A Thursday at 17:40 in a four-physio practice in Pinelands. The last patient of the day is on the bed in room two getting dry-needled. The receptionist went home at 16:30. The principal physio's phone is buzzing with WhatsApps — a knee patient asking if she can come tomorrow instead of Monday, a post-op shoulder confirming his slot, a mom asking whether her son's GEMS submission has been paid out. The owner is also halfway through writing up notes for the seven patients he saw today, because that part of the job has to happen after the patient has left. By the time the notes are done, the WhatsApp queue has fifteen messages on it and Discovery wants a response on six claim line-items from a Wednesday rejection.
That is the rhythm of an independent SA physiotherapy practice. The clinical work pays. What erodes the practice is the layer wrapped around it: bookings, rebookings, notes, claim follow-ups, second-guessing whether last week's hamstring patient needs an authorisation extension. Almost none of it requires a physio. All of it currently lands on the physio.
There is a sensible AI conversation to have here. It is not about replacing the practitioner. It is about pulling the predictable, structured, repetitive layer out of their evening.
Where SA physiotherapy and biokinetics practices actually lose time
Across the single-practitioner studios, the three-to-six-physio independents, and the biokinetics practices attached to running clubs and orthopaedic surgeons, the same handful of items dominates the non-clinical workload:
- Booking and rebooking on WhatsApp — the channel patients actually use, with the usual mess of unclear preferred times and chronic last-minute moves
- Session notes and SOAP write-ups — the HPCSA expects them, the medical aid expects them, and most physios are doing them after hours
- Medical-aid authorisations — chasing additional sessions on Discovery, Bonitas, Bestmed, GEMS, KeyHealth, Polmed and Profmed before the patient walks in for visit eight
- Claim rejections through Healthbridge or GoodX — the ICD-10 mismatch, the modality code the scheme reads as out-of-scope, the tariff band entered wrong
- No-show and late-cancellation management — a 15% no-show rate at R600 a session is real money walking out the door
- COIDA and RAF claims — slow, document-heavy, time-consuming for any practice seeing post-MVA or workplace-injury patients
- Exercise programme adherence for biokineticists — the patient does six sessions in-clinic and is meant to keep going at home for eight weeks; most stop after two
Some of this is clinical judgement. Almost none of it is, once you look closely.
Bookings and rebookings: the cheapest win in any practice
The patients in an SA physiotherapy practice want to book on WhatsApp. They do not want to phone. They do not want to use a portal. They want to send a message that says "can I come Tuesday afternoon" and have the booking land in the diary the same evening.
A booking handler sitting on the practice's WhatsApp Business number reads the message, pulls the available slots from whichever system the practice runs (Elixir Live, GoodX, Healthbridge BookingBridge), and offers two or three options that fit. Confirmation sets the slot and sends the reminder twenty-four hours and two hours ahead. A reschedule is the same flow in reverse.
The honest gain here is not the saved receptionist minutes, although those add up. The gain is the no-show rate. A practice running on WhatsApp confirmations with a twenty-four-hour and two-hour nudge typically cuts cancellations and no-shows by four to seven percentage points within a quarter. At a hundred sessions a week, that is meaningful.
One caveat. The handler has to know which patients are post-surgical, post-acute, or otherwise need a real human reading their request. The principal does not want the bot rebooking a shoulder-reconstruction patient three days post-op without a physio's eye on it. Build the flag into the patient file from day one and route flagged requests to the physio instead of resolving them in the bot.
Session notes: where AI helps, and where it doesn't
This is the part most physios secretly want help with, and it is also the part where the most damage gets done if it is rushed.
SOAP notes are the standard format the HPCSA expects. They get written either during the session (which most patients dislike) or after the patient has left (which most physios dislike, because that is what makes the day fifty minutes longer than it should be).
A practical AI layer does not write the note. The note is the physio's clinical record. It is also their legal record. What the layer does is convert a short spoken voice memo into a structured first draft. The physio dictates "Mrs Naidoo, day fourteen post-op right TKR, ROM 95 degrees flexion from 85 last week, palpable warmth reduced, gait with single crutch, no antalgic pattern in the corridor, continuing with the home programme, see again Friday" and the system produces a SOAP-shaped draft with the right sections populated. The physio reviews, edits, signs.
Two things make this work in an SA practice. The transcription model has to handle the accents the practice actually sees, and most off-the-shelf tools still butcher South African English, especially for non-mother-tongue speakers. The second is that the assessment and plan sections must leave the physio's clinical reasoning visible, not paraphrased. If the model rewords "patient demonstrating delayed quadriceps activation" into something softer, the note is worse than no help at all.
