Where AI genuinely helps in therapy practice, the decisions that must never be handed over, the consent and KVKK side, and how to assess a feature.
The question therapists face has changed over the past few years. Nobody asks “will AI change therapy” any more; they ask “should I turn this feature on”. The difference matters, because the second is a concrete decision: it has a scope, a risk and — if it has been built well — a way back.
This piece is meant to produce neither enthusiasm nor alarm. There are uses that genuinely work in practice today, there are decisions that must never be handed over under any circumstances, and between the two there is a line every therapist has to draw for their own practice. What follows are the distinctions that help draw it.
In a therapy context, “AI” today mostly means one thing: large language models. These models work on text. They take one piece of text and produce another in whatever form they are asked for. They summarise, rephrase, fit something to a template, turn a list into prose or prose into a list. That is the whole of what they do — and it is both less than people assume and more useful than people assume.
Less, because when a model builds a sentence it does not know whether that sentence is true; it only knows whether it looks like it fits the context. Saying “I don’t know” is the behaviour a language model is weakest at.
More useful, because a considerable share of the time spent in clinical practice goes not into thinking but into writing. The note in your head has already formed; what is missing is the twenty minutes it would take to get it into a file. A tool that works on text touches exactly those twenty minutes.
Setting this distinction up front makes every later question easier. If the work a feature offers to do for you is “writing” or “formatting”, the ground is solid. If the work it offers to do is “deciding”, the ground is not solid.
Note debt is the quietest problem in a working practice. Six sessions happen in a day, the notes are put off with “I’ll write them tonight”, two get written in the evening and the other four carry over to the next day. A note written three days later describes less of the session than of what the therapist remembers of it. The quality of the record drops, but before that the therapist’s confidence in their own memory drops.
A language model’s job here is simple: turn the scattered input the therapist leaves during or just after a session — a few bullets, half-sentences, the empty fields of a template that suits their modality — into a readable draft. The draft is not an account that stands in for the session; it is a tidied version of what the therapist has already said.
Transcribing an audio recording falls into the same category, but it is one step more sensitive: now the client’s voice is involved. We will come to the consent side of that shortly; for now it is enough to say that transcription is not a convenience but a separate processing activity, and it has to be treated as one.
The good thing about this use is that its errors are visible. If the draft is wrong the therapist can tell while reading it, because they were the one in the session. The bad thing is this: when a draft reads fluently, it is easy to skim it rather than read it. We will come back to that too.
Scoring a measure is not a job for AI. The PHQ-9 total, the GAD-7 cut-off, the subscale breakdown of the DASS-21 — these are all rule-based calculations and they should stay rule-based. There is no clinical situation that calls for a language model to do arithmetic.
The model’s contribution is not in the calculation but in the telling of it. Turning a three-point series into prose, turning a subscale profile into a paragraph ready for supervision, pulling out a summary to attach to a referral letter — these are writing jobs. The therapist has already looked at the curve and knows what they see; what is missing is the time to write it down once more.
The one thing to watch here is whose interpretation it is. The fact that a model wrote the sentence “a meaningful reduction is observed in the client’s anxiety” does not mean the assessment belongs to the model, and it must not. The moment the sentence passes under the therapist’s eye and is accepted, it is theirs. If it has not, it should never have been in the file at all.
Alerts raised when a clinically meaningful threshold is crossed are not a job for a model either. Item nine of the PHQ-9 coming back non-zero is flagged by a rule; there is no good reason to leave that to a probabilistic model. The alert is not a decision, only a marker that makes visible an item that could be missed on a busy day.
What slows a practice down is usually not the clinical work but the correspondence around it. Appointment confirmations, the polite reminder about the cancellation policy, the information text sent before a first consultation, the written explanation of an exercise given to a client, the summary prepared for a case handed to a colleague, the draft of a formal letter to an institution.
What these texts have in common is this: the cost of an error is low and the correction is easy. A reminder email written in the wrong tone gets rewritten; a risk assessment written wrongly does not get rewritten. So this is the most sensible place to start trying AI — not the clinical record, but the correspondence around the clinic.
In a small team this makes a difference of a few hours a week. For a therapist working alone it usually means something more specific: not still writing emails at eight in the evening.
