Is a Million-Dollar SaaS Not Enough? Startup Logic Has Changed in the AI Era

From Tax Tool to AI Assistant: A Million-Dollar Self-Negation

In an era where AI is reshaping the software industry, a SaaS business with annual revenue approaching one million dollars might have been a decent achievement five years ago. But Jean Denis, former CTO of Plaid and founder of Town, uses his personal experience to tell us: such results may not even be worth pursuing in the AI era.

After leaving Plaid, Denis spent a year building a tax automation product for small and medium-sized businesses. The product itself was successful—revenue approached one million dollars, with projections of 2.5 to 3 million dollars the following year. Before the AI wave, this number would have delighted both investors and founders. But Denis and his team chose to shut down the project after the first tax season ended and pivot to a completely new direction.

Behind this decision is a fundamental shift in startup logic in the AI era: technical feasibility of a product is no longer the biggest obstacle; the speed of channel penetration and the window for network effects are the real lifelines.

The Ceiling of the Tax Business: The Fatal Flaw of Selling Only Once a Year

Denis's tax product didn't fail because of product quality. LLMs excel in scenarios like tax processing that involve large amounts of text and structured input, and technical validation was already complete. The problem lay in the business model.

Tax is a typical "annual" business: you can only sell once a year, interaction frequency is extremely low, and it simply cannot become a daily-use tool. Worse is the paradox of word-of-mouth: no one boasts about their tax filing experience in social settings, and even if it comes up, it can only happen during the once-a-year window after tax filing is complete. By then, the listener is neither in a purchasing decision period nor likely to remember the conversation a year later.

Andreessen Horowitz partner Alex Rampell added an analogy in the conversation: edtech faces the same dilemma—only two sales windows each year, summer and year-end, and this cyclical discontinuity makes it nearly impossible for a company to achieve hyper-growth.

"We built a good business before AI, but after AI, it's not a venture-capital-grade business anymore," Denis summarized.

The key insight here is: the competitive clock for AI startups is accelerating. You not only have to compete with incumbents on product, but also on who can reach millions of users and build network effects before rivals use their distribution advantages to copy you. The once-a-year sales rhythm of the tax product made this goal seem unreachable.

Network effects refer to the phenomenon where a product's value increases as the number of users grows (e.g., social networks, marketplaces). In AI startups, network effects are seen as key to building competitive moats—once users form connections or accumulate data, it's hard for later entrants to replicate. Denis emphasizes that the window in the AI era is extremely short; if a product cannot form network effects through high-frequency interaction in a short period, it is easily copied by giants with distribution advantages. Because of its once-a-year low-frequency interaction, the tax product cannot generate connections between users or a data flywheel, making it difficult to build network effects and thus losing the hyper-growth potential that venture capital demands.

Execution Details of the Pivot: Timing Matters More Than Courage

Shutting down a still-growing business requires courage, but Denis's decision process reveals a more pragmatic factor: the time window.

The cyclical nature of the tax business actually became a catalyst for the decision. For about a month around October 15, the team could simultaneously live in two parallel universes: "continue doing tax" and "don't do tax." Once past this window, with partners signed and commitments made to customers, the door would close. This hard time pressure actually helped the team avoid the procrastination trap of "let's try one more month."

Denis admits the hardest part was facing employees and admitting failure. But he quickly realized: friends won't leave because of failure, employees can be treated well through proper severance and referrals to next opportunities, and "once you figure these out, the decision is not as emotionally difficult as the initial reaction."

The Birth of Town: Why Has No One Built a First-Class General AI Agent?

After the pivot, Denis's team seriously explored four directions. What ultimately led them to Town was a simple question: why has no one built a first-class AI for the most common workflows of knowledge workers?

The clue came from customers of the tax product. They repeatedly mentioned: automatic tax filing is great, but the bigger pain is forgetting meeting content before a meeting, clients always asking the same tax questions, and needing to prepare answers in advance. These needs are not complex, but there are no good products on the market.

Coinciding with the Opus model demonstrating stronger agentic capabilities, Town's product-market fit came very quickly. During the closed beta from December to April of the following year, the product grew steadily at 1% to 1.5% day-over-day, customers kept proposing new requirements, and the roadmap was exceptionally clear.

Agentic capabilities refer to AI models not only generating content but also autonomously planning, invoking tools, and executing multi-step tasks to achieve goals. The Opus model (Anthropic's Claude 3 Opus) improved reasoning and tool use, enabling AI to proactively handle complex workflows such as automatically scheduling meetings and answering customer questions. This capability is the technical prerequisite for Town's product-market fit—without agentic capabilities, AI can only respond passively; with them, AI can become a proactive "assistant" for knowledge workers, covering high-frequency scenarios like meeting preparation and customer Q&A.

Product Philosophy: Build for the User, Not the Buyer

Town made a key choice in product strategy: prioritize serving users rather than paying customers.

Denis drew a judgment from his experience at Dropbox: starting from a world-class product experience, it is feasible to later adapt to the enterprise market; conversely, starting from compliance and IT requirements, it is nearly impossible to build an excellent product experience.

Behind this strategy is a keen observation: consumer AI experiences are raising everyone's expectations. When CEOs and CIOs experience powerful AI assistants in their personal lives, they will demand the same level from enterprise software. Therefore, whoever can deliver a consumer-grade experience in enterprise scenarios gains an advantage.

Rampell reinforced this view with an analogy: BlackBerry died not because iPhone did something right, but because CEOs received iPhones as Christmas gifts, took photos of their kids with them, and then returned to the office saying, "Go get this thing into the company." Changes in enterprise software are often driven by executives' personal experiences.

Neutrality as a Moat: Not Locking into a Model Provider

Facing the threat of model providers like OpenAI and Anthropic expanding into the application layer, Town's differentiation strategy is absolute neutrality.

