AI use policy template: Healthcare edition

1. Purpose

This policy says what artificial intelligence (AI) tools may and may not be used for at [Company name]. It exists so that everyone on the team makes the same call, and so that a client, insurer or regulator who asks gets one clear answer.

Protected health information (PHI) makes the stakes higher here: an AI tool that sees PHI is a business associate, and the organization answers for what it does with that data.

2. Scope

• This policy applies to everyone who works for or with [Practice or organization name]: clinicians, administrative staff, billing and prior authorization staff, contractors, students and volunteers.
• It covers every AI tool: chat assistants such as ChatGPT, Claude, Copilot and Gemini, AI features inside software we already use, and any tool with AI in it that a team member signs up for.
• It covers work done on company devices and on personal devices when the work is for [Company name].
• It covers clinical, administrative, billing and patient-communication uses of AI.

3. PHI boundary: identified, de-identified, neither

Every workflow that could involve AI is placed in exactly one of three groups. A workflow that is not listed is in the third group until the policy owner moves it.

• May see identified PHI, only inside a tool that is approved for it and covered by a signed business associate agreement: [List the workflows, for example: clinical documentation inside the EHR's own AI feature; prior authorization letters inside the approved tool]
• May see de-identified data only, with all eighteen Safe Harbor identifiers removed before anything is entered (names, geographic subdivisions smaller than a state, all dates except year, phone and fax numbers, email addresses, Social Security numbers, medical record numbers, health plan numbers, account numbers, certificate and licence numbers, vehicle identifiers, device identifiers, web addresses, IP addresses, biometric identifiers, full-face photographs, and any other unique identifying number or code): [List the workflows, for example: drafting patient education material; summarizing published guidance; template letters with no patient details]
• May see neither identified nor de-identified patient data: [List the workflows, for example: general-purpose chat assistants on personal accounts; marketing copy; staff scheduling]
• Ages over 89 and dates that reveal them are an identifier. Free text that could identify a patient is treated as identified.

4. What AI may touch

• Workflows in the first two groups of the PHI boundary above, inside the tool named for them.
• Internal procedures, training material, staff communications and public-facing text that contains no patient information.
• Summaries of published clinical guidance, payer policies and regulations.
• [Add or remove lines to fit the practice]

5. What AI may never touch

• Identified PHI in any tool that is not approved for it under a signed business associate agreement.
• A medical necessity decision, a coverage decision or a clinical decision as the sole basis: a licensed person decides, and the record shows who.
• Patient communications sent without the disclosure that AI helped generate them, where state law requires that disclosure.
• Passwords, EHR credentials, API keys and security settings.
• Employee health information and personnel records.
• [Add any data category your state or your payer contracts restrict]

6. Approval of new tools

• [Name or role, for example the privacy officer] approves every new AI tool before anyone uses it for company work. Trying a tool with company or customer data counts as using it.
• Before approval, we check: where the tool stores what we type; whether our inputs train the vendor's models; who at the vendor can see our data; how we delete it; what it costs per month and per seat.
• For any tool that will see PHI: a signed business associate agreement that explicitly covers AI training, model inputs and downstream use, before the first use.
• For any tool that generates patient-facing text: the disclosure wording required in [State], and where it will appear.
• The list of approved tools, and what each one is approved for, lives at [Location of the approved tools list].

7. Human review

• A person checks every AI-written document before it goes to a client, customer, regulator or the public. The person who sends it is responsible for it, not the tool.
• AI output is a draft. We do not send an answer we have not read from start to finish.
• Facts, figures, names, dates and quotations in AI output are checked against the source before they go out.
• A licensed clinician reviews and signs any clinical content before it reaches a patient, a payer or the chart. The clinician's review is the supervision; the tool is not supervised by another tool.
• Patient communications generated with AI carry the disclosure required in [State]: [Disclosure wording].

8. Records

• When AI helped write something we sent out, we note which tool was used, what we gave it, and who checked the result. The record lives at [Location of the AI use log].
• Records are kept for [12] months, or longer if a contract or regulator requires it.
• For PHI workflows, the record also shows which business associate agreement covers the tool and the date it was signed.
• Records of AI-assisted clinical documentation are kept with the chart under the practice's retention schedule.

9. Vendor terms

• Before a tool is approved, [Name or role] reads its terms for four things: whether our inputs train the vendor's models; how long the vendor keeps our data; where that data is stored; and who owns the output.
• Where a vendor offers a setting that stops training on our inputs, it is switched on before the tool is used, and the setting is checked every [3] months.
• Terms are re-read every [12] months and whenever the vendor announces a change.
• No AI tool touches PHI without a signed business associate agreement that explicitly covers AI training, model inputs and downstream use. Agreements are kept at [Location]. A vendor that will not sign one is not used with PHI.

10. Training

• Every team member reads this policy when they join and again once a year, and confirms in writing that they have.
• [Name or role] runs a short session on safe AI use at least once a year: what the tools do with what we type, how to check an answer, and what never goes into a chat window.
• Training covers the three-group boundary, the eighteen identifiers, and the disclosure rules for AI-generated patient communications in [State].

11. Review date

• This policy is reviewed on [Review date], and sooner after any incident, any new tool, or any change in the law or in a client's requirements.
• The review checks that the approved tools list is current, that the records are being kept, and that the training has happened.

12. Sign-off

• Owner of this policy: [Name, role]
• Approved by: [Name, role]
• Date approved: [Date]
• Version: [1.0]
• Next review: [Review date]