Knowing where a tool does not help is as useful as knowing where it does. Some of what is sold under an AI label in clinical software is really rule-based work, and it should stay that way.
Sending an appointment reminder is a scheduling job. Checking the calendar for a clash is a comparison job. Scoring a measure is arithmetic. Working out whether the cancellation window has closed is a date calculation. None of these has any business involving a probabilistic model; rule-based code gives the same answer every time, and a model does not.
This distinction has a practical consequence: when you see “AI-powered” on a product, it is worth asking what sits underneath it. If the answer is “a rule-based check”, that is good news. If there is no answer, there may be nothing there to describe.
A language model can match the symptoms described to it against diagnostic criteria. A match is not a diagnosis. Diagnosis requires duration, functional impairment, differential diagnosis, ruling out medical conditions, developmental history and cultural context. Most of that is read in the consulting room, and most often from what is not said.
A list of “possible diagnoses” produced by a model is the most clinically harmful form of output there is, because it stops thought in proportion to how right it looks. Rather than narrowing the differential, it puts a name on the table early, and that name quietly shapes the language of the sessions that follow.
A formulation is built where the framework of a modality meets the client’s own language. A chain analysis in CBT, a mode map in schema therapy, an avoidance loop in ACT — each of them holds on to the particular words the client used in session.
A model imitates the form of these structures very well. It produces a formulation that looks tidy, uses the terms correctly and reads fluently. The trouble is that the formulation usually belongs to a generic case rather than to the client in front of you. “Fluent and wrong” is the hardest kind of error to spot in a clinical text.
Rather than having the model write the formulation, having it read the formulation you wrote and asking “what is missing here” is a far more defensible use. In the first case you hand over the thinking; in the second you test it.
Risk assessment is not a classification problem. Suicidal ideation, self-harm, a threat to a third party, suspected neglect or abuse — in none of these does a label of “low”, “moderate” or “high” mean anything on its own. What gives it meaning is what follows the label: a safety plan, closer monitoring, contact with the family, a duty to report.
A model errs in both directions here. A false alarm desensitises the therapist to alerts. A missed signal cannot be taken back. In either case the responsibility belongs to the therapist in a way that no contract can shift. That is not a legal detail; it belongs to the definition of the profession.
The therapeutic alliance is built between two people. A client holding an unsupervised “therapy conversation” with a model between sessions does not support the work; it moves the work outside it. The therapist does not know what the model said, still less how the client heard it, and in the next session the room carries the trace of three voices rather than two.
This does not mean every digital contact between sessions is a problem. Psychoeducational material, an explanation of a homework task, breathing-exercise instructions that the therapist has prepared and approved — these are the therapist’s voice, only in writing. The distinction lies in who approved the text.
The error in a hand-written note usually takes the form of an omission. You do not write something you do not know; you leave out a detail you cannot remember. The note is incomplete, but there is nothing in it that was not there.
A language model’s error is the exact opposite: it arrives as an addition. A sentence that was never said, an event that never happened, a wrong date, the name of a measure that was never administered — all of it written as naturally as the rest of the text. The invented sentence is written in exactly the same register as the true one.
This is why how you read the draft matters. Reading it for “is this right” is not enough, because fluent text looks right. The question that works is: did I say this? For every concrete claim in the text — a date, a name, a number, a sentence attributed to the client — you need to be able to remember where it came from. If you cannot, that sentence does not go into the file.
The one product feature that makes this easier is output that carries its provenance: which inputs it was generated from, which measurement result it rests on, when it was created. A summary that does not state its source is a summary that cannot be checked.
Session content reaching a model provider in any form is, from the client’s side, a separate and additional processing activity. A single sentence buried in the general privacy notice does not cover it.
What the client actually needs to understand is a short list: what is being recorded (audio, or only the therapist’s notes), where that record goes, how long it is kept, who can see it, and what happens if they say no. The last item is the one most often skipped and the most decisive of all: refusing must not cost anything. A client who does not want the recording carries on receiving the same service.
A way of putting it that works usually sounds something like this: “I use a piece of software that helps me prepare my session notes. A written summary of what we talk about goes into that system, and I review and approve the note myself. If you’d rather I didn’t, I can turn it off, and it won’t affect our work in any way.” Those sentences fit into the first five minutes of a first session.
