Key Takeaways

• Intercom prices its Fin AI Agent at $0.99 per outcome and counts an outcome when the customer confirms resolution, stops asking after Fin responds, or Fin completes a workflow including a handoff.

• Fin bills once per conversation, so one price covers every model call the agent made inside that thread.

• Attribution, not measurement, is the hard part: a human reply that settles the issue first usually ends the billable event.

• Tokens bill on every attempt while revenue lands only on successes, so the operating unit is cost per attempt divided by the success rate.

• At $0.12 of tokens per attempt, a 30% resolution rate spends $0.40 to earn one $0.99 outcome, against $0.15 at 80%.

What counts as an outcome in AI pricing?

An outcome is whatever the contract says it is, and writing that sentence down is the pricing decision. Our guide to outcome-based pricing covers the general model. The AI version runs harder because an agent produces a probabilistic result, and somebody has to decide when it counts.

Intercom's pricing page shows a working definition. Fin costs $0.99 per outcome, and Intercom counts one when any of these happen:

• The customer confirms their issue is resolved.

• The customer doesn't ask for more help after Fin responds.

• Fin completes a workflow, including a handoff to a human.

The second condition does the real work, since silence stands in for a confirmation the customer never gives. Intercom charges once per conversation regardless of how many questions Fin answers, and applies a minimum monthly commitment, 50 outcomes in its example.

Attribution follows the definition. Every billable event needs a record of which actor produced it and whether a human touched the thread first. That's the same event trail AI agent pricing runs on, and without it the invoice becomes an argument.

How does outcome pricing affect AI margins?

Outcome pricing moves the volume risk onto the vendor, because token-based pricing charges on inputs and outcome pricing pays on results. Failed attempts still cost tokens. The success rate, not the price, sets the margin.

The table runs a $0.99 outcome against an assumed $0.12 of tokens per attempt, varying only how often the agent succeeds. Intercom publishes a data point in that range: Ben Peak, Director of Technical Support at Robin, reports a 50% resolution rate with Fin.

The levers that move this margin aren't on the price tag. Raising the success rate cuts wasted attempts, and capping spend per attempt floors the worst case when an agent loops. Minimum commitments do similar work, stopping a low-resolution account from burning inference all month and billing almost nothing.

Related terms

These pages cover the neighbouring pricing mechanics.

• Outcome-based pricing is the general model, including the non-AI versions.

• AI agent pricing compares per-run, per-task, and per-outcome structures for agents.

• Token-based pricing is the input-side model that outcome pricing replaces on the invoice.

• AI margin management covers tracking inference cost against revenue per customer.

• Credit-based pricing is the prepaid structure teams often pair with outcome charges.

FAQ

How do you verify an AI outcome?

You verify it with a recorded signal the customer or the workflow produces, not the agent's own report. Confirmations, no follow-up inside a set window, and a completed workflow step all serve. Each needs a timestamp and an actor so the charge survives a dispute.

Is outcome-based pricing better than token-based pricing for AI agents?

It's better when the outcome is countable and the agent succeeds often enough to cover its failures. Token pricing protects the vendor's margin, outcome pricing protects the buyer's. Agents with unstable success rates give up more margin than the pricing wins back.

Who gets credit when a human finishes the conversation?

Most contracts stop crediting the agent once a human resolves the issue, which is why handoffs need their own rule. Intercom counts a completed workflow as an outcome even when it ends in a handoff, putting the boundary at the workflow.