Outcome-based AI pricing charges per resolution. Vendors decide what a resolution is.

Intercom's Fin AI Agent bills $0.99 per resolution. If a customer's problem gets fixed, that counts. If the customer simply stops replying a few minutes after Fin's last message, that counts too: an "assumed" resolution, built on the premise that silence means satisfaction. Outcome-based AI pricing is the industry's answer to two years of complaints about seat licenses that don't track AI usage and token bills that swing with how verbose a model happens to be that week. But every vendor building this model has to answer one question before anything else, and none of them answer it the same way. What, exactly, counts as an outcome?

Per-seat, per-token, and now outcome-based AI pricing

SaaS pricing has cycled through two failed AI-era models already. Per-seat pricing assumed a human sat at a desk using the software for a fixed slice of the day. An AI agent that resolves three thousand tickets a month doesn't fit that shape, and charging by seat count for something that runs unattended stopped making sense to buyers fast. Per-token and per-credit pricing followed, tying the bill to model usage instead of headcount. That solved the seat problem and created a worse one: a bill that moves with prompt length, retries, and model choice, none of which the buyer controls or can forecast. Finance teams that budget in fixed increments were suddenly reconciling invoices that looked more like a data centre's power bill than a software subscription. Some tried monthly credit pools to smooth the swings, and most buyers found the pools ran out mid-cycle at the worst possible moment, right before a renewal conversation.

Outcome-based pricing is the third attempt, and on paper it fixes both problems. You pay only when the AI actually does something: resolves a ticket, qualifies a lead, completes a refund. Fin charges $0.99 for a resolution, a procedure handoff, or a disqualification, and $9.99 for a qualified sales lead, with a stated minimum of fifty billable outcomes a month. Salesforce, Zendesk, and a growing list of vertical AI vendors have shipped some version of the same idea. Buyers report the appeal isn't really the price. It's the alignment. The vendor only gets paid when the thing you hired it to do actually happens, at least in theory.

Analysts who track enterprise SaaS purchasing describe outcome-based clauses moving from a rare ask two years ago to something close to a default request in new AI vendor negotiations today. That shift is coming from the buyer side, not the vendor side. Finance and procurement teams burned once by an unpredictable token bill are now writing outcome-based terms into their own RFPs before a vendor even proposes them. The pressure to define an outcome clearly, in other words, is arriving from the same people who will eventually dispute one.

What counts as an 'outcome' is a product decision, not just a pricing one

The alignment argument only holds if the outcome is defined the way a reasonable buyer would define it, and that's where the model gets complicated. Intercom's own documentation distinguishes a confirmed resolution, where the customer says something affirmative, from an assumed resolution, where the customer stops replying and Fin treats that silence as resolved.

Intercom's help centre spells out the mechanics of what counts as a billable outcome, including the confirmed-versus-assumed split.

That distinction matters more than it looks. A support conversation that gets genuinely resolved and one that gets abandoned mid-frustration produce an identical signal to a pricing engine that only reads for silence: no further messages. The AI agent has no reliable way to bill differently for a satisfied customer and an exhausted one, because both look the same in the transcript.

Vendors that have thought hardest about this split outcomes into two categories: hard and soft. A hard outcome is binary and provable from a system log, such as a refund that was issued or a ticket closed with a specific status code or a lead written into the CRM with a matching field. Nobody can argue about whether it happened. A soft outcome depends on a judgment call, such as whether the customer was actually satisfied, and judgment calls are exactly the kind of thing that produces a dispute whenever the vendor and the buyer read the same transcript differently. Most AI agent products today are priced on outcomes closer to soft than their pricing pages admit.

The dispute economy nobody priced in

Outcome-based pricing is structured like insurance: a variable bill tied to events, where the contract's real value depends on how narrowly or broadly those events are defined. Insurance disputes get settled in the definitions section of the policy, not the premium calculation, and outcome-based AI contracts are heading the same direction. The open question isn't whether vendors will keep billing for assumed resolutions. Most will, because assuming is cheaper than verifying. It's whether buyers get a real mechanism to contest an individual charge.

Today, that mechanism is thin. A buyer disputing a billed outcome is arguing against a system of record the vendor controls, using a transcript the vendor stored, applying a definition the vendor wrote. Few published contracts include an independent appeals process, a maximum dispute rate before charges are automatically waived, or an audit trail detailed enough for a buyer's finance team to reconstruct why a specific conversation was billed as resolved.

Not every AI vendor is convinced the model is healthy for the product itself. Siena AI, which sells a competing customer-service agent on a different pricing structure, has argued publicly that billing per resolution creates an incentive to close conversations quickly rather than well, the same problem that dogged per-minute call-centre billing for decades.

Three pricing models, three different places the risk lands

Laid out side by side, the risk allocation is the real difference between the three pricing eras, not the billing unit.

What a defensible outcome-based contract actually specifies

Buyers who negotiate these contracts well tend to insist on three things before they sign, not after the first disputed invoice arrives.

1. A binary, system-provable definition for every billable outcome type, with a separate and lower rate (or no charge at all) for anything that infers customer sentiment from silence.
2. A buyer-visible audit trail for every billed event: timestamp, transcript, and the specific system action the vendor treats as proof the outcome occurred.
3. A published dispute process with a real cap: a maximum dispute rate, for example, above which contested charges default in the buyer's favour, not the vendor's.

None of these are exotic asks. They're the same three things any buyer would want in a contract priced on an event they don't directly control. The difference is that outcome-based AI pricing is new enough that most contracts don't have them yet, and vendors have limited incentive to add them unprompted.

If you're the one shipping outcome-based pricing

The incentive runs the other way for a product team deciding whether to price its own AI feature this way. Outcome-based pricing is a strong pitch. It tells a buyer the vendor absorbs the risk of an AI feature that might not work, but only if the definition of an outcome can survive a customer asking to see the receipts. Before publishing a price next to the word "outcome," build the audit log that proves the outcome happened. Decide, in writing, whether silence counts as one, and if it does, at what rate compared with a confirmed result. Vendors that get this right end up with a pricing model that survives a renewal conversation. Vendors that don't spend one invoice at a time explaining why a customer who gave up got billed as a win.

Where this settles

Outcome-based pricing isn't going away. Buyers have made clear they would rather pay for results than for capacity, and vendors that can prove a result happened will keep winning deals against ones that can only point to a usage meter. What's still being negotiated, contract by contract, is who gets to define a result in the first place. The vendors publishing that definition plainly, backed by an audit trail a buyer can actually inspect, are the ones that will still be the default choice at the next renewal.