How should a startup price an AI product?

Start from the value you deliver and the cost you carry, not from what SaaS competitors charge. AI pricing is different because the product often does the work a person used to do, so you are competing with the cost of labor, not the cost of other software. The cleanest place to begin is per-unit pricing (per document processed, per call handled, per ticket resolved), which is easy for a customer to understand and ties revenue to real usage. As you learn what an outcome is worth, layer in value-based or outcome pricing. Whatever you pick, price above your cost of goods sold, which now moves with usage, and leave room for the token cost to fall over time.

Is usage-based or seat-based pricing better for AI?

For most AI products, usage-based pricing fits better than per-seat pricing. Per-seat pricing assumes value scales with the number of people logging in, but when an AI agent does the work, a customer may need far fewer seats while getting far more output, so seats stop tracking value. Usage-based pricing (per call, per token, per unit of work) tracks value more honestly and matches the variable cost you pay the model provider on every request. The trade-off is predictability: usage-based revenue is harder to forecast, so many founders run a hybrid, a base platform fee plus usage, to keep some floor under the number.

What is outcome-based pricing for AI?

Outcome-based pricing charges for a finished result rather than for access or usage: a resolved support ticket, an approved loan, a completed filing. It aligns your incentives with the customer's, because you only get paid when the work lands. That is powerful for AI products that deliver a discrete result, and it can command a premium because you are pricing against the labor you replace. The catch is that outcomes are harder to forecast and can expose you if your cost of goods sold per outcome is high or variable, so most founders reach for it once they understand their unit economics well enough to guarantee a margin on each result.

Why is AI pricing harder than SaaS pricing?

Classic SaaS had near-zero marginal cost, so once you built the software, another user cost you almost nothing and per-seat pricing printed high gross margins. AI breaks both halves of that. Every call has a real inference cost, so your cost of goods sold moves with usage instead of sitting near zero. And a lot of the value shows up as consumer surplus: the customer gets faster, cheaper work but does not always pay more for it, so productivity gains do not automatically become higher prices. On top of that, foundation models are commoditizing and competing on price, which pushes the durable margin up the stack into the layer you build on top.

AI Pricing: How to Price Your AI Product

If you are building an AI product, "how do I price this" is one of the first hard questions, and the old SaaS answer does not work. AI pricing is genuinely different from software pricing, and pretending otherwise is how founders end up with negative-margin customers they cannot afford to keep. Tech analyst Benedict Evans laid out the economics in an a16z conversation, "The Economics of AI Usage and What's Next For SaaS," and the founders building AI-native companies are converging on the same lessons in practice. Here is the founder-to-founder version.

The short answer

Price an AI product from two numbers: the value it delivers and the cost you carry to deliver it. That sounds obvious, but it is a real break from SaaS habits. In classic software you priced per seat, your marginal cost was close to zero, and gross margins took care of themselves. With AI, the product often does the work a person used to do, so you are competing with the cost of labor, and every call you serve has a real cost. The cleanest starting point for most founders is per-unit pricing tied to the work done, moving toward outcome or value pricing as you learn what a result is worth. The mistake to avoid is copying a per-seat price sheet from a SaaS competitor, because seats no longer map to value.

Why per-seat pricing is breaking

Per-seat pricing rests on an assumption: value scales with the number of people using the tool. That held when software helped people do their jobs. It breaks when the software does the job. If one operations lead points an AI agent at a workflow and it handles what used to take five people, the customer needs one seat and gets five people's worth of output. Charge per seat and you just told the customer that using your product better means paying you less. That is backwards.

This is the same shift we cover in service as software: when you deliver the finished outcome instead of a tool, you stop selling seats and start selling work. The pricing question follows the value. As the work moves from "people using software" to "software doing the work," the price has to move with it.

Your costs move with usage now

Here is the part that catches software founders off guard. In SaaS, your cost of goods sold was low and mostly fixed, so an extra user was almost free. In AI, every request carries an inference cost you pay the model provider, so your cost of goods sold moves with usage. The founders building AI-native services describe a cost stack with three parts: model costs, hosting, and any humans kept in the loop. Each of those is a real line you have to price above, not round to zero.

That changes the discipline. You cannot sign a flat unlimited plan and hope, because a heavy user can turn negative-margin fast. It also means the supply side is genuinely constrained right now. As Evans put it, describing the scarcity behind today's prices, "We are in this extreme scarcity. Like, we can't spend $10 trillion a year on AI infrastructure cuz there isn't $10 trillion a year there to spend on it." Compute is not free and it is not infinite, so your price has to respect the cost underneath it.

