A 10-point technical checklist for evaluating a healthcare AI agent

If you're the engineer who has to make a healthcare AI agent safe in production, marketing checklists are useless. Here's the technical version — what to actually verify.

1. Citations enforced at the output layer — a decision without a grounded source is rejected before it ships, not "usually cited."
2. Human gate as a durable interrupt — consequential actions halt and resume after sign-off, not a fire-and-forget callback.
3. PHI redaction on the input path — before the prompt is built; verify the order, not the checkbox.
4. Append-only audit log — request → retrieval → decision → citation → approver; exportable.
5. On-prem / air-gapped deploy — the governance model stays identical off-cloud.
6. Confidence-thresholded escalation — below-bar routes to needs_human with a structured reason.
7. Per-action autonomy metadata — reversible auto-executes; irreversible/compliance-bearing forces the gate.
8. Honest claims — "HIPAA-ready," and they can tell you exactly which controls map to which safeguard.
9. Real tool-use across systems — not a chat wrapper; it reads and acts via a control plane (MCP-style).
10. Inspect + override — you can see the reasoning, edit behavior, and kill a workflow.

One through five are non-negotiable. We build governed healthcare AI agents to this spec at IntelliBooks Studio — more at intellibooks.ai/overview.