Anatomy of a governed prior-auth agent: six steps, one human gate

Prior authorization is a great test case for agent design because getting it wrong has real consequences. Most 2026 vendors optimize for "no staff intervention." Here's the architecture I'd defend instead — six stages, with the human gate as a first-class step, not a callback.

1. Intake + PHI redaction on the input path. Redaction runs before prompt assembly, not after generation. Order matters: identifiers never reach the model.
2. Retrieval of live policy. Pull chart facts and the payer's current medical-necessity criteria at request time. Cached criteria go stale; read the live policy.
3. Match + draft with enforced citations. Each drafted element is grounded to the exact criterion it satisfies. No citation → flag the gap, don't paper over it.
4. Durable human interrupt before submission. The gate halts the run and waits for an approver — it's a state the workflow can't skip, not a "review later" flag. This is the step the category quietly drops.
5. Submit + track against SLA. File to the payer, track status against the CMS seven-day decision window (Jan 2026).
6. Append-only audit log. request → retrieval → draft → citations → approver → timestamp, exportable. "Who authorized this and why" must be answerable.

The design principle: autonomy is per-action metadata, and the consequential action (submission) is gated. That's what we mean by governed AI agents at IntelliBooks Studio — more at intellibooks.ai/overview.