Quick answerA discretionary goodwill-credit budget is a capped amount an AI agent can grant per conversation, per customer per period, and company-wide per period, without escalating to a human. Set three limits, not one: a per-incident cap small enough that a single bad judgment call is cheap, a per-customer cap over a rolling window to stop repeat requesters from draining the budget one small credit at a time, and a company-wide monthly ceiling with an alert well before it is reached. Review the caps monthly against actual usage, not once at launch and never again.
Why this is easy to get wrong
Teams tend to reach for a policy statement or a prompt instruction as the fix, when the real gap is usually structural: missing checks in the data or decision path, not a lack of good intentions in the system prompt. Treating this as a one-time launch decision instead of an ongoing operating discipline is the most common way it quietly regresses months after the original design review.
Related reading: our guide on how to test an AI agent before it issues refunds autonomously covers the adjacent infrastructure and process question in more depth, and it is worth reviewing before you finalize your own approach here.
A practical framework
Start with the smallest, cheapest control that closes the most likely failure mode, then add layers only where the data shows you actually need them. A control that looks thorough on a whiteboard but never gets exercised in production is worse than a simpler one your team actually reviews on a set cadence. Write the review cadence down and put an owner's name on it, since a control with no named owner tends to drift out of date within a quarter.
For teams that have already solved a related but distinct problem, see how much to hold in reserve for AI agent error costs company-wide, which shares some of the same underlying design pattern even though the two situations are not interchangeable.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
Where this connects to the rest of your AI agent program
This is not a standalone decision. It intersects with how you already handle keeping the evidence trail for AI agent decisions, and the two should be reviewed together rather than owned by completely separate teams with no shared visibility into each other's changes.
FAQ
How is this different from testing whether the agent can issue refunds at all?
Pre-launch testing and validation confirms the agent can execute a refund or payment action correctly and safely before it goes live. This budget question comes after that: once the capability is proven safe, how much can the agent spend on its own judgment calls, and how is that reviewed over time. One is a launch gate, the other is an ongoing operating parameter.
How is this different from a company-wide error-cost reserve?
The error-cost reserve is a finance function sizing how much capital the whole company sets aside to absorb AI agent mistakes across every category of error. The goodwill-credit budget is a narrower, day-to-day operating limit specifically for discretionary courtesy spend, reviewed by whoever owns support economics, not by finance's reserve-planning process.
What should trigger a mid-cycle change to the budget?
A sustained increase in the rate at which the agent hits its per-incident cap and has to escalate, or a sustained increase in total discretionary spend as a share of revenue, are both signals to revisit the caps rather than waiting for the next scheduled review.