Why AI pricing keeps getting rebuilt:

The Observation

Almost every AI company we work with has rebuilt its pricing at least once, and the average is closer to twice. This is a structural pattern with a timing repeatable enough to plan for: Series B is when most companies discover that the pricing they set at Series A does not survive contact with an enterprise procurement team.

Why Seat Pricing Fails First

Seat pricing was the operating standard for a decade of SaaS. It works when the software is a tool used by a person to do a job. It works less well when the software is a system replacing significant portions of the job itself. Bessemer's AI pricing playbook makes the point directly. Seat pricing under-captures value in categories where one AI seat replaces several human seats. The awkwardness becomes visible the moment a customer answers "how many seats?" with "we are still working that out."

Andreessen Horowitz's 2024 enterprise research documented the same shift across its portfolio, describing AI-native companies moving away from seat-based SaaS pricing toward newer models organised around usage, output, or verified outcome. What all three have in common is that they attach revenue to something the product produces.

Why Inference-Cost Pricing Collapses Next

Companies that pivot from seats often anchor next on inference cost. Inference-cost pricing is transparent to customers and defensible in a sales conversation. It is also quietly ruinous to gross margin. As foundation-model prices fall (which they have done at a compounding pace over the last two years), any pricing tied to inference cost falls alongside them. The effect is that the AI company transfers most of the value of model commoditisation directly to the buyer. Whether that is generous or accidental depends on the company, and in our experience it is usually accidental.

The Model That Tends to Hold

The pricing model that tends to hold across the rebuilds we have seen is one anchored on the outcome the buyer would otherwise have paid a person (or another system) to produce. Three separate strands of published work converge on this point.

Bain's 2024 pricing work argues that AI is finally forcing the shift from seat pricing to outcome pricing that the SaaS industry has been talking about for a decade. BCG describes the same model through a different frame in its 2025 work on B2B pricing in the agentic AI era, calling outcome pricing "payment on jobs completed." Stripe's guidance for AI companies is the most operational of the three: price against a defined business outcome (a ticket resolved, a meeting scheduled, a document produced) and instrument the deployment so the outcome can be measured on both sides of the contract.

The Reason Outcome Pricing Is Hard

Outcome pricing is harder to construct than seat pricing because it requires the AI company to know what its output is worth to a specific buyer segment in the buyer's own business terms. Cost-to-produce is a separate calculation that belongs elsewhere on the finance team's spreadsheet. Very few AI companies have done this piece of commercial work by Series A, which is why most of the rebuilds happen later.

It is also harder to sell in the first year, because enterprise buyers are still comparing against seat-priced comparators from earlier vendors. Metronome's 2025 field report on AI pricing notes that fully outcome-based pricing remains uncommon at scale, partly because procurement teams are uncomfortable tying spend directly to non-deterministic outputs. Uncomfortable does not mean unwilling. In our experience it means that the AI company has to do more of the framing work than a comparable SaaS company would have done five years ago.

What The Companies That Get This Right Have in Common

Across the AI companies we have watched navigate this well, three practices show up consistently.

The first is that they price against a specific business outcome from the first commercial conversation, and they have that conversation with the person whose budget is affected by the outcome. That is usually one or two levels up from the champion who will end up using the tool.

The second is that they hold the line on outcome pricing under procurement pressure. Procurement will always bring up seat-priced comparators, and most first-year AI sales teams give ground because they have yet to internalise that their product does not belong in the seat category. Companies that hold the line for the first eighteen months tend to compound margin as the category matures. Companies that give ground rebuild pricing later, at higher cost, having conditioned the market on the wrong anchor.

The third is that they instrument enough of the deployment to prove the outcome quantitatively at renewal. Outcome pricing that cannot be measured is a story the sales team tells. Outcome pricing that can be measured is a moat. The instrumentation work has to be part of the first deployment; treating it as a project for month ten is a common way to lose the argument at renewal and slide back into seat comparisons. Companies that instrument early tend to renew at north of 130% net revenue retention. Companies that do not tend to see renewal margins compress.

A difficult conclusion

Pricing is where most of the enterprise value of an AI company is either captured or given away, and it is one of the few pieces of the commercial motion that gets meaningfully harder as the category matures. It is worth rebuilding a second time, and often a third, because each rebuild is a sign that the market has moved.