How to Use AI for Clinical Notes Without Violating HIPAA

Key takeaways

• If there is no signed BAA, PHI cannot go into the tool. Everything else in this guide builds on that rule.
• BAA first, then minimum-necessary input, access controls, audit logs, clinician sign-off, and training. In that order.

How to Use AI for Clinical Notes Without Violating HIPAA

TL;DR:

• Get a signed Business Associate Agreement (BAA) before any patient data touches an AI tool. Consumer AI tiers do not offer one.
• Share only the minimum necessary information, and remove identifiers whenever the task allows it.
• A clinician reviews and signs every AI-drafted note. The AI writes the draft; the clinician owns the record.

AI can turn a 15-minute documentation task into a 2-minute review. But the moment a clinician pastes patient details into the wrong tool, the practice has made a disclosure of protected health information (PHI) that HIPAA does not allow. This guide walks through how to get the time savings without the compliance risk. For the bigger picture on choosing and governing AI across a practice, see our pillar guide, HIPAA-Compliant AI for Healthcare Providers: The Complete 2026 Guide.

Why Clinical Notes Are the Highest-Risk AI Use Case

Clinical notes are dense with PHI: names, dates of service, diagnoses, medications, lab values. Under HIPAA, any outside vendor that creates, receives, maintains, or transmits PHI on your behalf is a business associate, and the HHS business associate contract requirements say you need a written agreement with that vendor before it handles the data.

That makes the most common mistake easy to spot. A clinician opens a free AI chatbot, pastes in a visit summary, and asks for a SOAP note. The output may be excellent. The disclosure is still a violation, because no BAA covers it. Good security on the vendor's side does not fix a missing contract. (We break down the ChatGPT question in detail in Is ChatGPT HIPAA Compliant?)

What a HIPAA-Safe AI Notes Setup Needs

A compliant setup has four layers. Missing any one of them leaves a gap.

1. A BAA that fits AI. The agreement should limit how the vendor may use your data, require safeguards under the HIPAA Security Rule, require breach reporting, and extend to any subcontractors. For AI tools, also confirm in writing that your data is not used to train the vendor's models.

2. Technical safeguards. Look for encryption in transit and at rest, access controls by role, and audit logs that show who viewed or changed what, and when.

3. Administrative safeguards. Adding a new system that handles PHI means updating your security risk analysis. Your written policies should say which AI tools are approved and what may be entered into them.

4. Data retention you control. Ask where your data is stored, how long it is kept, and whether it is deleted after processing. A tool that keeps every draft forever is a breach waiting to happen.

A Step-by-Step Workflow for AI Clinical Notes

Step 1: Sign the BAA before the pilot. Do not "just test" a tool with real patient data. Have your compliance officer or counsel review the vendor's BAA first, then start the pilot.

Step 2: Limit who can use it. Give access to the clinicians and staff who need it for documentation, and record who has access and why.

Step 3: Share the minimum necessary. If the note only needs the chief complaint, vitals, and assessment, do not paste the full chart. Where the task allows, replace identifiers: "Patient A, 53-year-old male" instead of a name and birth date. The HHS guidance on de-identification lists the identifiers that count. Our guide on writing better prompts for HIPAA-compliant AI in healthcare shows how to do this without losing note quality.

Step 4: Keep the audit trail. Make sure the tool logs access, and review those logs on a schedule. Look for unusual patterns such as bulk exports or access at odd hours.

Step 5: Review and sign every note. The AI drafts, and the clinician checks accuracy and completeness, corrects, and signs. AI can misstate a dose or invent a detail, so the signature has to mean someone actually read the note.

Step 6: Train before go-live. Every user should know which tool is approved, what never goes into a prompt, how to review an AI draft, and how to report a mistake. Repeat the training each year.

How to Choose a Tool for Clinical Notes

Use this checklist when you compare options:

Consumer AI tiers are fine for tasks with no PHI, such as drafting patient education handouts or summarizing a published guideline. They are not safe for notes.

EHR add-on AI modules can work, but do not assume your existing EHR agreement covers them. Ask the vendor to confirm in writing that the AI feature and any subprocessors behind it fall under your BAA.

Purpose-built HIPAA platforms are built for this job. Hathr.AI is HIPAA-Compliant, built on FedRAMP High Infrastructure, following NIST 800-53 controls, and Powered by a Government-Approved Claude. A BAA is provided on signup, data retention is zero by default, and there are no usage limits and no seat minimums. Teams can upload up to 100,000 pages per workspace and analyze 1,500+ page files accurately, which matters when a note depends on a long patient history. Hathr.AI is trusted by 3,000+ organizations, including the U.S. Dept of Health & Human Services. Learn more about HIPAA-compliant AI tools for healthcare and government.

Common Mistakes to Avoid

Using a free tier "just this once." One pasted visit summary is a disclosure. Make the approved tool easier to reach than the unapproved one.

Pasting the whole chart. More input is not better input. Send what the note needs.

Assuming the EHR agreement covers the AI. Get it in writing for the specific feature.

Letting drafts leave the system. AI drafts copied into personal email, notes apps, or USB drives create a second breach risk. Keep drafts inside the covered tool until they are signed.

Skipping the risk analysis update. A new AI tool is a new system that touches PHI. Document it.

Frequently Asked Questions

Can I use ChatGPT or Claude for clinical notes?

Not on consumer tiers. They do not come with a BAA, so PHI cannot go into them. Use them only for tasks with no patient information, and use a tool covered by a BAA for anything that touches PHI.

Is a BAA enough on its own?

No. The BAA is the legal foundation, but you still need access controls, audit logs, a current risk analysis, staff training, and clinician review of every note.

What should never go into an AI prompt?

Leave out direct identifiers such as names, medical record numbers, Social Security numbers, and full dates of birth unless the task truly needs them. Use ages, years, or "Patient A" instead. For the information you do share, stick to what the note needs.

Does my EHR's built-in AI fall under my existing BAA?

Not automatically. Ask your EHR vendor to confirm in writing that the AI feature, and any third party that powers it, is covered.

Ready to Use AI for Clinical Notes Safely?

Try Hathr.AI free. HIPAA-Compliant, built on FedRAMP High Infrastructure. BAA provided on signup. Zero data retention by default. No usage limits, no seat minimums. Try Hathr.AI free