AI policy

This policy sorts AI tools into three categories, each with its own rules, so the safe uses are written down as well as the forbidden ones. Update Category 3 for the tools you actually use, and re-check vendors' AI features and business associate agreement (BAA) coverage at each annual review. For what the permitted uses look like day to day, see using AI to run the practice.

Practice: [Practice Name] Effective: [Date] Last reviewed: [Date] Reviewed by: [Your Name], Privacy & Security Officer Review cadence: Annual, and whenever AI tooling or U.S. Department of Health and Human Services (HHS) guidance changes materially

Purpose

Keeps the practice's use of artificial-intelligence tools within HIPAA's Privacy and Security Rules. AI tools fall into three categories:

1. General-purpose AI without a BAA: consumer or standard business tiers of ChatGPT, Claude, Gemini, Perplexity, and similar.
2. Clinical AI with a BAA: AI scribes and note assistants built for behavioral health that sign BAAs.
3. AI features embedded in tools the practice already uses: the electronic health record (EHR), email suite, and operating system. Coverage depends on whether the vendor's existing BAA extends to the feature.

Scope

Any use of AI in practice business (documentation, scheduling, billing, marketing, administration, professional communication, and research), including by any future contractor. Personal, non-practice use is out of scope.

Category 1: General-purpose AI without a BAA

No protected health information (PHI). None. That includes:

• names, initials, nicknames, dates of birth, addresses, phone numbers, emails
• session or appointment dates tied to a client
• diagnoses, presenting concerns, treatment-plan content
• direct quotes or close paraphrases from sessions
• payment or insurance details for a specific client
• any combination of details (city, occupation, and presenting problem) that could identify someone

Permitted:

• Drafting non-PHI content: website copy, blog posts, bios, directory profiles, colleague outreach, vendor emails, and templates.
• Research on clinical topics, interventions, theory, continuing education.
• Creating general psychoeducational handouts not tied to a client.
• Practice administration: drafting policies, bookkeeping structure, technology questions.
• Website and technical work.

De-identification standard: when drawing on case experience to ask a general question, remove all 18 HIPAA Safe Harbor identifiers (45 CFR §164.514(b)(2)) and make sure what remains couldn't identify the person in combination. When in doubt, leave the detail out.

Category 2: Clinical AI with a BAA

Permitted for the specific clinical use the BAA covers, only when all of these are true:

• A signed BAA is on file with the vendor.
• The BAA covers this specific use (transcription, summarization, note drafting).
• The tool has been added to the Security Risk Analysis system inventory.
• Adoption is recorded in the HIPAA Process Log.
• Clients are told in the Disclosure Statement/Informed Consent (including whether sessions are recorded) and can decline without losing access to services.
• The practitioner reviews every AI output before it enters the record. AI output is never the sole source of truth in the chart.

Tools in use:

Before adopting one: review its data flow, retention, sub-processors, and breach history, and update the risk analysis.

Category 3: Embedded AI features

Default posture: turn off AI features that send data outside the BAA-covered path, unless the vendor states in writing that the existing BAA covers the feature.

Document each one:

• [EHR] AI features (e.g. AI note-takers): if it processes session content, it's Category 2; apply that checklist before enabling. [Status: not enabled / enabled on date.]
• [Email suite] AI features (e.g. Gemini in Google Workspace, Copilot in Microsoft 365): check the vendor's list of BAA-covered services, which may exclude some surfaces (such as a browser-sidebar assistant). Record which is which. [Status.]
• Operating-system AI (e.g. Apple Intelligence, Windows Copilot): no BAA is offered. Do not deliberately run these on content that identifies a client. [Record your posture.]
• Browser extensions and writing assistants (AI summarizers, grammar tools): not installed in the work environment unless BAA-vetted, since many send what you type to their servers.

Cross-cutting rules

• No recording without explicit written consent, even with a BAA-covered tool.
• No PHI in search bars, URLs, or web-form free-text fields that aren't BAA-covered.
• No PHI in screenshots shared with AI tools for tech help.
• No AI in place of clinical judgment on high-stakes decisions. Output is advisory.

Practitioner-specific notes

[Describe your actual use honestly, e.g. "The practitioner uses [tool] for marketing copy, policy drafting, and administrative tasks. This is Category 1; no PHI is entered under any circumstance. Output is reviewed before use."]

Client-facing disclosure

If a Category 2 tool is adopted, update the Disclosure Statement to cover: the tool's name and purpose, what client information it accesses, whether sessions are recorded or transcribed, the BAA and where retention is handled, and the client's right to decline.

Incidents

If PHI is entered into a Category 1 tool:

1. Stop. Don't continue that conversation.
2. Run the four-factor breach assessment (HIPAA policies and procedures, under Breach notification).
3. Log it in the Process Log.
4. Use the vendor's delete function if one exists, knowing it may not reach backups.
5. Decide whether notification is required.

Review

Reviewed annually and on adoption of any new AI tool, material HHS guidance on AI, a vendor changing its HIPAA position, or any incident or near-miss.

Template shared in the Private Practice Toolkit. Not legal, tax, clinical, or compliance advice.