AI Agent Governance for Healthcare and Life Sciences
Healthcare organizations deploying AI agents face HIPAA, HITECH, and 21 CFR Part 11 requirements that demand provable data handling controls. An agent that touches patient data, clinical notes, or billing records must operate within a verifiable boundary.
What Goes Wrong Without Controls
A regional health system deploys agents to automate clinical documentation, insurance pre-authorization, and patient scheduling. Within weeks:
- An agent summarizing clinical notes sends a prompt containing a patient's name, DOB, and diagnosis to a third-party model API — a HIPAA breach reportable to HHS
- A scheduling agent accesses the EHR API to check availability but also pulls patient records it does not need — excessive access with no boundary enforcement
- An auditor asks for evidence that PHI was redacted before every model call during Q2 — the team has no receipts, only application logs that show the call was made
DLP redaction, egress allowlists, and proof bundles prevent all three scenarios and produce the evidence to prove it.
Regulatory Mapping
| Regulation | Requirement | Claw EA Control |
|---|---|---|
| HIPAA Security Rule | Access controls, audit controls, transmission security | DLP redaction, egress allowlist, proof bundles |
| HITECH Act | Breach notification, increased penalties | Secret boundary prevents credential/PHI leakage; receipts prove containment |
| 21 CFR Part 11 | Electronic records, electronic signatures | Ed25519 signatures on every receipt and bundle satisfy electronic signature requirements |
| State privacy (CCPA, etc.) | Data minimization, access logging | Scoped tokens enforce data minimization; tamper-evident logs provide access records |
Recommended Control Stack
DLP Redaction
Strip PHI (names, DOB, MRN, diagnoses) from prompts before they reach any model provider. Receipt proves redaction occurred.
Secret Boundary
Prevent credentials and sensitive tokens from appearing in model context. Enforced at the gateway layer.
Egress Allowlist
Restrict which external APIs and endpoints agents can contact. Unauthorized destinations are blocked and logged.
Audit Log Retention
7-year tamper-evident retention with hash-linked proof chains. Meets HIPAA 6-year minimum with margin.