Data Residency, HIPAA, and MCP: What Compliance Officers Need to Know Before Connecting AI Agents to a CRM

Letting Claude read PHI sounds scary until you understand the architecture.

Insurance CRMs hold PHI. Connecting AI agents to PHI without thinking through the compliance architecture is reckless. Done correctly, MCP is no riskier than any other API integration.

The Three Questions Compliance Always Asks

1. Where Is the Data?

MCP doesn't move data anywhere new. The agent calls the CRM API. The response goes to the LLM provider. From there, it depends on the LLM contract.

2. Does the LLM Provider Sign a BAA?

Anthropic offers a BAA for Claude Enterprise. OpenAI offers BAAs for ChatGPT Enterprise and API customers. If you require HIPAA, sign a BAA with whichever provider your team uses.

3. Are Logs Auditable?

unLocked CRM logs every API and MCP call with X-Request-Id, timestamp, key fingerprint, and resource. Combined with your LLM provider's logs, you have full call traceability for any audit.

Practical Controls

  • Use scoped API keys per agent type
  • Disable PHI fields for agents that don't need them
  • Mask SSN, DOB, and Member ID in MCP responses for analyst agents
  • Require approval for any write operation (Claude can draft; CSR sends)
  • Monthly access review of all issued keys

What unLocked Does Out of the Box

Field-level masking, scoped keys, audit logs, BAA-eligible architecture. Compliance review of MCP integration takes hours, not weeks.