Healthcare organizations are uploading patient data to external AI platforms faster than they're updating their vendor risk frameworks. A recent breach at Loma Linda University Health illustrates the consequences of inadequate AI platform governance: medical record numbers, dates of birth, and clinical information ended up in an uncontrolled environment. UCLA Health's separate disclosure incident, spanning December 2024 through April 2026, highlights a critical issue: your breach notification process is only as effective as your detection capability.
If you're a Data Privacy Officer watching your clinical teams adopt AI tools for research, diagnostics, or workflow optimization, you need a governance program that addresses the unique risks these platforms create. Here's how to build one.
What You Need Before Starting
Inventory your AI platform usage. You can't govern what you don't know exists. Before writing a single policy, survey your clinical departments, research teams, and IT groups to identify which external AI platforms are currently in use or under evaluation. Document the vendor name, platform purpose, data types processed, and sponsoring department.
Identify your regulatory baseline. For HIPAA-covered entities, you're working within the HIPAA Privacy Rule and HIPAA Security Rule. If you're subject to the Health Information Technology for Economic and Clinical Health Act, you have breach notification obligations triggered when you know or should have known about unauthorized access. If your organization operates in multiple jurisdictions, map applicable state breach notification laws as well.
Assign clear ownership. Third-party AI platform governance requires collaboration between your privacy office, information security team, legal counsel, and clinical leadership. Designate a single owner, typically the Data Privacy Officer or CISO, with authority to approve or reject platform use.
Establish your risk tolerance. Define what constitutes acceptable risk for different data categories. Protected health information used in research may have different thresholds than diagnostic imaging analysis or administrative workflow tools.
Step-by-Step Implementation
Step 1: Create a tiered vendor assessment framework
Build a risk classification system based on data sensitivity and platform functionality. Tier 1 platforms process identifiable patient data or perform clinical decision support. Tier 2 platforms handle de-identified or limited datasets. Tier 3 platforms operate on aggregated or non-patient data.
For each tier, define mandatory controls. Tier 1 platforms require Business Associate Agreements under HIPAA, evidence of SOC 2 Type II certification, and documented data residency commitments. Tier 2 platforms need BAAs if any re-identification risk exists, plus encryption-at-rest and encryption-in-transit validation. Tier 3 platforms still require basic security attestations.
Document these requirements in a vendor assessment template that your procurement and compliance teams can apply consistently. Include specific questions: "Does the platform train its AI models on customer data?" "Can you guarantee data deletion within 30 days of contract termination?" "What sub-processors have access to our data?"
Step 2: Build a pre-deployment approval workflow
No AI platform should process patient data without documented approval. Your workflow should include:
- Initial request form capturing business justification, data elements involved, and expected patient volume
- Privacy impact assessment reviewing HIPAA compliance, consent requirements, and Data Minimisation principles
- Security review validating encryption, access controls, and logging capabilities
- Legal review confirming contract terms, liability allocation, and breach notification obligations
- Clinical leadership sign-off confirming the platform serves a legitimate treatment, payment, or healthcare operations purpose
Store approved assessments in a central repository. When Loma Linda's incident occurred, their investigation had to determine what data was uploaded. If you've documented approved data elements in advance, your investigation starts with a baseline instead of forensic reconstruction.
Step 3: Implement technical controls at the boundary
Your network architecture should enforce your governance decisions. Deploy data loss prevention rules that flag or block uploads to unapproved cloud services. Configure your firewall to log all connections to known AI platform domains.
For approved platforms, create dedicated service accounts with Role-Based Access Control. A researcher uploading data to an AI platform shouldn't use their personal credentials; they should authenticate through a monitored service account that logs every transaction.
If your platform supports it, implement Just-in-Time Access for administrative functions. Researchers get read/write access only during active analysis sessions, not standing privileges.
Step 4: Define data preparation standards
Clinical teams often upload raw datasets without proper de-identification. Create a mandatory data preparation checklist:
- Remove direct identifiers (names, Social Security numbers, medical record numbers) unless clinically necessary
- Generalize dates to month/year or use date-shifting techniques
- Suppress small cell sizes that could enable re-identification
- Document the de-identification method applied and retain the mapping key separately under restricted access
For research studies, coordinate with your Institutional Review Board to ensure your de-identification approach aligns with approved protocols. The Loma Linda incident involved an IRB-approved study, but the data handling didn't match the approved risk profile.
Step 5: Build detection and monitoring capabilities
You need to know when unauthorized uploads occur. Deploy user and entity behavior analytics to flag anomalous data transfers. Configure alerts for large file uploads to cloud storage services, especially outside business hours.
For approved platforms, require audit logging and integrate those logs into your security information and event management system. Monitor for:
- Access from unexpected IP addresses or geographies
- Bulk data downloads by service accounts
- Changes to sharing permissions or access controls
- Failed authentication attempts
UCLA Health's disclosure incident ran from December 2024 through April 2026 before detection in July 2026. That's an 18-month window. Your monitoring program should detect unauthorized disclosures in days, not quarters.
Validation: How to Verify It Works
Test your approval workflow. Have a clinical department submit a request to use a new AI diagnostic platform. Track how long approval takes and whether all required assessments were completed. If the request bypassed security review or legal never saw the contract, your workflow has gaps.
Run a tabletop exercise. Simulate a scenario where a researcher uploads an entire patient registry to an unapproved AI platform. Walk through your detection, investigation, and breach notification process. Can you determine within 24 hours what data was exposed? Can you meet the 72-Hour Notification Requirement if the incident qualifies as a breach?
Audit platform usage quarterly. Pull logs from your approved AI platforms and compare them against your approved use cases. If a platform approved for orthopedic research is now processing cardiology data, you've got scope creep that needs immediate review.
Verify vendor compliance annually. Request updated SOC 2 Type II reports, penetration test results, and attestations of HIPAA compliance from your Tier 1 vendors. If a vendor can't produce current documentation, escalate to your risk committee for possible platform suspension.
Maintenance and Ongoing Tasks
Update your vendor inventory monthly. New AI platforms launch constantly, and clinical teams adopt them faster than procurement cycles. Make vendor discovery a standing agenda item in your privacy committee meetings.
Refresh workforce training every six months. Loma Linda indicated they're reviewing workforce training on external technologies. Don't wait for an incident. Train clinical staff on approved platforms, data preparation requirements, and how to request new tool evaluations. Include real examples of what goes wrong when researchers upload raw datasets.
Review and update contracts annually. AI platform terms of service change frequently. Schedule annual contract reviews to confirm your Business Associate Agreement still covers the platform's current functionality and that liability terms haven't shifted unfavorably.
Track regulatory developments. The EU AI Act creates new obligations for high-risk AI systems, and NIST AI Risk Management Framework provides guidance even though it's not mandatory. Monitor proposed regulations and adjust your governance framework accordingly.
Measure your program's effectiveness. Track metrics that matter: number of platforms assessed, approval cycle time, incidents detected, and time-to-detection for unauthorized use. If you're not detecting any violations, your monitoring probably isn't working.
Your AI governance program isn't a one-time project. It's an operational capability that evolves as your organization adopts new technologies and as regulations tighten. Build it now, before your next research study becomes your next breach notification.
