Free resource · Anjaneyullu Tamma

AI Governance Readiness Checklist

A practical, board-ready way to pressure-test whether your AI is safe to put into production.

Print / Save as PDF

Use this checklist before — and during — any project that puts AI in front of real data or real decisions. It is deliberately tool-agnostic: it asks whether a control exists, not which product provides it. If you can’t confidently tick an item, that gap is where your next incident is hiding.

1. Strategy & ownership

• Use cases are prioritized by value and risk, not novelty.
• Each use case has a named owner accountable for outcomes.
• Build-vs-buy and model choices are deliberate and revisited.
• There is a target architecture, not a pile of disconnected pilots.

2. Data protection & privacy

• Every category of personal data is mapped to a controller/processor role.
• Erasure reaches every store — including vector DBs / embeddings and backups.
• Sensitive fields are redacted at the gateway before reaching any model.
• Crown-jewel identifiers are encrypted/tokenized at the field level.
• Cross-border data transfers are registered and assessed.

3. Security & tenant isolation

• Identity comes from verified claims, never caller-supplied input.
• Every request is scoped from verified identity (not the request body).
• Data access is tenant-scoped, with row-level security as the final net.
• Agents run least-privilege; tools and actions are allow-listed.
• Model inputs/outputs are treated as untrusted (prompt-injection aware).

4. Trust & data integrity

• AI output is treated as a proposal, not a fact.
• A deterministic validation layer runs before any write to a system of record.
• Data provenance — not model confidence — is the primary trust signal.
• Data is classified by cost of failure to tier automation vs. review.
• Every decision is recorded in an immutable audit trail.

5. Human oversight & reversibility

• Consequential or irreversible actions sit behind an explicit human gate.
• People affected by automated decisions can contest them.
• The system can pause, roll back, and recover a prior state.
• The system escalates its own uncertainty instead of guessing confidently.

6. Evaluation & monitoring

• An evaluation suite defines and continuously measures expected behavior.
• Behavioral telemetry surfaces drift before silent corruption spreads.
• Human corrections feed back into thresholds and validation rules.

7. Compliance & documentation

• Records of processing, DPAs, and a sub-processor list exist and are current.
• A retention schedule names statutory minimums; deletion respects them.
• Disclosure that AI is involved is provided where required.
• Runbooks exist for breach notification and data-subject requests.

Most teams find the AI-specific controls (sections 3–6) are the strongest part of their system — and the gaps live in the unglamorous lifecycle work (sections 2 and 7). Get the boring parts right and the AI parts turn out to be the easy half.