The 12 fields a defensible AI audit trail actually needs (healthcare edition)

A reference spec, not a pitch: what a defensible AI-decision audit-log entry actually needs to contain, and how long you have to keep it.

The 12-field baseline (2026 guidance):

1. NTP-synced timestamp, UTC
2. Unique decision ID
3. Authenticated human user identity
4. AI system identity + version
5. Model identity + version
6. Inputs, with source attribution
7. The specific policy/rule invoked
8. Human-readable reasoning trace
9. The output
10. The downstream action taken
11. Human review/approval, when applicable
12. Tamper-evident integrity proof

Retention isn't one number: 6 months (EU AI Act) / 1 year (SOC 2) / 6 years (HIPAA-covered workloads). Design for the longest applicable requirement, not the shortest.

Design note: fields 6–8 (inputs with source attribution, policy invoked, reasoning trace) are what actually get scrutinized under challenge — they're the difference between "the model said X" and "the model said X because of Y, sourced from Z." If your schema doesn't capture that chain, the other 9 fields don't save you.

Why this matters now: Colorado, Utah, Texas, and California have active AI enforcement in 2026 — a state-by-state inventory of deployed AI tools is a realistic near-term ask, not a hypothetical. This is the audit-trail schema behind governed AI agents at IntelliBooks Studio — more at intellibooks.ai/overview.