One action, fully traced.
A single autonomous workload — one reasoning session that holds authority, reads data, and calls tools — commits an action that touches every domain of enterprise trust at once. Here is one, followed through the evidence chain that has to hold before it's allowed, and stays reviewable after.
A customer-service AI agent drafts and sends a commitment email to a customer — promising a refund, a delivery date, and a service credit.
None of the control data here is illustrative. Every control below — its plain-language description, its required evidence, and its blocking effect — is fetched live from the published control matrices when this page loads. The step-to-control mapping is the only editorial layer; each control links to its full definition so you can check the fit. This is the proof structure the action requires (mapped). Whether a given deployment satisfies it depends on the real evidence at runtime — mapped ≠ satisfied. Recompute the corpus hashes →
The one action, as an evidence chain — select any link to jump to the controls that govern it. Counts are live from the corpus.
Loading the live control matrices…