The AI use cases that actually hold up in a therapy practice are not especially exciting.
Practice owners seem to be drifting toward two extremes. Some are putting AI into everything they can find, sometimes without really understanding where the data is going or what the system is doing. Others hear "AI," see the risks, and decide the safest response is to avoid it entirely.
Neither approach is particularly useful.
The better question is not whether your practice should use AI. It is what kind of work you should trust it to do.
For me, the most useful dividing line is between tasks that involve processing information and decisions that require professional judgment. AI can be extremely helpful with the first. I am much more cautious about the second.
Start With the Boring Work
Think about how much administrative work in a group practice involves taking information that already exists and turning it into something more usable.
A prospective client submits an inquiry containing their insurance, availability, presenting concern, and preferences. Someone has to read it and determine which clinicians might be a good fit. A therapist finishes a session with rough notes and needs to turn them into appropriate clinical documentation. An insurance explanation of benefits arrives and someone has to pull out the relevant payment information and reconcile it against client accounts. A meeting produces several decisions and someone needs to turn those decisions into actual tasks. Or an employee needs an answer that exists somewhere inside dozens of policies, procedures, emails, and reference documents.
These are different problems, but they have something important in common: the information already exists. A person is spending time reading it, organizing it, categorizing it, summarizing it, or moving it somewhere else.
AI can be very useful for that kind of work — not because it has some magical understanding of your practice, but because current AI systems are unusually good at working with language and patterns. Used well, that can save a meaningful amount of administrative time.
The important question is what happens after the AI produces an answer.
AI Should Usually Assist a Workflow Before It Runs One
Suppose your practice receives 50 new client inquiries in a month. You could give an AI system information about your clinicians — their specialties, insurance panels, schedules, populations, and current availability — and ask it to identify likely matches. I think that could be genuinely useful. I would still want a person reviewing the recommendation.
The reason is fairly simple: the system may be working with incomplete information. A clinician's availability might have changed yesterday. The client's presenting concern may include an important nuance. A therapist might technically work with a particular issue but not actually be the best fit for this person.
AI systems are very good at working with the information they have. They are much less reliable at recognizing the importance of information they do not have. That becomes a problem when a tool performs well often enough that people gradually stop checking it.
For low-risk administrative work, some amount of automation may be perfectly reasonable. As the consequences of a mistake increase, I want more human judgment in the process. For a lot of therapy-practice workflows, the design I prefer is pretty straightforward: AI handles the first pass, a person makes or confirms the decision, and the system records what happened. Over time, you can automate more if the evidence supports it.
That may sound less impressive than promising a fully automated AI practice. It is also much more likely to work.
Clinical Documentation Is Useful — and Complicated
Documentation is probably the most obvious AI use case for therapists. There is a good reason for that. Writing notes takes time, most clinicians do not particularly enjoy doing it, and a large part of the task involves organizing information that is already known into a consistent format. AI can help organize rough notes, improve clarity, reduce repetitive writing, and make documentation more consistent.
But this is also an area where practice owners need to be careful.
Clinical documentation contains protected health information. If identifiable client information is going into an AI product, you need to know what product you are actually using, how that information is handled, whether the vendor will enter into an appropriate Business Associate Agreement when required, and whether the workflow itself fits your practice's privacy and security obligations. AI does not create an exemption from HIPAA. And a vendor describing its product as "secure" is not enough information to make that determination.
There is another problem here that gets less attention: AI can generate language that sounds clinically appropriate even when it is inaccurate. A bad sentence often looks like a bad sentence. A confidently generated clinical statement can look completely reasonable.
The clinician still has to review the note and make sure it accurately reflects what happened in the session, the interventions actually used, the client's presentation, and the clinician's own judgment. AI can help write the note. It did not conduct the session.
Where I Would Be Much More Cautious
There are parts of clinical work where efficiency should not be the main objective.
Diagnosis is one. Risk assessment is another. Treatment decisions, clinical interpretation, crisis response, and decisions about appropriate levels of care all require professional judgment and accountability.
AI may still have a role around those decisions. It can help organize information, surface something worth considering, or help a clinician think through possible explanations. That is very different from handing the decision over to the system.
This distinction matters because AI is persuasive. It gives complete answers. It explains itself. It often sounds confident even when the underlying answer is uncertain. Clinicians are trained to evaluate context, history, behavior, conflicting information, and subtle changes in presentation. More importantly, we are responsible for the decisions we make. The software is not. That should affect where we place it in the workflow.
HIPAA Is Not the Only Risk
Privacy understandably gets most of the attention when therapists talk about AI, but it is not the only operational risk.
Dependence is another. If an important workflow only functions because one AI product behaves a particular way, what happens when the vendor changes the product, the pricing, or the feature you built around? Accuracy matters too — who is responsible for noticing when the system gets something wrong? So does accountability: who reviews the output before it affects a client, clinician, payment, or business decision?
And then there is a simpler question that often gets skipped entirely: should this process exist in its current form? I have seen plenty of technology added to workflows that were already unnecessarily complicated. The software did not fix the underlying problem. It just made the bad process faster and harder to understand. AI can do exactly the same thing.
Before I automate something, I want to understand why the task exists, whether it should exist, and whether the underlying process works. Sometimes removing three unnecessary steps is more valuable than automating all three of them.
The Standard Should Be Useful, Not Impressive
What I've noticed is that the practices using AI well tend not to talk about it much. They found a few places where it saves real time, built simple workflows around those, and kept humans in the decisions that matter.
Sometimes AI belongs inside a good system. Sometimes a spreadsheet, a template, a checklist, or a competent employee is the better tool.
There is still a significant opportunity here. Group practices generate enormous amounts of repetitive information work through inquiries, scheduling, documentation, billing, credentialing, supervision, communication, reporting, and administration. Used carefully, AI can reduce some of that burden and make good systems easier to operate.
I don't think the goal should be removing humans from the practice. The better goal is removing unnecessary cognitive and administrative work from people so they have more time for the work that actually requires human judgment.
So when I evaluate an AI tool or workflow, I start with the operational problem. I look at the information involved and what happens if the system gets something wrong. I decide where human judgment still needs to remain, and I make sure the privacy and security requirements are actually being addressed.
The goal is not to have an AI-powered practice. The goal is to have a practice that works — and AI may be one of several things that helps with that.