Seth's Blog : Building a second-wave AI business

Building a second-wave AI business

A very long rant, riffing on the opportunity for bootstrapped startups who seek to create value using AI. I wrote it a while ago, thought it was too long, but in arguing with an AI today (everyone needs a hobby) I realized it was worth sharing:

Most AI success stories to date are about cost reduction or speed improvement. A startup offers businesses a way to get more done with fewer people, replacing customer service or programming teams with bots. The upside of cost reduction is that it’s a very easy sale—give the client a free sample, once it’s demonstrated to work, they have an instant benefit in switching.

The downsides: it’s difficult to win a race to the bottom, since someone can always promise more savings than you. And it’s finite—once the savings are made, there’s no incremental value left to create.

The opportunity lies in something generative. A use of AI that doesn’t reduce costs, it creates value. It opens new opportunities, leads to growth, connection, and utility.

Worth paying for: Most bootstrappers target price-sensitive customers and then wonder why growth is hard. But people and organizations with expensive problems and real resources don’t just put up with paying more for things they value—they prefer it. Premium pricing signals seriousness. Look for a market where the problem is real, the budget exists, and the solution creates something they couldn’t get otherwise.

What people actually pay for: At the foundation of almost every premium purchase are three drives:

status (I matter, people like me see me as significant),

affiliation (I belong, there are people like me and they accept me), and

freedom from fear (I am safe, the threat is not coming).

Freedom from fear may be the most primitive—you can’t pursue status or affiliation while in survival mode. And most premium purchases are a quest for freedom from fear pretending to be something else.

Built on those roots is a middle layer of things that offer one or more: legitimacy, transformation, belonging to a narrative, control, certainty, protection, trust, health and longevity, leverage.

And the outer layer that’s easier to measure—things people buy because they deliver the middle layer: access, capital, time, attention, convenience, efficiency, delight, new experiences, beauty.

Commodities—food, shelter, sex, addictive substances—are often outside this hierarchy. They don’t really build toward the three roots; they allow survival or temporarily suppress the anxiety that comes from not having them.

An AI business worth building delivers something from the middle layer, justified by the outer layer. Nobody goes shopping for transformation.

What businesses actually pay for: The hierarchy for individual consumers doesn’t translate directly to organizational purchases. In B2B, the customer is spending someone else’s money. That means that the dominant question they’re asking is, “what will I tell my boss?” Three desires sit at the foundation of almost every business buying decision:

Avoid blame — if this goes wrong, it won’t be my fault. The IBM principle: nobody ever got fired for buying the market leader. The champion inside the organization often needs a defensible story before they’ll act.

Claim credit — I brought something in that worked and people noticed. The flip side of blame avoidance, and the engine of the internal champion. If your solution lets someone look good, they’ll sell it to their peers, you won’t have to.

Reduce uncertainty — we can plan around this, the chaos goes down. Organizations pay significant premiums for the ability to forecast, commit, and stop worrying.

Built on those roots is a middle layer of things organizations reliably spend on: growth, efficiency, compliance, competitive advantage, talent, morale, resilience, optionality, speed, legitimacy, relationships.

And an outer layer that justifies the middle: cost savings, time savings, data, access, convenience, integration, reporting, support.

Mechanics without a story is the race to the bottom, and being the cheapest is not the best use of your time.

New vs. repeat purchases require different approaches: Repeat purchases are won by switching costs, relationships, and relentless incrementalism—you’re replacing someone, which means you need to be cheaper, easier, or have a better story and sales force. New purchases require someone inside the organization to become a champion, which means they need a story that serves their career, not just their company’s interests.

Not all problems are equally interesting: Some purchases—like gaining market share or entering a new category—are chaotic and interesting, with room for narrative and ambition. Others—like cheaper materials or faster processing—are grinding commodities where the only story is price. Commodity buyers fear paying too much. Buyers in chaotic spaces fear making a wrong choice.

The forcing function: Businesses rarely lead the way on new purchases without a crisis compelling them. Without a forcing function, even a perfect solution sits in the pipeline forever—committees form, pilots stall, and champions get reassigned.

Three kinds of crises create forcing functions:

Competitive crisis — a rival did something and now there’s urgency. “They have it and we don’t” is a sentence that ends a discussion and starts the buying process.

Technology crisis — the old way stopped working, or a new capability made the old way look reckless. AI itself is currently creating this for many industries simultaneously. This time, the forcing function and the solution are the same thing.

Public/market upheaval — regulatory change, cultural shift, a collapse in input costs, a pandemic. These are the most powerful and least predictable. They create entirely new categories of buyer.

