A portable AI Trust Passport: cross-company assurance without re-auditing everyone

Here is a AI governance problem that deserves more rigour than it usually gets.

The mechanism

Governance defines who is accountable, which policies bind the system, and which decisions require human sign-off. The operational backbone: classify the system against EU AI Act risk tiers (Art. 6 and Annex III for high-risk), then discharge the high-risk obligations — risk management (Art. 9), data governance (Art. 10), technical documentation (Art. 11 / Annex IV), record-keeping and logging (Art. 12), and human oversight (Art. 14). Layer the NIST AI RMF functions (Govern, Map, Measure, Manage) for process, and operate the whole programme inside an ISO/IEC 42001 AI management system (risk treatment under clause 6.1, operational controls under clause 8).

Why it compounds

As deployments move from assistive (the model suggests) to agentic (the model acts), errors stop being isolated outputs and become state changes — money moved, records altered, messages sent — that propagate downstream. A small per-action failure rate, multiplied across thousands of autonomous steps, becomes material operational, legal and financial exposure. The differentiator is not adoption speed but the ability to bound, observe and reverse what the system does.

Controls that apply

• A live AI system registry keyed to risk tier, owner and intended purpose (EU AI Act Art. 6).
• Approval gates and explicit decision rights for high-impact use cases (Art. 14 human oversight).
• Immutable, queryable logs of model and tool actions (Art. 12 record-keeping).
• Technical documentation maintained as you build (Art. 11 / Annex IV), not reconstructed before an audit.
• Policy expressed as automated checks in the deployment pipeline, versioned alongside the model.

What to test

• Instrument one high-impact use case end-to-end with per-action logging.
• Define quantitative thresholds (error rate, autonomy level, blast radius) and an automatic stop.
• Run a bounded pilot; compare incident rate, review latency and residual risk against the baseline.
• Capture near-misses as structured signals — they are the cheapest telemetry you have.