AI Integration Risk Register

Use this before giving an AI workflow access to email, calendar, CRM, documents, chat, or other business tools.

Integration Inventory

Risk Assessment

Permission Rules

• Use a provider-supported workload identity where available; never share a personal login to bypass a provider limitation.
• Use least-privilege OAuth scopes or API keys.
• Prefer read-only access first.
• Enable write access one action at a time.
• Store credentials outside prompts, model context, ordinary logs, and generated output.
• Rotate credentials according to provider guidance and the organisation’s risk-based policy.
• Revoke unused integrations.

Human Gate

Choose and test a gate for each write action. Irreversible, high-consequence, or externally visible actions require explicit review before the effect:

• Approve before act.
• Delay with a named monitor and a tested cancellation mechanism, only where the action is reversible.
• Automatic action within a documented low-consequence policy boundary, with exception review and a tested stop control.

Required Logs

• Timestamp.
• Agent or workflow name.
• Trigger.
• Pseudonymous input record ID or another approved stable reference.
• Tool called.
• Minimum approved arguments or stable references; never credentials or unnecessary personal data.
• Result code and reconciliation status; store payloads only where separately justified and protected.
• Human approver if applicable.
• Error or warning.

Stop Conditions

Pause the workflow immediately if any of these occur:

• Wrong recipient, contact, customer, or account.
• Duplicate business action.
• Missing approval.
• Sensitive data sent to an unapproved destination.
• Authentication or permission error.
• Unexpected cost or rate-limit spike.

Launch Decision

• Approved by:
• Date:
• First review date:
• Rollback path:
• Credential revocation path:
• Kill-switch owner and test date: