The line lands on the first slide and disarms the whole room: "you only pay when the agent resolves it." The CFO looks up for the first time in twenty minutes. Nobody asks anything else, because there doesn't seem to be anything else to ask. Who argues with a price that only bills when it works.
The thesis in one line: paying per outcome sounds like shared risk and is almost always transferred risk, because whoever defines what counts as "resolved" keeps the margin —and in almost every contract I have read, you are not the one writing that definition—.
AI agent pricing per outcome or license in 2026, with real numbers
This is no longer an analyst trend: it is the price list. Intercom sells Fin at $0.99 per outcome —a resolution or a completed handoff— with a public price and no platform fee. Salesforce charges $2 per resolution for Agentforce, with the caveat that if the case escalates to a human or the customer leaves angry, there is no charge. And Sierra, which hit a $15.8 billion valuation in May 2026 and works with close to 40% of the Fortune 50, also sells per outcome, but from a very different floor: its platform and annual implementation start above $200,000 before a single outcome is counted.
That last figure is what orders the debate. All three say "per outcome" and mean different things: in one case it is a public unit price with no floor, in another it is a unit price sitting on a platform that already costs you a quarter of a million before you start. Same label; the cost structure looks nothing alike.
- Pure per outcome. Public unit price, no fixed fee. Your invoice is proportional to the volume resolved. The vendor eats the cost of the failed attempts.
- Per outcome with a floor. Annual platform fee plus a negotiated, non-public price per outcome. You pay the floor whether it resolves or not.
- Per license or per seat. You pay for capacity, not usage. If the agent resolves three times as much, you don't pay more; if it resolves nothing, you don't pay less.
The honest argument in favor: someone has to carry the risk that it doesn't work
Before taking it apart, it is worth admitting that per-outcome pricing solves a real and expensive problem. For fifteen years software shipped per seat, which meant the vendor got paid the same whether the product was used or left dead in a tab. Anyone who has seen the invoice for a tool nobody opens knows exactly what I mean: it is the same mechanism by which adoption, not the model, decides AI ROI.
Charging per outcome fixes that incentive at the root. The vendor only earns when its agent works, so it has its own reasons to make it work: tuning the model, improving the knowledge base, cutting escalations. And some back it with money —Intercom goes as far as guaranteeing up to $1 million if its agent misses the resolution targets—. That is not marketing: it is real alignment of interests, and it is more than almost any license offers.
The model isn't the problem. The word is.
Whoever defines "resolved" keeps the margin
A per-outcome contract is, underneath, a contract about a definition. And "resolution" covers an enormous range, all of it defensible on paper: does it count as resolved if the customer doesn't write back within 24 hours? If they don't ask for a human? If the agent marks the case closed? If the customer says yes when asked whether it helped?
Each of those definitions produces a different invoice with exactly the same agent and exactly the same behavior. And some produce a different invoice with worse service: if "resolved" means the customer doesn't write back, a customer who gives up and walks counts as a win. If it means the agent closes the case, you are paying for its own opinion of its own work.
This is where the model splits into two different businesses. In the good one, the definition is observable and verifiable by you: the order was created, the return was processed, the appointment landed on the calendar. In the bad one, the definition is an internal state of the vendor's system that only they can measure and only they can audit. The difference between the two never shows up on the first slide; it shows up in schedule three.
And there is a second, quieter effect: a loose definition doesn't just inflate the invoice, it contaminates your metrics. If your resolution rate is calculated by the party billing on it, you have outsourced the work and the thermometer at the same time.
The four questions that decide the contract
You don't need a law firm to lock this down. You need four questions asked before signing, and the quality of the answers tells you more about the vendor than the demo does.
- What exactly counts as resolved, written with an example and a counter-example? If the answer is a marketing sentence and not a rule you can check inside your own systems, you don't have a price: you have their estimate.
- Who measures, and from what data? The healthy version is a metric you can rebuild from your side —your CRM, your helpdesk, your ERP—. If it only exists in the vendor's dashboard, you are accepting their word as a billing line.
- What happens with false positives? A case marked resolved that the customer reopens two days later should be deducted. Whether a reopen window exists —and whether it subtracts— is the acid test of whether the model was built to align or to bill.
- Where is the cap? A per-outcome price with no ceiling turns a demand spike into an invoice nobody budgeted. A monthly cap, or tiered pricing that drops as volume grows, is what stops success from getting expensive.
When outcome pricing wins and when the license does
The answer isn't that one model is better. It is that each one wins on different ground, and the ground is set by your operation, not by the vendor.
Outcome pricing wins when volume is irregular and you don't want to pay for idle capacity, when the outcome is objectively verifiable from your systems, when you are starting out and need the vendor to have skin in the game, or when the cost of it not working would land entirely on you. The license wins when volume is high and stable —past a certain point the unit price always loses to the flat fee—, when the outcome is fuzzy or hard to measure, when the agent does many small things instead of one big countable one, or when you need budget predictability more than you need cost optimization.
The arithmetic of where the crossover sits is the same as ever, and it is the one almost nobody runs before signing: expected volume × unit price against the fixed fee, with the full cost of both options —integration, maintenance, human hours of supervision— inside the calculation. It is worked through in the guide on calculating the ROI of an AI automation, and the underlying pattern is identical to the one that decides the bill on automation platforms: what matters isn't the entry price, it is which unit they actually bill you on.
What to ask for before you sign
Three things, none of them exotic. First, the definition of resolution written into the contract, not on the website —websites change without version control—. Second, a parallel measurement period: two or three months where you count outcomes your way and they count theirs, and the two get compared before the invoice depends on it. Third, the reopen clause and the cap. A vendor confident in its agent takes all three without blinking; one that gets nervous about the second is telling you something it will never say out loud.
And a warning underneath all of it: no pricing model fixes an agent that doesn't fit your operation. An agent that resolves 30% of cases because it can't reach the data it needs will keep resolving 30% whether you pay per license or per outcome; the only thing that changes is who eats the loss. So the first job isn't negotiating the price of a resolution: it is building the agent so it resolves, with the systems connected and the real work in front of it. That is what we ship as AI employees —we leave the agent running inside your operation—, and only once the work comes out right does it make sense to argue about how it gets billed.
Per-outcome pricing is a good idea with fine print that decides the entire deal. Don't sign it off the first slide. Read the schedule where somebody, at some point, wrote down what "resolved" means —and check whether that somebody was you—.