-
Start with root cause and a defensible scope
Begin by establishing the root cause of the finding and a scope you can defend. That means a complete inventory and named owners for every in-scope AI system and agent.
-
Break the MRA into workstreams
Break the finding into workstreams so that each gap has a named owner and can be tracked to closure.
-
Map each finding to a control and an evidence artifact
Design one control per finding, and define the evidence artifact that will show the control is working.
-
Sequence work against the time remaining
Sequence work by dependency against the committed dates. Controls need to run long enough to generate testable samples, so schedule them to start early enough to build an operating history before independent testing.
From finding to closure evidence
The work moves through four stages, each with a distinct outcome.
| Stage | Outcome |
|---|---|
| Scope | Complete inventory and named owners for every in-scope AI system and agent. |
| Design | One control per finding, with a defined evidence artifact. |
| Operate | Controls run long enough to generate testable samples. |
| Validate | Independent testing confirms controls work, not just exist. |
Test whether controls are operating before examiners do
Use these criteria as a closure readiness check:
- Every finding has a mapped control, a named owner, and a defined evidence artifact.
- The inventory reconciles against independent sources, such as cloud accounts and API gateway records.
- A sampled AI action can be traced from request to policy decision to outcome.
- Policy exceptions are logged, approved, and time-bound.
- Evidence covers a defined operating period rather than a single point in time.
- Someone independent of the control owner has tested the controls and documented the results.
Where runtime controls produce testable evidence
A written policy states intent. A runtime control evaluates each AI action against that policy as it happens and records the outcome. This distinction matters for closure, because the record is generated by the system and does not depend on someone remembering to document compliance. Useful runtime mechanisms include:
- Policy enforcement on agent actions
- Distinct agent identities with least-privilege permissions
- Approval workflows for sensitive tool calls
- Audit logs of requests, decisions, and outcomes
Together these give a reviewer something concrete to test. They can pick an agent action, trace it to the identity that made it, the permission that allowed it, the policy that was applied, and the result. They can also check that blocked actions and approved exceptions appear in the record.