Agent governance Proctoring
Agent Privileging · Component 2

Proctoring

New clinicians are proctored: they practice under observation until the record supports independence. An agent's proctoring period is the same instrument — a defined window in which every action is human-reviewed, ending in an explicit graduation decision rather than a quiet lapse into trust.

Duration and dosage

Fix the window in advance — 30 days is a defensible default; high-risk scopes warrant 60–90 — and define the review dosage: what fraction of actions get human review during proctoring (100% for anything that leaves the organization; sampling is defensible for internal drafts). An open-ended proctoring period is not caution, it is an ungoverned deployment wearing caution's badge.

Graduation criteria, written before day one

Privileges are granted when the record shows: zero out-of-scope actions; an error rate at or below the level the granting committee accepted in writing; and no unresolved incident. The decision is made by a named human, recorded with rationale. If the criteria cannot be stated before deployment, the organization is not ready to deploy — that discovery is the proctoring design doing its job.

What the record must contain

Dated action log with scope compliance, the review sample and its findings, every incident however minor, and the graduation decision with grantor and rationale. This record is the direct analogue of the proctored case list in medical staff credentialing — and like it, its value is that it existed before anyone asked for it.

Part of the complete healthcare AI governance guide.