An AI assistant for therapists: what AI should not do in a practice
The ads for AI tools for therapy practices are getting bolder: "AI will understand the client", "AI will lead the conversation", "AI will save you the thinking". Our advice: do not believe such ads. Not even ours: check what sits behind the slogan. We build an AI assistant for therapists, and that is exactly why we start with what AI in a practice should not do. There are five limits.
1. AI does not run therapy
This is the most important limit and it cannot be crossed. Therapy is a relationship between a client and a professional, with training, supervision and professional responsibility. An AI assistant that "talks to the client about their problems" is not help for the practice, it is a risk for the client.
In MentAI this rule is written into the code, not only into declarations. An example: on a practice's booking page a client can handle things by voice, book an appointment, move it. But when the conversation turns to clinical or crisis topics, the system declines and points to the emergency number 112 and the 116 123 helpline. The assistant knows its place: the calendar, yes; the soul, no.
2. AI does not make decisions for the therapist
The second limit: no AI action happens "on its own". MentAI follows a rule we call human-approval internally: AI proposes, a human approves. A note draft after a session goes into the records only once the therapist has accepted it. Natural-language commands (the "Ask" feature: around twenty actions in Polish, from "issue an invoice" to "move the appointment") always end with a confirmation screen. Nothing enters a client's file without your click.
Does it slow things down? By seconds. In return, every entry in the records has a human author, and that is the right order of things.
3. AI does not need to know the content of sessions
The popular idea of an "AI note" is recording the whole session and sending it to the cloud. We believe that in a therapy practice this is the wrong default. MentAI works on a minimum principle: after the session the therapist writes down a few words (they can also dictate them), and AI expands them into a draft in the chosen approach: CBT, psychodynamic, systemic, humanistic, Gestalt or integrative. Drafts are processed ephemerally: they are not stored or logged.
How this relates to professional confidentiality and GDPR (the data processing agreement, the record of processing activities, the client's right to information) is covered separately in our article on AI-assisted notes and professional confidentiality.
Update, 20 August 2026: since August there is a deliberate exception to this limit: an optional session transcription module, in which AI does learn the content of the conversation, but only with the explicit consent of every participant, and the recording and the transcription are deleted as soon as the transcript is returned (a note draft is created only at the therapist's request). The limit does not disappear, it changes its guardian: from "technically impossible" to "only with consent, without a trace". Details in the updated article on professional confidentiality linked above.
4. AI does not take the data
The fourth limit concerns what happens to data "under the bonnet". In MentAI the whole AI and voice route, from dictation, through note drafts, to the voice reception, runs on infrastructure in the European Union, and practice data is not used to train models. On top of that, clinical records are a separate, optional module with encrypted storage and an access log; the reception never sees them.
5. AI does not pretend to be human in front of the client
When a client of the practice is dealing with a machine, they should know it. On the booking page the voice reception is labelled plainly as "Voice reception", and next to the form it says in black and white: nothing is booked automatically, the practice always confirms the appointment. Trust in a practice is built over years, and a machine posing as a person can ruin it in a single conversation. The limit is simple: the machine handles the matter and says it is a machine. It does not play a human.
So what is actually left for AI?
The paperwork. And that is exactly the part of the job no therapist dreamed about at university:
- A note draft after the session, from a few of your words, in your approach, with AI corrections per section; you approve.
- Documents: online GDPR consents, a data processing agreement and a record of processing activities generated from the panel.
- Appointments: SMS and e-mail reminders, and a slot proposal via WhatsApp to which the client replies "works for me" and the matter closes itself.
- Commands in plain language: "Ask" understands Polish and prepares the action for your approval.
- Dictation: it also works in Brave, Firefox and Safari, because not everyone works in Chrome.
AI takes the paperwork. You stay with the client.
That sentence comes from our advertising, and at the same time it is our whole product philosophy in nine words. If you ever catch us breaking any of the five limits above, remind us of this article. That is why we wrote it.
Your practice. Your decisions. AI only helps.
Frequently asked questions
Will an AI assistant replace a psychologist or psychotherapist?
No, and a well-designed assistant does not even try. Therapy is a relationship between a client and a professional. The job of an AI assistant is to take the administration off the therapist: note drafts, documents, reminders and bookings, not to hold a therapeutic conversation.
Are AI-assisted notes safe for professional confidentiality?
It depends on how the tool is built. In MentAI the draft is generated from a few words written by the therapist, it is processed ephemerally, it is not stored or logged, and only what the professional approves goes into the records. The legal details are covered in a separate article.
Do you have to use the AI features in MentAI?
No. The AI features are optional: notes can be kept entirely by hand, and every AI action requires approval by a human. Bookings, invoices, reminders and documents work regardless of whether you use AI.
Where is data processed when I use AI in MentAI?
The whole AI and voice route runs on infrastructure in the European Union, and practice data is not used to train AI models. AI drafts are not stored or logged.
See the limits in practice: the demo, no account needed
The product interface is in Polish. Prefer to start for real? Sign-up is open: 30 days of the full version, no card, at mentai.pl.
Related articles: AI-assisted notes, professional confidentiality and GDPR · Where the name MentAI comes from · Choosing practice-management software: a comparison.
New posts every week. Follow MentAI on Facebook or Instagram: one thing a week that takes paper off the practice's desk.
MentAI