Consent is not a one-off either. A client may agree at the start and change their mind in the third month; they may ask for it to be switched off for a particular session — especially one where heavy material is going to come up. “Not today, please” has to be possible to say, and asking must not be a hassle.
On the product side, the counterpart of this is not a promise in prose but a technical constraint: with no consent record, the processing never starts. Asking a provider how they have built that tells you more than reading their privacy policy.
When recording or AI comes up, the questions clients ask are remarkably similar. Being prepared keeps that conversation from swallowing the session’s agenda.
“Is a computer reading what I say?” The honest answer is usually “yes, the text goes into a system”. Trying to soften it damages trust more badly later on. What should be added is who reads the text and who can get at it.
“Who can listen to the recording?” The answer should be concrete: you, and — in a multi-therapist setting — the people whose access has been defined. That reception staff cannot reach the clinical record is a boundary worth spelling out.
“Is this recording just sitting somewhere?” If you do not know the retention period, finding out and coming back is better than guessing.
“Is the AI saying something about me?” Being clear here matters: the text produced is not an assessment but a draft of the therapist’s note, and it goes nowhere until the therapist approves it.
None of these questions sits outside the work. More often than not they belong to a wider pattern about the client’s sense of control, and they can be taken up that way.
Under the KVKK, Turkey’s data-protection law, psychotherapy records are special categories of personal data within the meaning of Article 6. When an AI feature comes into play, that data reaches another actor — the model provider — and the chain gets longer. What needs clarifying:
For a fuller treatment of this topic, see Writing session notes; the questions there apply independently of AI.
Marketing copy is the source that tells you least about a feature. The questions below reveal in practice whether a provider has taken it seriously.
On top of these questions, the single test that works best on its own is this: for one week, write your notes as you normally would and keep the draft generated for the same sessions alongside them. Reading the difference tells you more than any sales demo.
The most common mistake when bringing AI into a practice is trying everything at once. A sturdier route:
Start with a single use. Preferably somewhere outside the clinical record — drafts of correspondence, or summaries of measures. Move on to note drafts once you know the interface and its limits.
Set a criterion. “Is it saving time” is a vague question. A more concrete one: is the gap between a session and the note being closed getting shorter, and is the quality of the closed notes holding up? You can only answer the second by comparing against your own older notes.
Say so in supervision. Mentioning it when you bring in a case whose note draft was prepared by a model lets your supervisor read the text with a different eye. It is not something to hide; it is methodological information.
Review the habit. The question to ask yourself after three months is this: am I still reading the drafts, or am I skimming them? The answer changes over time, and it usually changes for the worse. Noticing is the only remedy.
For a therapist working alone, using AI is a personal choice. In a clinic of three it is an organisational decision, and it does not work until it is written down.
What needs settling is a short list. Which roles may use the feature — reception should never enter the clinical record, so generating note drafts is not their job either. Who accepts an interpretation of a measure. Whether one therapist can generate a draft for another therapist’s client, and if so in what circumstances.
Underneath all of this the real issue is consistency. If, in the same clinic, one therapist edits drafts carefully and another accepts them as they stand, the quality of the records varies from person to person. That stays invisible for a while and surfaces at the first supervision session or the first handover.
Agreeing a shared minimum rule within the team is usually enough: which fields are always written by hand (risk, plan, discharge criteria), and where a draft may be used. Fitting that rule onto one page and talking it through in a team meeting does more good than a long policy document.
AI enters clinical practice as an instrument, not as a colleague. The good thing about an instrument is that it is predictable: you know what it does, you know where it stops, and you can tell when it breaks. The bad thing is that when it is poorly designed, none of that holds.
The practical upshot is plain: no sentence a model produces should enter a client’s file before you have read and accepted it; no clinical decision should rest on an output; and the client should know what is being done with their data before they have to ask you.
With those three rules in place, the rest is a matter of convenience — and there is nothing wrong with convenience as such. What is wrong is failing to notice the moment convenience takes the place of clinical responsibility.
Let's talk through how this would work in your own practice.Book an intro call
In about fifteen minutes we'll walk you through the product on a demo built around your practice, and answer whatever is on your mind.
30-day money-back guarantee.