The opportunity for a bootstrapper: sell into a forcing function that already exists, don’t try to create one. Organizations already feeling the crisis don’t need convincing—they need a solution that lets their champion say “I found it.”

NOTES:

Naked AI is a trap. If all you’re doing is building a gateway to Anthropic or ChatGPT, your token costs eat a significant portion of your revenue—and you have no defensible position.

Hidden prompts are insufficient. Breakthrough prompting can create real value, but there’s no protectable, reliable way to sell it as a business. If one of the frontier companies made it a business model, the mechanics would work in the bootstrapper’s favor, but I haven’t seen this.

The network effect matters. Selling benefits one person at a time is brutally expensive. The breakthroughs come with projects that have the network built in—where interactions work better when your colleagues are using them too.

Asymmetric information is worth seeking out. Some of the most durable advantages come not from network effects but from knowing something others don’t, or from helping a cohort work together to pool what they know against a party that currently has structural information advantage over them. Let all of Walmart’s vendors see information that they currently hoard, for example.

So, a theory of profit—a framework for the kind of project that becomes a business:

Creates its own useful data stack. The data doesn’t need to be large to be valuable—it needs to be specific and trusted. Over time it informs the AI. It belongs to users and the project, not to Anthropic or competitors. And it’s built to work for users, not to trap them.
Has a built-in network effect. Either an engaged peer-to-peer community (where users see each other, not just the platform) or an obvious benefit to spreading the word.
Solves an expensive problem for people with resources. The value delivered goes beyond saving time or money—it might be education, reduced fear, joy, reassurance, connection, or capability expansion. And it’s priced accordingly.
Is bootstrappable. Specific and conceptual rather than infrastructural. No data centers, no thousand-person teams required to get started.

Bonus:

A note on data stack reality. A network built on user data is only as good as the willingness of users to populate it. And willingness requires two things: it has to be frictionless enough that people don’t have to think about it, and it has to feel safe enough that people don’t have to worry about it. These two conditions are almost always in tension. The more automatic the data collection, the more it feels like surveillance. The more control you give people, the more friction you add.

The most promising data stacks are ones where people are already generating the data, are already comfortable with it existing somewhere, and the innovation is simply giving them better access to what’s already theirs. The forcing function for consumer data sharing may be the simplest one of all: I already feel watched. I might as well get something back.

The cautionary version of this is the email surveillance tool—a business reads all internal email and gets a report on who’s helpful, who’s toxic, who’s looking for a job. The value is real and obvious. The fear is also real and obvious. And in most organizations, the fear wins. Any data stack business has to answer the question: who controls this, and what happens if it goes wrong? If the answer isn’t immediately reassuring, the business doesn’t get built.

The sponsored model. Not every valuable AI business needs the end user to pay. When the problem is real but the affected population lacks resources, a foundation, brand, or institution with aligned interests can fund the miracle instead. The economics flip entirely: instead of acquiring thousands of customers one at a time, you close one relationship with one institution that already has the distribution, the mission, and the budget. The user gets the miracle for free. The sponsor gets impact, data, or loyalty.

This model works when three things are true: the population being served is large and underserved, the value created is legible to an institution that cares about it, and the data generated serves both the individual user and the sponsor’s mission.

For example, a foundation pays $2,000,000 and 40,000 families of the incarcerated have access to a tool that generates a ten-page legal document instead of a bushel of random papers—and the shared data starts identifying patterns in the system (bad actors, defective paperwork) that no single case could surface alone.

A bank funds a personal finance tool for its own customers. A health brand funds a fitness coach for an underserved population. The viral problem is much easier to solve: once it’s free, you don’t need to work hard to persuade users to recruit each other, you need one institution with existing distribution to say yes.

The cheap inference model. Not every AI application needs a frontier model. The problems worth looking for here aren’t the ones that require reasoning or nuance—instead, look for structure, pattern recognition, aggregation, and organization at scale. Form filling. Document organization. Transcription plus summarization. Matching similar records across large datasets. These problems are unglamorous but enormous in volume and largely underserved.

Moore’s Law is on your side. The models that feel too limited today will evolve to become adequate in eighteen months. Building on cheap open-source inference means your margins improve as the technology does, without changing your product (which is the data stack and the network). And “huge” doesn’t mean huge—a single business school graduating class is enough to populate a meaningful census of what jobs actually lead where. The data stack doesn’t need to be large. It needs to be specific, trusted, and ahead of what anyone else has assembled.

This was a particularly long rant, thanks for hanging in. I started writing it for a friend six months ago (with many inputs from others), but it’s more true now than then.