AI vendors love to pitch outcome-based pricing. In Orb's count, 3.8% actually sell that way.
Billing platform Orb looked at 80 AI agent companies and found outcome-based pricing at 3.8%, while usage-based hit 91.3%. Even the per-resolution deals sold as outcomes mostly measure whether the customer stopped asking.
Every AI agent company has some version of the same pitch now: you only pay when it works. It's a good pitch, because it answers the question every buyer actually has. Is this thing going to do the job, or just sit in the stack costing money? That said, the pricing data tells a different story than the pitch decks.
Orb, a billing platform built for usage-based pricing, analyzed the pricing models of 80 AI agent companies for its 2026 report. Outcome-based pricing came in at 3.8%, down slightly from 4.5% in 2025. On a sample of 80, that's about three companies. Usage-based pricing climbed from 83.3% to 91.3%, and 95% now use hybrid models. Orb sells the plumbing for usage pricing, so read it with that in mind. But no margin of error closes a gap that wide.
Orb's explanation is the useful part: "The gap is not about interest. It is about instrumentation." Defining, measuring, and attributing an outcome reliably is hard, and most billing systems were never built for it. Every PM knows a version of this from writing acceptance criteria. An outcome has to be something the customer would recognize and agree happened, not just something your own system decided to count.
Look at two companies that sell per-resolution pricing. Zendesk charges $1.50 per automated resolution, triggered when its AI agent resolves an issue with no human involved. Zendesk says a resolution is often defined by a 72-hour quiet period where the customer doesn't reopen the ticket. Intercom's Fin charges $0.99 per resolution. That counts when a customer confirms the answer helped, or when they leave without asking for more help. Intercom calls the second kind an "assumed resolution."
Both are real improvements over paying for seats nobody uses. And Zendesk's waiting period is a smart touch, because it checks that the answer held up. But both measure the same thing: the customer stopped asking. That's a support event. Whether the customer's problem actually got solved, or whether they just gave up, is a different question, and the invoice doesn't ask it.
I built product in micro-lending, where "outcome" has a meaning everyone agrees on without a meeting. The loan got repaid, or it didn't. Nobody prices lending risk off how confident the underwriting model felt when it approved the loan. The outcome gets checked later, by an event the lender doesn't control. Per-resolution pricing gets closer to that than per-seat pricing does. It's still the vendor's own system deciding what counts.
So if you're building the pricing model, three questions decide whether you can honestly call it outcome-based. Would the customer independently agree the outcome happened? Is it checked after enough time that it actually stuck? Can someone outside your company audit the count? A no on any of them means you're selling usage, priced per unit of visible work. That's fine. Just call it that.
If you're the one buying, run the same test before you sign. Ask what counts as the outcome, who verifies it, and whether that check happens before or after the customer has had a chance to disagree.
Pricing is a product decision before it's a finance one. Call an event an outcome and the argument about what "done" means moves from the roadmap to the invoice.