Outcome-based pricing for AI agents: why failed tasks should cost zero

Every pricing model is an incentive machine wearing a rate card. For AI agents, the industry has cycled through three machines in about four years, and each one teaches the platform a different behavior. The third only works if you can verify outcomes, which is why almost nobody runs it honestly yet.

Three machines

Per-seat pricing carried over from SaaS, and it fits agents badly. A seat measures a human's presence, but agents are not present; they act. Charging per seat while agents do unbounded work either overprices light users or forces usage caps that make the product feel broken. More importantly, the platform's revenue is disconnected from whether any agent action worked.

Per-task pricing gets closer: pay for what runs. But it quietly teaches the platform a bad lesson. If every completed task is billable and the platform decides what counts as completed, then the cheapest revenue lever is a generous definition of done. The bar drifts. Customers end up paying for attempts dressed as outcomes, which is how the category earned its reputation for marking work done that never shipped.

Per-outcome pricing is the one everyone gestures at: pay only when the work verifiably succeeded. Stated as a slogan, it is easy. Operated honestly, it has a hard dependency most platforms lack: an outcome test the seller cannot game.

Verification is the precondition

Outcome pricing without independent verification is just per-task pricing with better marketing, because the same party still grades the outcome. The pieces that make it real are mechanical. The definition of done must be fixed before the run: a deploy means this URL answers, a plan means output parses against this versioned contract. The check must run outside the agent's reach, in a process the agent cannot write to. And the verdict must be recorded as an artifact the customer can audit later.

With those pieces in place, billing can bind to the verdict. On Peragam the binding is atomic: the credit charge and the verification result settle in one step. A passed check charges the task's credits. A failed check charges zero, not as a refund with a form attached, but because no charge ever existed. There is no window where a failed task has billed you and a make-good is pending.

Why zero is the right number

The obvious objection: the failed run still cost the platform real money in model tokens and sandbox time. Why should the customer pay nothing rather than cost?

Because of where that puts the incentive. When failure costs the customer zero, the platform eats every failed run, which makes failure a line item on the platform's books instead of the customer's. A platform that absorbs its failures is forced to invest in fewer failures: better verifiers, better task design, better models. A platform that passes failure costs through can stay comfortable with a mediocre success rate forever, because someone else is funding it.

Zero also does something subtler: it makes the platform's completion claims believable at no auditing cost to you. We charge on passed checks only, so every optimistic checkmark would be a checkmark we lose money on. You do not have to trust the green pixels. You can trust that we cannot afford to fake them.

The arithmetic, concretely

Numbers make the contrast plain. Take a month with 40 attempted tasks and a 90% pass rate. Under attempt-based billing you pay for 40. Under verified-outcome billing you pay for 36, and the four failures cost you nothing while costing the platform its own run costs. Now drop the pass rate to 60%. Attempt billing charges you the same 40 while delivering 24 outcomes; your effective price per outcome quietly rose 67%. Verified billing charges 24, and the platform just inherited sixteen failures worth of motivation to fix itself. Same agents, same month. Only the incentive moved.

This is also why cost transparency belongs on the receipt. Every verified task on Peragam records the run's dollar and token cost next to the credits charged, so you can see the margin you are paying and we cannot claim outcome pricing while hiding attempt economics. How that receipt is produced, stage by stage, is on the product page.

Where this goes

As verifiers mature, we expect outcome pricing to become table stakes, the way uptime SLAs did for infrastructure. The platforms that get there first will not be the ones with the best slogans but the ones whose verification actually holds up under audit. Ask any vendor one question before believing an outcome-pricing pitch: show me a failed task, and show me what it charged. The honest answer is a receipt that reads zero.