Training Your Staff to Use AI Without Leaking PHI

Your biggest AI risk this month probably isn't a fancy agent. It's a tired staff member pasting a note into a free chatbot to fix the grammar.

Nobody in that story is a bad person. They're trying to get their work done. That's exactly why training matters more than rules on paper.

Why this happens

AI tools are everywhere now. They're on phones, in browsers, in email, in keyboards. Staff use them at home and naturally reach for them at work. Most people have never been told that the free version of a chatbot isn't covered by a Business Associate Agreement, or that typing a client's situation into it can be a disclosure.

In behavioral health, the stakes are higher. Substance use disorder records may fall under 42 CFR Part 2 in addition to HIPAA. HHS's Office for Civil Rights began accepting Part 2 complaints and breach notifications on February 16, 2026, per its announcement. And a leak in this field can cost someone a job, custody, or their safety.

The one-page policy

Adapt this with your counsel. Keep it to one page. If it's longer, nobody reads it.

Approved tools. List them by name, with what each is approved for. "Tool A: approved for session note drafting, BAA on file. Tool B: approved for marketing and general writing, no patient information."

The green list. You can use approved general AI tools for:

• Drafting emails, flyers, handbooks, and blog posts with no patient details
• Learning and research questions with no patient details
• Formatting reports from totals and averages
• Brainstorming, outlines, and editing your own writing

The red list. Never put these into a tool that isn't approved for them:

• Names, initials paired with details, phone numbers, emails, addresses
• Dates of admission, discharge, or birth
• Diagnoses, drug screen results, medications, treatment details
• Insurance information
• Anything from a chart, note, call recording, or text thread
• Photos or screenshots of any of the above

The stop list. No tool does these, approved or not:

• Responding to a crisis
• Making a clinical decision
• Counseling a patient or resident
• Promising coverage or cost
• Anything involving payment for referrals

If you make a mistake. Tell your supervisor or privacy officer the same day. You won't be punished for reporting. Hiding it is the problem.

Personal devices. Say plainly what is and isn't allowed on personal phones, including AI keyboards and assistant apps.

Why "removing the name" isn't enough

This is the most common misunderstanding, so train on it directly.

HIPAA's Safe Harbor de-identification method lists 18 kinds of identifiers that have to come out, including dates, small geographic areas, phone numbers, and record numbers, according to the HHS de-identification guidance. "My 34-year-old client from Salmon who relapsed after her custody hearing on the 12th" has no name in it and still points at one person.

A simple rule for staff: if someone who knows your community could figure out who it is, it doesn't go in.

A 45-minute training session

Ten minutes: why. One realistic scenario, like the grammar fix. Ask the room what could go wrong. Let them get there themselves.

Ten minutes: the policy. Walk through the green, red, and stop lists. Hand out the one page.

Fifteen minutes: practice. Give people five short prompts written on cards. They sort them: green, red, or stop. Discuss the ones people disagree on. Those are your real training moments.

Examples you can use:

• "Write a flyer for our family night." (Green)
• "Summarize this group note for the chart." (Red unless the tool is approved for notes)
• "A caller says she wants to die. What should I tell her?" (Stop. Get a person, give 911 and 988.)
• "Rewrite our visitor policy in plain English." (Green)
• "Help me appeal this denial for John's residential stay." (Red unless approved, and remove what you can)

Ten minutes: how to report a mistake. Practice saying it out loud. Make it normal.

Train the people who get forgotten

House managers in sober living. Peer specialists. Front desk staff. Volunteers. Night techs. These people often handle the most sensitive information and get the least training. They're also often using their own phones.

Keep their version shorter. The red list and the "tell someone same day" rule are the core.

Training on the agents you deploy

If you put an AI agent on your phones or chat, staff need to know:

• What the agent does and doesn't do
• How a handoff reaches them and what information comes with it
• How to flag a bad conversation
• Who owns the agent and can shut it off
• What to do if the agent says something wrong to a patient or family

Have staff read a sample of real transcripts together once a month. It's the fastest way to learn what the agent is actually doing.

Where the privacy line sits

Safe with de-identified or aggregate data. General writing, learning, formatting totals, training materials built on fictional cases.

Needs a BAA and an approved tool. Any patient, client, or resident information, in any form. Add Part 2 review for SUD records.

Don't automate. Crisis response, clinical decisions, counseling, coverage promises, referral payments.

After a mistake

Someone will slip. When they do, the way you respond teaches the rest of the team whether to report next time. Thank them for telling you. Then work with your privacy officer and counsel to decide whether it's a reportable breach and what to do next. HHS's risk analysis guidance is a good reminder that every new tool belongs in your risk analysis, including the ones staff found on their own.

For the contract basics behind all of this, read the PHI and BAA primer. When you're ready to approve a new tool, use the vendor evaluation guide.

What good looks like

A staff member stops before pasting something and asks, "Is this one okay?" That pause is the whole goal. Build a culture where people pause and ask.