PHI Protection for AI Workflows

How we keep PHI out of external models and aligned with HIPAA

The PHI Risk in AI Tools

What happens when staff paste PHI into external AI models

Common Scenario

Common Scenario

A nurse copies a patient note into ChatGPT to generate a discharge summary: "Patient: John Smith, DOB 3/15/1967, MRN 445521. Admitted 12/1/24 for acute chest pain. EKG shows ST elevation..."

  • What Just Happened: PHI sent to OpenAI servers (no BAA).
  • Patient name, DOB, MRN exposed.
  • Clinical data in third-party training corpus.
  • No encryption controls.
  • Zero audit trail.
  • HIPAA violation.

Regulator Lens

Regulator Lens

OCR (Office for Civil Rights): No BAA with AI vendor = impermissible disclosure. Lack of safeguards = failure to implement HIPAA Security Rule. No risk assessment = administrative noncompliance.

  • Potential Consequences: $100 - $50,000 per violation (can be per record).
  • Mandatory corrective action plan.
  • Multi-year monitoring period.
  • Breach notification requirements.
  • Reputational damage and media scrutiny.

Our PHI Protection Approach

Detect → Cleanse → Send → Rehydrate

1

Detect

Automatic scanning for all 18 HIPAA identifiers in real-time as users type or upload content. Technical: Pattern matching, NER models, contextual analysis.

2

Cleanse

Replace PHI with randomized placeholders while preserving clinical context and meaning. Technical: Token replacement, semantic preservation, reversible mapping.

3

Send

De-identified content sent to AI model (ChatGPT, Claude, etc.) with zero PHI exposure. Technical: Encrypted transport, no PHI in logs, no training corpus exposure.

4

Rehydrate

AI response returned to user with original PHI restored in secure environment. Technical: Local rehydration, no PHI stored, audit trail maintained.

Technical Controls

De-identification

All 18 HIPAA identifiers detected: names, dates, addresses, MRNs, SSNs, phone numbers, emails, account numbers, biometric identifiers, device IDs, and any unique identifying information.

Placeholder Logic

Randomized, non-reversible tokens that maintain clinical context and preserve semantic meaning. Consistent within session, no PHI in external systems.

Transport & Storage Security

End-to-end encryption (TLS 1.3), zero PHI in vendor logs, no training corpus exposure, data residency controls, and complete audit trail.

Organizational Controls

Technology is only part of the solution — governance requires people and process

Written Policies

Written Policies

Acceptable use policy for AI tools. PHI handling standards. Approved tool library. Role-based access controls. Incident response procedures.

Training & Awareness

Training & Awareness

Staff onboarding on AI governance. PHI protection best practices. How to use governed AI workspace. Reporting shadow AI discoveries. Ongoing compliance education.

Governance Structure

Governance Structure

AI governance committee. Executive sponsor (typically CIO/CISO). Compliance liaison. IT/Security enforcement. Regular review cadence.

Monitoring & Enforcement

Monitoring & Enforcement

Usage analytics dashboard. Policy violation alerts. Audit log review. Quarterly compliance reports. Continuous improvement feedback.

Compliance Alignment

HIPAA

PHI protection, BAAs, safeguards

SOC 2 Type II

Security, availability, confidentiality

HITRUST

Healthcare information security framework

State Privacy Laws

CCPA, state-specific regulations

For Your Security & Compliance Team

We understand your team needs detailed documentation. Available on request: Architecture diagrams & data flow maps, Security controls matrix, PHI de-identification methodology, Penetration testing results, SOC 2 Type II report, Business Associate Agreement (BAA), Incident response procedures, Vendor risk assessment questionnaire, HIPAA alignment documentation, Data processing addendum (DPA), Audit log specifications, Encryption standards & key management.

During the 6-Week Governance Pilot, we schedule dedicated sessions with your security and compliance teams to review technical controls, answer questions, and provide all documentation needed for internal approvals.

We don't ask you to "trust us" — we show you exactly how PHI protection works and provide evidence that meets your standards.

Frequently Asked Questions

Does PHI protection work with all AI models?

Yes. Our approach is model-agnostic — it works with ChatGPT, Claude, Gemini, or any AI tool your staff wants to use. The de-identification happens before content reaches the external model.

What about false positives or missed PHI?

Our system uses multiple detection layers (pattern matching, NER models, contextual analysis) to minimize false positives and negatives. We also provide user feedback mechanisms and ongoing tuning based on your specific workflows.

Can we get a BAA with the AI vendors directly?

Some vendors (like OpenAI Enterprise) offer BAAs, but most don't. Even with a BAA, you still need de-identification as a safeguard. Our approach gives you defense-in-depth: PHI never reaches the vendor in the first place.

How do you handle edge cases like rare diseases or unique identifiers?

Our system learns from your environment during the pilot. We tune detection rules for your specific context (e.g., department names, facility identifiers, rare conditions) to ensure comprehensive coverage.

What if a user tries to bypass the system?

Technical controls (enforced at the network/application layer) prevent circumvention. Plus, governance training and monitoring create accountability. We also provide usage analytics to detect workarounds.

Ready to Protect PHI in AI Workflows?

Schedule a security review or start with a Shadow AI Risk Check.