AI in Behavioral Health: Where It Helps, and Where Humans Decide

Most behavioral health CEOs I talk to have the same three things on their desk. Vendors in the inbox promising AI that will fix documentation. A staff split between people already using it and people who won’t touch it. And a board asking what the organization is doing about AI.

That’s a lot of pressure to buy something.

Here’s what I tell them first. AI in behavioral health works when you know exactly what it drafts, what it sorts and what stays human. Most organizations skip that step. Then they wonder why the tool they paid for sits switched off.

Your staff are already using AI

You probably don’t get to decide whether AI enters your organization. It’s there.

In the American Medical Association’s survey of nearly 1,200 physicians, 66% said they used health AI in 2024, up from 38% in 2023. Across all U.S. jobs, Gallup found 45% of employees used AI at work at least a few times a year by the third quarter of 2025. In the same survey, 23% didn’t know whether their own organization had implemented AI at all.

That last number is the one I’d worry about. It means people are using tools nobody has approved, with data nobody has drawn a line around.

So the real question isn’t “should we use AI?” It’s “which work do we hand it, and who checks it?”

AI is a mirror

AI gives you back what you give it. A messy workflow in means a messy result out.

Most people expect a super intelligence. What they get is a system trying to produce the best-looking answer to the question it was asked. Sometimes that’s the right answer. Sometimes it only looks right.

That’s why I start with the workflow, not the tool. When the workflow is clear, you often find you’re three settings away from the thing a vendor was about to sell you.

The 4 S’s: how I run AI work

I walk every organization through the same four steps. Skip one and it comes back to bite you.

1. Surface. List every tool you already pay for and every AI feature inside it. Is it on or off? Who owns the switch? If it’s off, why? This takes about an hour. If nobody asked legal, I can almost guarantee it’s off.
2. Sort. Decide what AI takes off your plate and where the human stays. More on this below.
3. Set guardrails. Write down what prevents harm before it gets bigger.
4. Supervise. Name who checks the AI, how often, and what happens when it’s wrong.

Sort: three questions before AI touches anything

Ask these about every task.

Would a human judgment call change the outcome? If yes, a person keeps the call. If you can write down how the decision gets made, the way you’d train a new hire with a rubric, AI can help you get there faster.

Would the person affected be upset to learn AI did it? Nobody minds if AI schedules an intake. People mind if AI decides their level of care.

How fast does a mistake surface, and how bad is it? A wrong reminder gets fixed in minutes. A wrong denial, a wrong discharge note or a wrong rejection might never come back.

Then make three piles.

AI can draft	AI can sort	Stays human
Intake packets, reminders, first drafts of notes and summaries	Missing paperwork, duplicate entries, items that need a checkbox	Clinical decisions, hard conversations, anything a client would want a person for

The human is primary. AI supports the team so the team can do the human part better.

Set guardrails: five sentences, not a binder

You don’t need a binder-sized AI policy to start. You need five decisions, made at a senior level.

1. Disclosure. What do we tell clients and staff about how AI is used?
2. Data boundary. What can go into this tool, and what can’t? In behavioral health that depends on your business associate agreement, and on 42 CFR Part 2 if you run a substance use program it covers.
3. A named decider. One role, not one person, owns the workflow.
4. Records. What do we keep, and for how long?
5. Sunset. When do we look at this again? Put the date in the calendar now.

Fill in the detail as you learn. The five sentences get you moving safely.

Supervise: put your AI on a performance improvement plan

You would never let a new hire work for six months without a check-in. Don’t do it to your AI either.

For the first 30 days of any new tool, a human reviews its output. Not a token sample. Real review. After that, pull a set of recent decisions on a regular schedule and ask: would I have done the same thing? Why or why not?

This is more than good practice. Federal rules under Section 1557 of the Affordable Care Act require covered entities to make reasonable efforts to identify patient care decision support tools that use race, color, national origin, sex, age or disability as inputs, and to reduce the risk of discrimination from their use. “The vendor handles it” is not a plan.

What good looks like

The evidence on AI in clinical documentation is early but real. In a 2025 study across six U.S. health systems, 263 physicians and advanced practice providers used an ambient AI scribe for 30 days. The share reporting burnout fell from 51.9% to 38.8%.

Two cautions. That study wasn’t specific to behavioral health. And it measured clinicians who chose to try the tool. Your results depend on your workflow, your staff and your supervision. Which brings us back to the 4 S’s.

Where to start this week

• Spend one hour listing every tool you pay for and every AI feature in it.
• Pick one low-risk, high-friction task. Intake reminders are a good first one.
• Write your five sentences.
• Name the person who checks the output, and put the 30-day review in the calendar.

Then poke holes in it. Ask your team what would go wrong. I need the holes, and so do you.