Outcome-based pricing for AI: examples and margin math

What is outcome-based pricing?

The customer pays per finished result instead of per seat or per unit of usage. If the AI does not deliver the result, there is usually no charge, so the vendor carries the cost of failed attempts.

Sierra, which prices its agents by outcome, says the model only works where the software is highly autonomous and the outcome can be cleanly attributed to it. If people still do most of the work, you are selling a tool, and tools price by seat or usage. If the result cannot be traced to the software, every bill becomes an argument. Its essay places four pricing models on those two axes.

Which companies charge per outcome?

Terms from vendor pages. Prices checked November 10, 2026.

How do I price an outcome so it covers failed attempts?

Start from the cost of one attempt, then divide by the share of attempts that succeed. The successful attempts pay for the failed ones.

Example: a support conversation runs 6 model turns, each with 4,000 input tokens and 300 output tokens, on Claude Sonnet 5.5. At $2 and $10 per million tokens, input costs $0.048 and output $0.018. Two Firecrawl searches through Locus Pro add $0.0034 at 1.7 credits each. One attempt costs $0.0694.

Long conversations break the average. The model part of a 30-turn conversation costs five times as much as a 6-turn one. Cap the cost of a single attempt, or your worst customers set your margin.

What goes wrong with outcome-based pricing?

Common protections are a monthly minimum, like Intercom’s for standalone Fin, and a usage fee for work that does not fit an outcome. Sierra describes pricing routing or greeter conversations by conversation count for that reason.

• Definitions: "resolved" has to be measurable from data both sides can see, within a stated time window.
• Attribution: if a customer also phones or emails, the AI can get credit for a result a person delivered, or the reverse.
• Weak accounts: a customer where only 10% of attempts succeed costs you ten attempts per paid outcome.
• Cash timing: you pay model and API costs at once, but bill only after the outcome is confirmed.
• Complexity: Sierra calls outcome pricing harder than seat or usage pricing to run, to contract and to account for.

How do I measure outcome costs with Locus Pro?

Locus Pro bills per call, not per outcome. Use it to run and cost each attempt, and charge the outcome fee from your own billing system.

Run the agent's paid tool and model calls through a Locus Pro Enterprise workspace. Tag each call with X-Locus-Attribution, carrying your customer ID and conversation ID, and with an agent loop ID. The ledger and customer analytics then show the cost of every attempt and every customer, successful or not.

An agent key's per-loop limit caps what one attempt can spend. With a limit of 250 credits, Locus refuses further paid calls in that loop once it reaches $0.25. Compare each customer's cost with the outcomes you billed them, and you see margin fall as soon as their resolution rate drops.

• Control AI agent tool spending
• Enterprise plan

When should I use outcome-based pricing?

Use it only where both sides can count the result without a dispute.

• Use it when the agent finishes the job without a person, the buyer already counts the result, and shared data proves it.
• Use it when your success rate is high and stable enough to price against.
• Avoid it when results take weeks to confirm or depend on people outside your product.
• Avoid it for tools that help a person work faster; price those by seat or usage.
• Consider pricing per conversation or per action as a middle step. Buyers can count those without arguing over results.