Denis's sales pitch to mid-sized enterprise customers hits the pain point directly: you cannot predict which model provider will be strongest in three to five years. If you buy OpenAI's product, you will forever be limited to OpenAI's models; if OpenAI raises prices, you have no choice. Town, as a neutral application, selects the most suitable provider in real time based on the specific task—text, voice, video, slides.

"We are a neutral application. We will use the provider we believe is best for you at the moment," Denis said. "Our job is to help you judge, from all AI capabilities, what is the right choice for your business."

The second layer of the moat is data ownership. Town's position is: the context, memory, and contextual assets generated by the enterprise should belong entirely to the enterprise. If a better solution emerges one day, switching should take only a day.

This contrasts sharply with the FDE (Forward Deployed Engineer) model. Denis believes that the FDE model essentially increases customer switching costs, and as model capabilities improve, such highly customized solutions will actually become worse. Town chose a path more reliant on productization and automation, betting on the long-term advantage brought by the learning flywheel.

The FDE (Forward Deployed Engineer) model is a model that provides deep customization services to clients, with engineers embedded on-site or remotely in client teams to develop solutions for specific needs. This model can quickly meet the needs of large clients, but the cost is high customer switching costs due to high customization and difficulty in productization. Denis believes that as model capabilities improve, customized solutions will gradually fall behind because general products can continuously improve through automated learning flywheels. Town chooses the productization path, betting that standardized products can serve a broad customer base more efficiently than relying on labor-intensive custom services.

From Horizontal to Vertical: Let the Market Tell You What to Do

Town's market strategy is to start as horizontal as possible, entering through PLG and SMB, mid-market, and discovering vertical opportunities through user behavior data.

The specific approach: do not directly look at user data (an enterprise privacy requirement), but analyze which users are power users, then interview them to find common needs, and manually build prototype automations (called routines). When new users sign up, the product determines the user type through email and calendar and recommends corresponding templates.

A vivid example is a restaurant owner: every month, wholesalers send Excel sheets with prices for thousands of items, and almost no one seriously compares price changes. But Town can run an analysis once a month, and if in one out of twelve months it finds a price adjustment opportunity that saves a few hundred dollars, the cost of one dollar per month is completely worth it.

This case reveals the root cause of uneven AI product diffusion: the world needs enough tinkerers (people willing to tinker) for AI to take hold, but most industries lack such people. Giving non-tinkerers the same value as tinkerers is a huge business opportunity.

The Future of the Agent Economy: Humans Are No Longer Slow Routers

The most forward-looking part of the conversation was about how agents will change enterprise information flow.

Town's vision is to build an internal CRM-like feature: when an employee asks, "Does anyone know someone on this prospect's finance team?" the AI can converse with other employees' agents and instantly find connections. In the human mode, this question would be posted on Slack or Teams, and everyone in the company would read, think, and reply—what Rampell calls "humans as slow routers."

Denis acknowledges this requires extremely careful privacy design, but his judgment is: in future work scenarios, human in the loop will be significantly reduced.

On the relationship between agent traffic and advertising models, the two discussed Amazon's dilemma. When agents shop on behalf of users, Amazon's advertising revenue is threatened. Rampell believes that leading platforms should build their own personal agents and decouple monetization from the agent experience, rather than simply blocking agent traffic. Denis points out that the essence of the problem is the declining value of "supply aggregation"—when human attention can be automated, a business model that merely aggregates supply will no longer be solid.

Human in the loop refers to retaining human review or decision-making steps in automated processes to ensure safety and controllability. In the agent economy, as AI agent capabilities grow, much information transfer and decision-making can be automated by agents, and humans gradually shift from executors to supervisors. Denis predicts that human in the loop will be significantly reduced in future work scenarios because agents can handle cross-departmental information queries and coordination more efficiently. But this also brings privacy and security challenges, requiring finely designed permission and audit mechanisms to ensure agent behavior is transparent and controllable.

Lessons for AI Entrepreneurs

Denis's experience offers several counterintuitive lessons:

1. Million-dollar revenue is no longer a safety signal. A business that looked healthy before AI may not be large enough after AI. The criterion shifts from "is the product viable" to "can it achieve network effects within the window."
2. Cyclical businesses are inherently unsuitable for venture capital. Products that can only be sold once a year with extremely low frequency, no matter how AI optimizes them, will find it hard to establish a hyper-growth curve.
3. Product experience takes precedence over enterprise compliance. Build a great product first, then compromise for enterprise needs; the reverse almost never works.
4. Neutral positioning is a strategy. In an environment of rapid model iteration, not locking into a single provider can be a value proposition for enterprises.
5. The bottleneck of AI diffusion is product experience, not model intelligence. Most people don't need smarter models; they need someone to package AI capabilities into concrete services: "Do you need this? If so, I'll help you do it."

One-line Summary

Former Plaid CTO abandons a million-dollar tax SaaS to pivot to general AI agent Town, revealing the shift in startup logic in the AI era.

Paragraph Summary

This article tells the story of Jean Denis, former CTO of Plaid, who proactively shut down a tax automation product with annual revenue approaching one million dollars amid the AI wave and founded the general AI assistant Town. The article analyzes the fatal flaw of the tax business—its low-frequency interaction prevents network effects—and the reality of an accelerating competitive clock in the AI era. Town enters through high-frequency scenarios for knowledge workers and quickly achieves product-market fit thanks to the Opus model's agentic capabilities. Its product strategy prioritizes serving users over buyers, builds a moat through a neutral model strategy and data ownership, and discovers vertical opportunities through a horizontal-to-vertical market strategy. Finally, the article explores the impact of the agent economy on future work information flow and summarizes five lessons for AI entrepreneurs.