It is Sunday at 20:10. A clinical psychologist with a small practice off Belvedere Road in Claremont is opening her laptop for the third time today. Two cancellations have come in on WhatsApp for Tuesday morning, both inside the twenty-four-hour window. She is allowed to bill the fee. She will not. A new client is coming in tomorrow at nine and the intake pack still has not gone out. A Discovery rejection is sitting in her email because the ICD-10 was filed as F32.9 when the scheme wanted F33.1. The supervisor letter for the intern she signs off on is due to the HPCSA on Friday and is half-written.
None of this is therapy. All of it is the price of doing the work.
There is an AI conversation worth having for South African psychology and therapy practices. There is also a louder, sillier one being sold on LinkedIn: the "AI therapist", the "AI co-therapist", the model that "supports clinical decision-making" in real time. Most of what is being pitched is either a complaint to the HPCSA Professional Board for Psychology waiting to happen, or a POPIA Section 26 breach, or both. The honest scope is narrower, and usually enough.
Where the hours actually go in an SA therapy practice
Walk into the diary of a generalist clinical or counselling psychologist running a private cash and medical-aid mix in Cape Town, Johannesburg or Durban, and the same admin shows up. Some of it is annoying. Some of it is genuinely costly, because every cancellation slot is also a clinical decision (do I bill, do I phone, was there something going on?). The recurring pieces:
- Booking, rebooking and the cancellation tax — WhatsApp is the channel everyone uses, and the cancellation conversation is the one most therapists handle badly because they hate the money part
- Intake paperwork — informed consent including the limits of confidentiality, demographic and history form, fee agreement, in some practices a baseline PHQ-9, GAD-7 or K10 screen
- Medical-aid claims and PMB code traps — F32 versus F33, the modality code for psychotherapy, the difference between the in-hospital and out-of-hospital tariff, the chronic application for major depressive disorder
- Authorisation chasing — Discovery for sessions thirteen onwards, GEMS for PMB cover, Bonitas for behavioural-health out-of-network
- Referral letters in and out — to a psychiatrist for medication review, to a GP, to a school counsellor with the parents' consent
- HPCSA admin — annual fees, CPD points, the supervision logs early-career colleagues need signed off
- Invoicing and tax — SnapScan and Yoco reconciliation, the monthly run for the bookkeeper, the provisional submission
None of this is the part of the job the work pays for. All of it makes the work possible.
Booking, cancellations, and the fee no one wants to charge
A WhatsApp booking handler on the practice number is the lowest-risk project in any therapy practice. The client messages on Sunday night. The handler offers two slots from the diary the therapist has marked as bookable, sets the slot on confirmation, and queues reminders twenty-four and two hours ahead. A reschedule runs the same path in reverse.
What changes the numbers is not the booking. It is the cancellation flow. The pattern I see most often: the therapist quietly absorbs the cost because charging feels punitive and the patient is already struggling. Over a year, in a practice billing R1,400 a session, that is real money. Usually one to two sessions a week walking out the door unbilled.
A handler can do the awkward part. The reminder at twenty-four hours states the fee policy in the patient's own language, not as a threat but as a fact ("if cancelled inside 24 hours the session is billed; if you need to reschedule, reply RESCHEDULE"). The patient who genuinely cannot make it reschedules earlier. The silent no-show gets the invoice. The therapist is no longer the one sending it.
One firm caveat. The handler must know which patients are in active risk management, recently discharged from inpatient care, or working through acute trauma. None of them should get a chirpy automated cancellation reminder. Build the flag into the patient record on day one. Flagged messages route to the therapist.
Intake and risk screening: the genuine sweet spot
This is where AI quietly earns its place in a therapy practice, because the intake layer is structured, it is repetitive, and it is the part most therapists end up rushing before the first session.
A sensible intake flow sends the new client an informed-consent document that names, clearly and in their language, the limits of confidentiality under South African law. The duty to report under the Children's Act if a minor is at risk. The duty to disclose under the Mental Health Care Act 17 of 2002 where there is a serious threat to self or others. The fact that medical-aid claims contain ICD-10 codes that disclose the broad diagnostic category to the scheme. None of this is news to the therapist, but most clients arrive having read none of it.
The same flow can collect the demographic and history form, capture the fee agreement, and run the standard screening instruments the practice already uses: PHQ-9, GAD-7, K10, where licensed and appropriate. The therapist walks into the first session with a summary on screen and the consent record in the file. The first fifteen minutes of session one are no longer spent on paperwork.
What this layer must not do: triage. A client in crisis on the intake form does not get a chatbot reply. They get a calm message that the therapist will phone within two hours, the SADAG helpline (0800 567 567, 0800 21 22 23 for the suicide line), and a route to emergency services. The system flags the file and pings the therapist. It does not assess risk.
