Practical software guides
An AI registry should identify every enabled workflow and the systems it can reach. An entry describing “helpful assistant” is too vague to review or revoke.
| Field | Example |
|---|---|
| Workflow/version | Policy draft assistant v1 |
| Business owner | Operations role and named backup in the actual deployment |
| Sources | Approved policy set; version and access scope |
| Actions | Draft only; no send or write credential |
| Provider/retention | Reviewed terms and configured retention |
| Stop procedure | Revoke credential; cancel queued work; confirm denied call |
| Change review | New source/action/model triggers review and repeat evaluation |
A case record worth checking
A fictional staff request concerns a cancellation policy. The assistant links policy version 4, drafts supported wording and flags that an exception is not covered. The case records the source, draft, human correction and release decision. If the policy is superseded, the old draft must not be presented as current.
A searchable knowledge base or deterministic decision tree can be better when questions and rules are stable. AI is an option for variable wording, not a required next step. Test unsupported questions, contradictory sources and text that attempts to grant itself tool permissions.
Enforce the entry
Use role-scoped retrieval and least-privilege credentials. A disabled send permission must remain disabled outside the prompt. Revoke the credential and verify the next request fails; include queued jobs and cached access in the review. Record the result, not merely a checked “kill switch” box.
The permission budget defines action approval; the drafting tests examines source accuracy. NIST’s voluntary framework supplies risk-management context without certifying this template.