AI governance checklist

Inventory

• Every internally built AI system is listed with an owner.
• Vendor and embedded AI features in existing SaaS are listed.
• Shadow AI usage has been surveyed via egress or SaaS admin data.
• The inventory has a named maintainer and a review cadence.
• Agentic systems and autonomous tool-calling agents are inventoried separately from single-turn LLM features.

Classification

• Each system has a documented role: provider or deployer, matching the EU AI Act's definitions.
• Each system has a risk tier with recorded reasoning; Annex III high-risk categories (biometrics, employment, education, critical infrastructure, essential services) are checked explicitly.
• Personal-data involvement is recorded per system.
• Systems affecting individuals' rights are flagged for legal review.
• Systems in scope for AI Act transparency duties (chatbots, synthetic content, deepfake generation) are flagged for labelling obligations.

Policy and process

• Acceptable-use policy published and communicated.
• Intake gate exists for new AI use cases.
• Approval authority and escalation path are defined.
• Human oversight requirements are specified per system.
• A regulatory calendar tracks applicable deadlines and is reviewed each quarter, since AI Act implementation dates have moved before and will likely move again.

Evidence

• Prompt, decision and tool-call logging with defined retention.
• Change control covers model and prompt changes.
• Vendor terms reviewed for training use and data residency.
• Incident path defined for AI-specific failures.
• Mapping exists to NIST AI RMF and, where third-party assurance is required, to ISO/IEC 42001's Annex A control set.