Ideas · Practical implementation

The first mistake is making AI an IT project.

IT can tell you which tools are safe to use. Your agency still has to decide what it needs AI for.

When an agency starts taking AI seriously, the first questions tend to be about the tool. Which one should we buy? Is it secure? Who gets a license? Those questions need answers, and I want IT and compliance in the room answering them. They know things the rest of us don’t about where data goes and what can go wrong.

The trouble starts when AI gets handed to IT and everyone else waits for the rollout. IT can make sure a tool is safe. They have no way of knowing that intake is typing the same information into two systems, or that one supervisor answers the same question ten times a week, or which of the agency’s rules wins when two of them disagree. The people who know those things are spread all over the building, and most of them were never asked.

Bringing in AI changes how work gets done, who does it, who checks it and what happens with the time. Behavioral health already knows a lot about changes like that. It’s what I know best, so it’s where my examples come from, and I’d say the same thing to a school district or a county agency that doesn’t know where to start.

The tool

The IT questions

  • Which tool is approved?
  • Where does the data go?
  • Who gets access?
The work

The agency questions

  • Which work is taking up the week?
  • Which rule applies, and who decides?
  • Where does the time go?

We already know how this goes

While I was in grad school for social work, I worked as a psychiatric care tech on a child and adolescent inpatient unit. The unit was moving to trauma-informed care, right around when that idea was starting to get real attention, and I helped support management through it. It was the first time I really saw a culture shift happening. Years later, as a statewide trainer, a lot of my work was helping teams implement a model in their own agencies, with their own constraints.

Anyone who has done that work knows a training doesn’t change practice by itself. You can send everyone to the two-day training, hand out the manual and check the box. Practice changes in the supervision afterward, and when leaders keep asking about it, and when someone notices that the model and the real day don’t fit together yet.

AI is a culture shift too, and it’s going to work the same way. Buying the licenses is the two-day training. It can be a good start, and by itself it won’t change much about anyone’s week. Agencies need champions for this: people at different levels and in different departments who use it themselves, talk about what worked and keep it going after the kickoff is over.

Buying the licenses is the two-day training.

Start with the work

The next thing people tend to want is a brainstorm: get everyone together and ask what AI could do for us. I’d hold off. Without a plan, a brainstorm gives you a long list with no way to choose between the ideas on it, and the ones that win are usually the most exciting ones or the ones from the most senior person in the room.

Start by finding out what the work is. Ask people across departments which tasks take up their week, and then check the hours, because what people dislike most isn’t always what takes the most time. A ten-minute task that happens forty times a week can matter more than the monthly report everyone complains about.

Then write down how that work actually gets done. A lot of SOPs describe how something was supposed to go a few years ago. Sit down with three people who do the same task and you may find three different processes and a workaround nobody told you about. You need to know that before AI gets anywhere near it, because AI will speed up whichever version you hand it. And some of what you find won’t need AI at all. Sometimes the fix is a form, a shared folder or one decision about who owns it.

This doesn’t have to mean someone types for a month. I’d meet with the people from the different departments who touch that task, have them walk through how it really gets done, and record the meeting. Tell everyone it’s being recorded, use a tool your agency allows and keep client information out of the conversation. Then AI can help pick it apart and make it better: where the same information gets entered twice, or where two departments each think the other one owns a step. The people who were in the room correct what it got wrong. People know a lot more than they’d ever sit down and type.

Put the right people at the table

I’d form an AI core team with people from different levels and departments. Front desk, billing, a supervisor, someone who works directly with the people you serve, someone from IT, someone from compliance. And include the people who are already using AI.

Some of your staff are. Some of them have stayed quiet about it because no one has said whether it’s allowed, and if the first thing they hear from leadership is a warning, they’ll keep staying quiet. They’ve already learned what it does well and where it makes things up. You want that on your team, and they’re often your first champions.

A group like this also keeps everyone honest about where the burden really is. A director may think the problem is the reports. The person doing intake can tell you it’s something else.

Someone at the top has to say what this is for

I’ve gone back and forth on how much of this has to start with the CEO. I do believe CEOs need to lead it, and staff need to see that they’re all in. All in doesn’t mean signing the contract. It means using AI yourself, telling people what you tried and what you had to fix, and saying out loud what it’s for in your organization.

It also means answering the question staff are asking whether or not they say it: what happens to me? I wrote about that in my first note. AI takes tasks, and then someone has to decide on purpose where the time goes. If an agency can’t say where the time goes, people will assume the answer is more work, and they may be right.

If an agency can’t say where the time goes, people will assume the answer is more work.

I also don’t think you have to wait for your CEO to get started. Leaders at every level can learn on their own work, in a tool the organization allows, with nothing from a client record. Prep for a one-on-one. Draft the agenda you’ve been putting off. Count the whole job while you’re at it: the setup, the reading, the fixing and the follow-through. Then bring one real example to your CEO. “I used this to prep for my one-on-ones, this is what it saved and this is what I had to fix” is a much easier conversation than “we should have an AI strategy.”

Where to start

You can do the first part of this without buying anything.

  1. 1

    Say what it’s for

    One sentence from leadership on the problem AI is here to help with. If no one can write that sentence yet, that’s your first conversation.

  2. 2

    Find out who’s already using it

    Ask in a way that makes it safe to answer.

  3. 3

    Check the hours

    Ask a few people in different departments which tasks take up their week. Look for one that gets done the same way every time.

  4. 4

    Pick work someone can check

    Choose a task where the person doing it will know right away if the result is wrong. Anything that touches client records or clinical decisions comes later, inside your agency’s approved tools and policies.

Do that before anyone picks a tool and your conversation with IT gets a lot better. You’ll be able to tell them what you need AI to do, who needs it and what it has to stay away from. I suspect they’ve been waiting for someone to bring them that.

If you’d like help finding your starting point, I built a five-minute assessment for that. AI First Step shows you where your organization should start and gives you a 30-day plan to share with your leadership team.

About the author

Joyce Agne is a licensed independent social worker with 15+ years in behavioral health: inpatient care, trauma therapy, home-based work with families, statewide training and regional crisis leadership. She builds practical AI systems for behavioral health organizations and experts. More about Joyce

Get the next note

Short, practical notes on AI in behavioral health, with the caution beside them.

Send me the notes

Working through this with your team?

Tell me what you're trying to work through.

Start the conversation