AI Startup Business Models: Why the Economics Aren't SaaS

Most founders building on AI borrow their business model from software, and most of what they borrow still works: recurring revenue, seat-based plans, a free tier that feeds a paid one. But one assumption underneath all of it does not survive the move, and it changes how the whole business should be priced.

Traditional software costs almost nothing to use once it is built. The thousandth customer query costs the company roughly what the first one did, which is close to nothing. An AI product is different. Every time a customer uses it, the company pays for the computation — the inference — that produces the answer. Usage is a cost line, not just a sign of engagement.

That single difference explains most of what goes wrong with AI pricing. It means your most enthusiastic customers can be your least profitable ones, that a flat monthly price quietly subsidises heavy users, and that falling model costs help you and your competitors at the same time. What follows works through each of those, with the arithmetic, and ends with the questions worth answering before choosing a model.

The cost that software never had

In a classic software business, gross margin is high because the cost of serving one more customer is small: some hosting, some support. Pricing can therefore be almost anything the market will bear, because nearly all of it falls through to gross profit.

When each use triggers a paid model call — whether you run the model yourself or pay a provider per token — cost of revenue rises with usage. The more valuable customers find the product, the more they use it, and the more it costs to serve them. That is the opposite of the relationship most software pricing is designed around.

None of this makes AI products bad businesses. It makes them businesses whose margin depends on the relationship between price and usage, which is something a founder has to design rather than assume.

What flat pricing does to your best customers

Hypothetical example. An AI writing tool charges $30 per seat per month. Assumptions, all illustrative: each task the tool performs costs the company $0.04 in inference; half of users run about 100 tasks a month, 35% run about 300, and 15% run about 900. No other variable costs are included.

The light user costs $4 to serve, leaving a gross margin of 86.7%. The median user costs $12 — a 60% margin. The heavy user costs $36, so the company loses $6 a month on every one of them: a margin of minus 20%.

Blended across all three groups, the average cost to serve is $11.60 and gross margin is 61.3%. That number would look acceptable in a board pack. It hides the fact that the product's most committed users — the ones most likely to recommend it, and the ones whose usage tends to grow — are the ones losing money. At $0.04 a task, any seat that runs more than 750 tasks a month is served at a loss.

This is the same lesson that shows up in what one retailer's accounts say about how businesses make money: the blended figure tells you what the business sells, and the segment tells you where it actually earns — or loses.

Three ways to price it, and what each one costs you

Each pricing model decides who carries the cost of heavy use. None removes that cost; they only move it.

Per-seat pricing is simple to buy and easy to forecast, which is why it survives. Its cost is the one in the example: it transfers usage risk entirely onto you. It works best when usage per seat is naturally bounded, or when you can cap it with fair-use limits that customers accept.

Usage-based pricing — per task, per document, per thousand tokens — aligns revenue with cost, so heavy users become your most profitable customers rather than your least. Its cost falls on the customer's side: bills become unpredictable, finance teams dislike that, and some customers will ration their own use of your product to control spend, which undercuts the adoption you wanted. The trade-offs of recurring versus per-use billing are covered in more depth in subscription pricing: what you gain and what you owe.

Outcome-based pricing — charging per resolved support ticket, per qualified lead, per completed review — is the most persuasive to a buyer, because they pay only for results. Its cost is that you carry both the usage risk and the performance risk, and that every dispute about what counts as an "outcome" becomes a billing dispute. It suits products whose results are unambiguous and measurable, and few are at the start.

Most AI products that price well end up with a hybrid: a base subscription that includes a usage allowance, with metered charges above it. That keeps the bill predictable for most customers while stopping the heaviest from being served at a loss.

Falling costs cut both ways

The cost of inference has fallen dramatically. Stanford's 2025 AI Index reports that the price of querying a model performing at the level of GPT-3.5 fell from $20 per million tokens in November 2022 to $0.07 in October 2024 — a reduction of more than 280-fold.

For a founder, that trend is good news and bad news in the same sentence. In the example above, halving the cost per task would lift the heavy user from minus 20% to a 40% margin. But the same decline is available to every competitor, and it steadily lowers the price customers expect to pay. A business whose only advantage is access to a capable model has an advantage that is getting cheaper for everyone else to copy.

That is why the durable part of an AI business is rarely the model call itself. It is the workflow around it, the proprietary data it learns from, the integration into how a customer already works, and the trust that it does the job reliably. Pricing should reflect that value, not the cost of the tokens underneath it — otherwise every cost decline becomes a price cut you did not choose.

It is also worth being clear about what this trend is not: a forecast. Past declines in inference costs say nothing reliable about the pace of future ones, and a business plan that depends on costs continuing to fall is carrying a risk it cannot control.

Where the demand actually is

Founders often assume AI adoption is already universal among their prospective customers. The best available US data says otherwise. The Census Bureau's Business Trends and Outlook Survey found that between December 2025 and May 2026, overall AI use among US businesses hovered between 17% and 20%. Use was far higher among larger firms — 37% of firms with 250 or more employees — and did not change significantly among firms with fewer than 20 employees.

That matters for the business model, not just the sales plan. A product aimed at small businesses is selling into a market where most prospective customers do not yet use AI at all, which means longer sales cycles, more education, and higher customer acquisition costs. A product aimed at larger firms meets buyers who already use AI and are comparing options, which means more competition and more pressure on price. Neither is easier; they are expensive in different ways. The arithmetic of what it costs to win a customer is set out in customer acquisition cost, explained with real numbers.

These are US figures from one survey, and adoption elsewhere will differ. They are also a snapshot of a regularly updated series, so check the latest release before relying on them.

Three questions before you choose

What does your heaviest user cost you? Not the average — the top 10% or 15%. If that group is loss-making at your planned price, the model needs a usage allowance, a metered tier, or a different price.

What would a competitor with the same model charge? If the answer is "less than you, and soon," the price needs to rest on something other than the model: data, workflow, integration, or reliability.

What happens to your margin if your model provider changes its prices? Dependence on a single provider is a supplier concentration risk like any other. Know how much of your cost base it represents and how quickly you could switch.

None of these has a universal answer. They are the questions that decide whether an AI product is a software business with unusually good distribution or a services business with a software interface — and those two have very different economics.