Why session notes are different from physio SOAP notes
This is the section that most needs to be read twice.
A physiotherapist can dictate a thirty-second voice memo after a session and let a transcription layer produce a SOAP-shaped draft. Risk low, privacy stakes manageable. A therapy session is a different object. The content is not "patient performed left shoulder external rotation against red theraband"; it is the patient's marriage, their relationship with their father, the suicidal ideation they have not told their husband about. Recording it, transcribing it, or feeding it to a third-party model is a category of risk most therapists are right to refuse.
What AI can do, carefully, is help the therapist clean up the notes they were going to write anyway. The therapist types or speaks a short summary into their own device, without naming the client. The layer turns it into a structured progress note in the practice's format. No cloud transcription of the session. No model "listening". No data leaving the therapist's controlled environment without a written processing agreement that holds up under POPIA Section 19 and Section 26.
In my experience, the therapists who get value here keep the recording question completely off the table. The shortcut is tempting and the trade-off is bad. Notes that read fluently but came from a session transcript stored on someone else's servers are a HPCSA complaint waiting to be lodged the first time a client subpoenas the file.
Medical aid, PMB codes and the chronic application
Mental-health claims into South African medical aids have their own grammar. A diagnosis of F32.1 (moderate depressive episode) bills differently from F33.1 (recurrent depressive disorder, currently moderate). Major depressive disorder, bipolar disorder, schizophrenia and a handful of others are listed as PMB conditions and the scheme is legally required to fund treatment to a defined minimum. The application still has to be lodged, with the right ICD-10 code, the diagnostic criteria summary, and the GP or psychiatrist's report attached.
An admin layer that watches the Healthbridge or GoodX feed picks up rejections, reads the code, and surfaces the likely fix in plain language. ICD-10 in the wrong sub-code. Authorisation requested past session twelve when Discovery wanted the application by session ten. PMB application missing the psychiatrist's confirmation. The layer does not refile. It drafts the resubmission, pre-fills what it can, and queues it for the therapist to send.
A second use is the chronic-application workflow. A patient with a confirmed major depressive disorder diagnosis qualifies for PMB funding the scheme is required to provide. The application is paperwork-heavy and most therapists put it off, which means the patient pays out of pocket for months they did not have to. A layer that flags likely PMB-qualifying patients, drafts the application against existing notes, and queues it for sign-off is sensible. The therapist still makes the diagnostic call. The system just stops the application from sitting in the drawer.
Where AI has no business being in a therapy practice
The Professional Board for Psychology cares most about scope-of-practice and the recording of confidential material. From implementations I have seen, the line sits at six places.
Doing therapy. No model does psychotherapy. Anything pitched as an "AI therapist", "AI counsellor" or "between-sessions clinical companion" is a scope-of-practice claim a registered psychologist cannot delegate. The HPCSA does not register software.
Live session transcription. Recording the therapy session itself, to any cloud service, for any purpose, with any "anonymisation", is a category of risk that almost never pays back what it costs. Do not do it.
Risk assessment. A new client writing "I have been thinking about how I would do it" on an intake form is a clinical event. The layer flags and routes to a human. It does not score, classify or reply.
Child and adolescent records. Section 26 of POPIA treats mental-health information as special personal information, and Section 35 adds children to the category. Any flow touching a minor's file needs the right consenting adult's explicit written consent, captured against the file, and a privacy posture you can defend to the Information Regulator without flinching.
Forensic and medico-legal work. If your practice does court-directed assessments, the evidentiary chain matters more than the time saved. Keep AI out of those reports until your supervisor and your medical-defence cover say otherwise in writing.
Anything that goes to a foreign jurisdiction without a Section 72 agreement. If the model or the storage sits outside South Africa, the transfer condition under POPIA Section 72 has to be actually satisfied, not assumed. Most off-the-shelf consumer tools do not satisfy it.
A sensible first project
Start with intake-and-booking. Lowest risk, fastest return, no contact with the clinical content of a session. The impact shows inside two weeks: new clients arrive with consent signed and screens completed, Sunday-night admin shrinks, the cancellation conversation runs without the therapist having to drive it.
After that, the next move depends on the practice. A solo cash practitioner gets the most from the invoice and reconciliation layer. A group practice with a heavy medical-aid book pays back the rejection layer inside one billing cycle. A practice supervising interns gets real value from a supervision-log and CPD-tracking layer.
None of this is the work. The work is the fifty minutes in the room.
What an honest AI layer does in a South African therapy practice is hand back the Sunday evening, the unbilled cancellation, the rejected claim, the intake the therapist was going to do tired at the start of session one.
For most practitioners I speak to, that is more than enough.