Privilege and governance: using AI without losing protection
Answer in brief: AI doesn't create privilege and doesn't destroy it on its own; how and why the analysis is made decides that. The medical record is never patient safety work product, so a chart timeline is protected only if it is made inside your patient safety evaluation system or at counsel's direction. Settle the route with counsel before the first upload, use only tools under a BAA, and write the rules down.
Limits: We read each vendor's own web pages and the cited studies on 2026-11-14 and did not upload records to any tool, ours included. Nothing here is legal, actuarial or clinical advice; privilege and reserving rules vary by state and by program.
3 protections risk teams rely on, and what AI changes
Hospital risk work leans on 3 different protections. They have different sources, cover different things and fail in different ways. This is a map for the conversation with counsel, not legal advice.
| Protection | Source | Covers | Doesn't cover | What AI changes |
|---|---|---|---|---|
| Patient safety work product | Federal: Patient Safety Act and 42 CFR Part 3 | Analyses such as root cause analyses assembled for reporting to a PSO, or deliberations within a patient safety evaluation system | The medical record, billing and discharge information, and information collected separately from the system | Where the AI output is made and recorded: inside the system, or outside it |
| Peer review privilege | State statutes, which differ | Committee deliberations and records, as each state defines them | Usually the underlying facts and records available from original sources | Whether committee material is shared with an outside tool and under what terms |
| Attorney-client privilege and work product | Common law and court rules | Communications with counsel; material prepared in anticipation of litigation | Facts themselves; ordinary business records | Whether the AI work is done at counsel's direction and kept in counsel's file |
The federal rule is precise on 1 point that matters for AI. Patient safety work product "does not include a patient's medical record, billing and discharge information, or any other original patient or provider information", nor information "collected, maintained, or developed separately" from a patient safety evaluation system (42 CFR 3.20, read 2026-11-14). Where it applies, the privilege is strong: 42 CFR 3.204 shields it from subpoena and discovery, subject to listed exceptions.
Where an AI timeline sits
Take the same output, a page-cited chart timeline, made 3 ways:
- Made inside your patient safety evaluation system, for reporting to your PSO, and documented as such with the date it entered the system. It may qualify as patient safety work product. The chart it was built from does not.
- Made by the risk department for the claim file, outside the safety system and without counsel's direction. It is likely an ordinary business record, and may be discoverable.
- Made at defense counsel's direction after a claim is anticipated, kept in counsel's file. It may be work product, subject to the usual limits.
Many hospitals will want 2 separate tracks, 1 for safety learning and 1 for the claim, with separate outputs. Running 1 AI job and copying its output into both is the easy mistake.
Questions to settle with counsel before the first upload
- Which track is this work on: patient safety evaluation system, claim file at counsel's direction, or ordinary operations?
- Who engages the vendor for claim work: the hospital or defense counsel?
- How are outputs labeled and where are they stored, so they stay on the chosen track?
- Does the vendor agreement keep our files confidential, bar model training on them, and limit who at the vendor can see them?
- Can the vendor delete files and outputs on request, and confirm it in writing?
- Does the incident report go into the tool at all, or only the chart?
- What does our state's peer review statute say about sharing committee material with a service provider?
- If a reviewer or expert uses AI, how will that be described if asked at deposition?
Security and contract checks for AI vendors
Before any chart leaves your EHR, the vendor is a business associate under HIPAA and needs a signed Business Associate Agreement. Consumer chatbot accounts don't come with one. Beyond the BAA, ask:
- Do our records train any model, ours or anyone else's, and where does the contract say no?
- Which subprocessors touch the records (OCR, AI model hosts, storage), and in which countries?
- Can we see a current SOC 2 report under NDA, not just a badge on the website?
- What access controls exist per matter, and is every access logged?
- Is there a deployment that keeps records inside our network, if our policy requires it?
AI use policy template for the risk department
Copy and adapt. Keep it to 1 page so people read it.
AI USE IN RISK MANAGEMENT AND CLAIMS, [hospital], effective [date] 1. Approved tools: only tools on the approved list, each with a signed BAA and a contract term that our records never train a model. No consumer chatbots for any patient, claim or committee information. 2. Tracks: before upload, the user records the track: patient safety evaluation system, claim at counsel's direction, or operations. Outputs are labeled and stored on that track only. 3. Who uploads: [roles]. Claim files at counsel's direction: [roles]. 4. Checking: no AI line goes into an RCA, ECA memo, reserve rationale or committee packet until a person has read its cited page. 5. Judgment: AI output may list facts, gaps and questions. It may not state whether care met the standard, assign fault, or set a reserve. 6. Retention: outputs follow the retention rule of their track. Files are deleted from the vendor when the matter closes; deletion is confirmed in writing. 7. Logging: the tool's access log is kept with the matter. 8. Review: the approved list and this policy are reviewed every [12] months with counsel and information security. Owner: [risk manager]. Approved: [counsel], [CISO], [date].
Questions about privilege and AI
Does using AI waive patient safety work product protection?
Not by itself. Protection turns on whether the material is part of a patient safety evaluation system and reported to a PSO, and on how it is disclosed. Sharing it with a vendor under a BAA and a confidentiality agreement is different from pasting it into a consumer chatbot. Ask counsel how your program's rules apply.
Is an AI-made chronology of the chart privileged?
The chart never is. A chronology may be protected if it is made inside the patient safety evaluation system or at counsel's direction in anticipation of litigation, and kept on that track. The same chronology made for routine operations probably isn't.
Should experts disclose that they used AI?
Expect the question at deposition. A reviewer who used a tool for sorting and a timeline, read the cited pages and formed the opinion personally has a simple answer. Agree with counsel in advance how it will be described.