevenly distributed—engineers in Silicon Valley have long crossed the chasm, while most SMB owners around the world are still at the stage of treating ChatGPT as "a better Google search." Glez concludes that the real opportunity lies in giving ordinary people who don't like tinkering the same AI value as tinkerers.

Distribution Is Never Uniform

The two discuss a subtler issue: distribution is never uniform. Rampell says many people talk about crossing the chasm, but this isn't a technology that exists in one place and not others—people all over the world use it, just unevenly. Glez gives the example of a vacation rental company: someone uses Town to compare photos of beach boards or beach buggies for damage, to decide whether to refund the deposit afterward. This wasn't on his wish list at all; what unlocked this need was a tinkerer at that small business who got annoyed by this and then worked with the Town assistant on how to automatically compare images. They see such usage problems in all sorts of odd corners of the economy, and this is exactly where the world lacks enough tinkerers—people with time to tinker with their own business. That's why AI hasn't spread. Even though Claude or ChatGPT have existed for 12 months and many small businesses could use them to save more time, they just treat them as a better Google search. The reason is that the product experience can't just be a text box; you can't teach people—you can only show them the value proposition and do it for them.

So why keep users on Town? They connect email and calendar, ideally also a business record system, and then build a custom integration that isn't out-of-the-box but is magical. That's also why distribution is uneven: in Silicon Valley everyone works with engineers and is closer to that natural hands-on ability, while most of the world doesn't have that. If you can bring a non-tinkerer into AI and deliver the same value as a tinkerer, you could build an extremely large company. AI can solve many business problems; intelligence is readily available—the key is how to deploy it effectively at scale.

Rampell mentions a very smart professor he met at Harvard who, after learning about AI, said he'd tried it a few years ago—it was okay, not great. Rampell had him rewrite a paper he'd written over 20 years ago using AI, and he said it was fantastic; upon learning that the AI was ChatGPT, he couldn't believe it. AI is progressing so fast that many people judge it by today's standards, while Silicon Valley tends to judge where it will go tomorrow, and a large group of people only see it as it is now. Part of the reason is ironic: early adopters tried it once two years ago, found it hallucinated, made mistakes, and wasn't good, and then got stuck in that version of reality. If they tried it again now, the result would be completely different.

Glez's judgment is that the ultimate victory lies in continuous development, constantly adapting the business to the best AI of the moment, and people's perception of what's possible, what's hard, and what's not will be completely overturned. He cites Computer Use and Browser Use as examples: over the past two and a half months, intelligent assistants have been able to do things that were almost impossible six months ago, with success rates rising from about 30% to 90%.

So when demoing with Town, he's picking things for you, and you'd be surprised—doesn't this require a user meeting? But these things aren't easy to sell; you have to be very specific: for example, I can use your existing paid Canva subscription to make the kind of posters you've always made with Canva—do you want me to? Five minutes later, for just 25 cents, Town has produced something very valuable. Those increasingly impossible-looking things are really just about riding the tide of what current models can do for you.

Agent Traffic Rewrites the Advertising Contract

Rampell then steers the conversation to agent traffic. He says broadly defined agent traffic has already far exceeded human traffic. He uses Yahoo Finance as an analogy: Yahoo pays to get stock quotes from places like Nasdaq and offers them to you for free, on the condition that you give it your attention and see E-Trade ads. But that contract has been broken; the two parties to the bundled contract are in opposition, but there's also an opportunity for synergy. This is like a whole new SEO: if you can become the recipient of agent traffic, not blocking it but thinking about how to welcome it, that's a huge opportunity. By this logic, Amazon's loss should become Walmart's gain, because Walmart has no advertising business and its incentives are different.

Glez follows up: the most interesting part is what happens if the agent's information is no longer neutral? For example, if you use Meta's agent—Meta has a huge number of advertisers, and United Airlines spends more on Meta ads than any other airline—when you ask it to book a flight for you, what happens once you feel it's no longer giving neutral information? Human motives always exist; there will always be people who want to sell and market. So he believes you need a neutral agent that belongs to you, with no other source of funding—your wallet is its only source of funding—so it has the incentive to give you the best information, undistorted by marketing forces.

He also talks about adversarial versus non-adversarial, framing it within the experience: some friction is enjoyable, like browsing between stores while shopping; other friction he doesn't want to spend time on, like ordering toilet paper at home—he just wants to get the same item into the cart as quickly as possible. But people aren't always maximizing efficiency. If he were Amazon, facing roughly $70 billion a year in advertising business, he'd say you shouldn't block agents, but not blocking them would create greater pressure, so you need to create a different experience where you can actually see ads. However, AI is developing so fast that it will change distribution and people's habits at a pace no one can predict, and there will surely be some spectacular failures along the way.

Another analogy he finds interesting is Muse, a free personal agent. People will think Muse will eventually have to run ads, but maybe it doesn't need to at all—maybe Muse just provides data so it can better target you on Instagram. When users open Instagram, it's not for efficiency but for fun, so monetization and agent experience are separated. And platforms like Amazon are at a disadvantage, at least on the agent side. If you're just a supply aggregator, that may no longer be a good business model; the real response should be to own your own personal agent.

He adds an example: there's a website called CamelCamelCamel that tracks Amazon prices, used by 20 million people in the US alone. You tell it you want to know if Dr Pepper soda drops from $8 to $6, and to tell you when it does—it won't place the order for you, just send you a document. Some people have more time than money, some have more money than time; those with more money than time won't use CamelCamelCamel, only those with more time will. If you now hand these tools to the agent of every person with money but no time, of course they'll use them. This is practically an IQ test. What all this has in common: the value of supply aggregation has greatly diminished, because human attention is limited, and now it can be automated.

Looking at the whole conversation, Glez's path is actually very clear: a good business that worked in the old world may not be able to grow into a venture-scale company after AI compresses the advantage window, so he chose to shut it down proactively, complete the pivot in one month, and turn that customer complaint into Town. His judgments about the industry are also on the table: distinguish customers from users, keep model supply neutral, leave ownership of data, memory, and context to users themselves, and the traffic brought by agents is rewriting the old advertising contract supported by eyeballs. Whether all these judgments hold true will take time to verify, but one question is worth leaving to everyone: when your agent's interests conflict with the platform's agent, what will you choose? Continue using that free assistant that can be distorted by marketing, or be willing to pay a little for neutral intelligence?