Turning an AI risk appetite into enforceable limits

Appetite, tolerance, and limit are different things

Appetite is the qualitative statement approved by the board or risk committee. Tolerance expresses that statement as measurable indicators with warning and hard thresholds. A limit is a specific technical or procedural boundary set for each threshold.

Keeping the three terms separate makes it clear which parts of the appetite statement have been made measurable and which remain qualitative.

From appetite to evidence

A six-step translation method

1. Break each qualitative statement into a measurable indicator.
2. Set tolerance thresholds, including warning and hard thresholds.
3. Map each threshold to a preventive, detective, or responsive control.
4. Decide whether each limit is enforced at runtime or through process.
5. Define monitoring and exception handling.
6. Define the audit evidence that shows deployed systems stayed within approved limits.

Limit categories for AI systems and agents

Each threshold needs a specific limit, which can be technical or procedural. Every limit is then mapped to a preventive, detective, or responsive control with a named owner. The same approach applies to AI systems and to agents.

Runtime enforcement versus process enforcement

The enforcement point is chosen per limit: a runtime control or a process control. Record which limits are enforced at runtime and which rely on process controls, so that reviewers can see how each approved limit is actually upheld.

Monitoring, exceptions, and audit evidence

Warnings and breaches should be detected, escalated, and recorded. Exceptions should be approved, time-limited, and reviewed. Evidence takes the form of logs, approvals, and reviews showing the limit held, which lets you demonstrate that deployed systems and agents stayed within approved limits.