In my experience the practices that get value here use the layer purely as a stenographer. The ones who try to skip the review step end up with notes that read fine until the medical-aid auditor reads them, at which point they read like they could belong to any patient on the books.
Medical-aid follow-up and the BHF switch
Every physiotherapist in this country has lost a Friday afternoon to medical-aid follow-up. A claim went in on Monday, came back rejected by Wednesday, the rejection reason was ambiguous, and now the practice is on hold to Discovery while a patient is on the bed.
A structured layer that watches the Healthbridge or GoodX feed picks up rejections automatically, reads the rejection code, and surfaces them with the likely reason and the next step in plain language. ICD-10 mismatch where 075.81 should have been 075.82. Modality code rejected because the scheme requires authorisation past visit six and the auth was not requested. Tariff billed at the wrong rate band. The system does not refile. That decision still belongs to the physio or the practice manager. What it does is lay out the response and pre-fill the resubmission, which is most of the work.
A second use is authorisation tracking. The principal at a busy musculoskeletal practice has fifteen patients across various schemes who will reach their plan limit inside two weeks. The layer knows which ones, drafts the motivation against the clinical notes already in the file, and queues it for the physio to review and send. The patient finds out about the extension before their session eight, not after a rejected claim two weeks later.
What the layer must not do: file authorisation requests without the physio's review, generate clinical motivations that go beyond what the notes support, or commit the practice to a tariff structure with the scheme. These are professional decisions with regulatory weight. The BHF, the schemes, and the HPCSA care that a real practitioner is behind the submission.
Biokineticists specifically: the home-programme problem
Biokinetics has a structural problem pure physiotherapy does not. The six clinical sessions are the easy part. The next eight weeks at home are where the patient is meant to do the work, and most do not. The biokineticist gets a returning patient at three-month review who has not done a single set of the prescribed programme since session four.
Light, sensible AI nudging earns its place here. After session three, the patient gets a WhatsApp every other day with the day's prescribed exercises: the actual ones the biokineticist put on the programme, in plain language, with a single short video clip per exercise from the practice's own library. A check-in question on Sundays asks how the week went. The patient sends back "knee was sore on Wednesday" and the message routes to the biokineticist's review queue, not to the bot's discretion.
Two boundaries. Programme content stays the biokineticist's. The bot delivers, it does not design. Anything that smells like a clinical concern (a flare-up, a fall, a new pain) routes to a human within the same hour. Get those right and the rest is sensible nudging.
Patients who complete their programmes refer. Patients who refer are the practice's growth engine. The arithmetic is straightforward.
Where AI should not be in your practice
The HPCSA's Professional Board for Physiotherapy, Podiatry and Biokinetics will care most about what gets delegated to a system without a practitioner in the loop. From what I have seen in implementations, the line sits at roughly five places.
Initial clinical assessment. The first visit is the practitioner's. A system cannot triage a new shoulder complaint into a treatment plan. It can take the demographic intake, the medical-history form and the consent. The clinical decision is the physio's.
Diagnosis and treatment-plan formulation. A model that suggests "this looks like a frozen shoulder, here is a six-week protocol" is exactly the kind of pitch to walk away from. Treatment plans live with the practitioner whose registration number is on the file.
Children and elderly patients under POPIA. Section 32 of POPIA covers children's personal information; older patients with cognitive impairment carry their own considerations. Any automated flow touching those files needs explicit guardian or family consent, captured in writing, tied to the patient record.
Manual-therapy notes generated without the physio's voice memo. A model that fabricates manual-therapy specifics it did not witness is a complaint waiting to happen. The voice memo is the source. The system has nothing to invent.
Cross-scheme price quoting. Do not let the system commit to private-pay top-up amounts or guarantee that a specific scheme will pay a specific tariff. Those conversations are the practice manager's, not the bot's.
A sensible first project
Pick the booking-and-no-show layer first. It is the lowest-risk, highest-frequency win, and the impact is visible inside thirty days. The diary fills more evenly, cancellations get filled, the principal stops being the one chasing rebookings on WhatsApp at 21:00.
After that, the next sensible step depends on what hurts most. For practices losing Fridays to claim rejections, the medical-aid follow-up layer pays for itself in one rejection cycle. For practices where notes are eating the practitioners' evenings, the dictation-to-draft layer matters more. For biokineticists where adherence is the bottleneck, the home-programme nudge layer changes the conversation entirely.
Nothing here replaces a physio. Nothing should. What an honest AI layer does in a physiotherapy or biokinetics practice is pull the structured, repeatable, predictable workload out of the practitioner's evening, so the principal can go home at six and the notes are done at six-fifteen, not nine-thirty.
That hour and a half is the entire pitch.