The good news is that the cost floor is falling. Token prices trend down over time even as usage rises, which the AI-native services founders call AI operating leverage: as you build more of the product and lean less on expensive models and humans, your cost of goods sold drops and your gross margin improves. So price for the margin you can defend today, and let falling costs widen it, rather than pricing at cost today and hoping.

The consumer surplus problem

The hardest truth about AI pricing is that a lot of the value you create never reaches your bank account. Evans's point is that AI can deliver enormous consumer surplus: the customer gets work done faster and cheaper, but increased productivity does not always translate into a higher price you can charge. If your product saves a customer ten hours and you can only capture a sliver of that, the rest is surplus they keep.

That is not a reason to despair, it is a reason to price deliberately. The founders who capture more of the value tend to price against the labor they replace rather than against other software. Jensen Huang of Nvidia frames AI as valuable precisely because it does work: "We're basically paying AI a lot of money today, the fastest-growing software business in the history of mankind." When your product genuinely does a job, anchoring the price to the cost of that job, an hour of a paralegal, a support agent, a loan processor, captures more of the value than anchoring to a SaaS seat ever will.

Models are commoditizing, so charge for the layer above

Evans's structural call is that foundation models are likely to become commodities rather than products, and that the durable value sits further up the stack. The founders comparing the leading models say the same thing from the ground: over time the base models converge on similar behavior and compete on price. That has a direct pricing consequence. If your product is a thin wrapper whose only advantage is the model underneath, you will get squeezed as that model commoditizes and cheaper equivalents appear.

The margin you can defend lives in what you build on top: the workflow, the reliability, the data, the trust with the customer. That is also where a moat lives, a point we make in why AI is compressing every moat. Price for the value of the layer you own, not for access to a model anyone can rent, because the model is the part that gets cheaper and more contested every quarter.

Four ways founders are pricing AI right now

No single model has won, and the market is still finding equilibrium. These are the four patterns founders actually use:

Per-seat. Still fine when a human uses your product as a tool and value really does scale with headcount. It breaks the moment the software starts doing the work instead of assisting it.
Usage-based. Charge per call, per token, or per unit of work. This tracks your variable cost and the customer's real usage most honestly. The cost is forecasting: revenue is lumpy and harder to predict.
Outcome-based. Charge for a finished result: a resolved ticket, an approved claim, a completed filing. It aligns incentives and can command a premium, but it exposes you if your cost per outcome is high or variable, so you need to know your unit economics cold.
Hybrid. A base platform fee plus usage or outcomes. Most founders end up here, because the base fee gives you a predictable floor while usage captures the value of heavy customers.

The right answer depends on how discrete your output is and how well you understand your costs. Start simple with per-unit pricing you can explain in one sentence, watch the margin on every customer, and move toward outcome pricing as you learn what a result is worth.

What to do this week

Write down your cost of goods sold per unit of work: model cost, hosting, and any human in the loop. If you cannot, you are not ready to set a price.
Check whether your current price survives your heaviest user. If a power user is negative-margin, fix the structure, not just the number.
Pick the unit you charge for: a seat, a call, or an outcome. Choose the one that tracks the value the customer actually gets.
Anchor your price to the labor you replace, not to a SaaS competitor's seat price. Capture more of the surplus you create.
Assume the model cost falls and competitors appear. Make sure the value you charge for lives in the layer you own, not the model you rent.

Pricing is one of the highest-leverage decisions an AI company makes, and it is the kind of cross-functional, AI-native call we work through in AI Operating System for Startups: understand the economics, price for the value you deliver, and keep a human on the judgment that matters.

Sources

The Economics of AI Usage and What's Next For SaaS, the a16z conversation with Benedict Evans that this article distills. Evans is an independent technology analyst and former a16z partner (2014 to 2019) whose weekly newsletter reaches a large founder and investor audience.
How to Build an AI-Native Services Company (Y Combinator), on per-unit versus outcome pricing, competing with labor cost, the three-part cost of goods sold, and AI operating leverage.
NVIDIA's Jensen Huang on why AI is valuable (Sequoia Capital), on paying for AI because it does work.
Differences Between The Leading AI Models (Y Combinator), on base models converging and competing on price.