AI Governance Checklist for Secure, Reviewable AI
A practical checklist for teams moving one AI workflow from idea to a controlled pilot or production decision. It is a planning aid, not a certification or a claim about any customer implementation.
Use this checklist before scaling an AI workflow
Work through the controls below with the business owner, delivery team, security, legal or compliance stakeholders as relevant. The aim is to identify evidence gaps early—not to add process for its own sake.
1. Name an accountable owner
Document who owns the business outcome, who can approve a change, and who can stop the workflow if it behaves unexpectedly.
2. Define the decision and risk tier
State what the AI may recommend, draft, classify, or automate. Identify whether an incorrect output could affect money, safety, rights, regulated records, or customer commitments.
3. Classify data and access
List data sources, sensitivity, retention needs, residency constraints, and who or what may access them. Apply least-privilege access rather than broad shared credentials.
4. Select and evaluate the model
Record the intended model, hosting location, known limitations, evaluation set, acceptance threshold, and a fallback if quality is below the agreed standard.
5. Protect prompts and instructions
Separate trusted system instructions from untrusted user or document content. Test for prompt injection, data leakage, unsafe tool calls, and instruction conflicts.
6. Set human review boundaries
Specify which outputs require review, who reviews them, what evidence they need, and when the workflow must escalate instead of continuing automatically.
7. Constrain agents and tools
Give agents only the tools and permissions required for the defined job. Require confirmation for consequential actions such as external messages, record changes, payments, or deletes.
8. Keep audit evidence
Capture the request, relevant version identifiers, tool actions, reviewer decision, and final outcome in a way that supports investigation without retaining unnecessary sensitive data.
9. Monitor quality and cost
Track outcome quality, error patterns, latency, usage, cost, drift, and failed safety checks. Decide who receives alerts and how often the controls are reviewed.
10. Prepare incident response
Define a pause or rollback path, incident owner, customer communication path where needed, and a process for fixing and retesting the workflow before restart.
Evidence to gather for a first assessment
- One representative workflow and its current manual steps
- Sample inputs with sensitive data removed or protected
- Known policy, security, regulatory, and residency requirements
- A measurable success definition and an acceptable failure boundary
- The systems, teams, and approvals needed for a bounded pilot
Where evidence is incomplete, that is an assessment finding—not a reason to promise a solution prematurely.
LLM API governance: minimum controls
When a workflow calls an external or internal language-model API, document the model provider, purpose, approved data classes, authentication method, rate and spend limits, retention terms, and fallback behaviour.
Gateway and identity
Use named service identities, scoped keys or workload identity, server-side secret handling, and a controlled gateway where appropriate. Do not embed broad provider keys in user-facing applications.
Input and output controls
Filter or transform sensitive inputs where required, separate trusted instructions from untrusted content, validate structured outputs, and define human review boundaries.
Evidence and change control
Record the model/version, prompt or policy version, tool actions, evaluation evidence, and approval for material changes.
These controls help a team assess API use responsibly; they are not a substitute for legal, security, or compliance advice.