HIPAA Compliance Audit for AI Deployments — From Zero to Audit-Ready

You: We're deploying an AI agent for patient appointment
scheduling and billing inquiries at a 50-physician
practice. Generate a HIPAA compliance framework.

## HIPAA Compliance Framework — AI Agent Deployment
## 50-Physician Practice | Scheduling + Billing

### 1. PRE-DEPLOYMENT GATE

☐ Business Associate Agreement (BAA)
  - Required between practice and AI vendor
  - Must specifically address: AI model training
    restrictions, data retention, breach notification
  - Template clause: "Vendor shall not use PHI for
    model training, fine-tuning, or improvement of
    AI systems without explicit written authorization"

☐ PHI Data Flow Mapping
  Patient → Scheduling system → AI Agent → Response
  ├── PHI accessed: name, DOB, insurance, appointment
  │   history, provider preferences
  ├── PHI NOT needed: diagnoses, medications, lab results
  │   → BLOCK these from AI agent access
  └── Minimum Necessary standard: Agent receives ONLY
      scheduling-relevant data

☐ Risk Assessment (45 CFR 164.308(a)(1))
  - Document ALL PHI touchpoints
  - Classify by risk level (see matrix below)

### 2. TECHNICAL SAFEGUARDS (45 CFR 164.312)

ACCESS CONTROLS:
☐ Unique service account ID for AI agent
☐ Role-based permissions (scheduling-only, no clinical)
☐ 15-minute session auto-logoff
☐ Emergency access procedure documented
☐ Multi-factor authentication for admin access
☐ IP allowlisting for API connections

AUDIT CONTROLS:
☐ Log every PHI access: timestamp, user/agent, action, data fields
☐ 6-year log retention (HIPAA minimum)
☐ Anomaly detection on access patterns
☐ AI decision audit trail (what was asked, what was answered)
☐ Monthly audit review procedure

TRANSMISSION SECURITY:
☐ TLS 1.3 minimum for all connections
☐ End-to-end encryption for patient communications
☐ Certificate pinning for API connections
☐ NO PHI in URLs, query strings, or log messages
☐ At-rest encryption for any cached data (AES-256)

### 3. AI-SPECIFIC RISK MATRIX

| Risk | Likelihood | Impact | Mitigation |
|------|-----------|--------|------------|
| Prompt injection → PHI leak | Medium | Critical | Input sanitization, output filtering, sandboxing |
| Model training on PHI | Low (if BAA) | Critical | BAA prohibition, audit vendor compliance |
| Hallucinated medical info | High | Critical | Restrict to scheduling/billing, no clinical responses |
| Patient impersonation | Medium | High | Identity verification before PHI access |
| Shadow AI with PHI | Medium | High | Approved tool registry, DLP rules, staff training |
| Conversation data retention | High | Medium | Auto-purge after session, no persistent storage |

### 4. BREACH RESPONSE TIMELINE

0-1 hours: CONTAIN
☐ Disable AI agent immediately
☐ Preserve all logs (do not modify/delete)
☐ Notify Privacy Officer and IT Security

1-24 hours: ASSESS
☐ Determine scope of PHI exposure
☐ Identify affected individuals
☐ Determine if breach is "unsecured PHI"

24-48 hours: DOCUMENT
☐ Root cause analysis
☐ Number of individuals affected
☐ Types of PHI involved

Within 60 days: NOTIFY
☐ HHS notification (via HHS breach portal)
☐ Individual notification (written, first-class mail)
☐ Media notification (if 500+ individuals in one state)
☐ State attorney general notification (if required)

30-90 days: REMEDIATE
☐ Patch vulnerability
☐ Update controls
☐ Retrain staff
☐ Document corrective actions

### 5. DEPLOYMENT RISK CLASSIFICATION

| Use Case | Risk Level | Required Controls |
|----------|-----------|-------------------|
| Appointment scheduling | MEDIUM | Standard safeguards |
| Billing inquiries | HIGH | Enhanced audit + DLP |
| Insurance verification | HIGH | Data minimization + encryption |
| Patient communication | HIGH | Consent + encryption + retention limits |
| Clinical decision support | CRITICAL | Full controls + human-in-loop |
| Medical records access | CRITICAL | Full controls + access logging + alerts |

Your deployment (scheduling + billing) = MEDIUM-HIGH.
Manageable with standard technical controls.

### 6. PENALTY REFERENCE

| Tier | Per Violation | Annual Cap |
|------|-------------|------------|
| Unknowing | $141-$71,162 | $2,134,831 |
| Reasonable cause | $1,424-$71,162 | $2,134,831 |
| Willful neglect (corrected) | $14,232-$71,162 | $2,134,831 |
| Willful neglect (not corrected) | $71,162 | $2,134,831 |

Average healthcare data breach cost: $10.93M (2025).

skills:
  - afrexai-hipaa-compliance
  - afrexai-medical-billing  # For billing-specific compliance