A few years ago, if you asked most clinicians what they wanted AI to do, ‘write my notes’ would have been pretty high on the list. Which is fair enough; nobody became a physiotherapist because they had a deep passion for clinical documentation. AI scribes have become remarkably good at solving that problem. They listen to our consultations, turn conversations into structured notes and draft letters before we’ve finished our coffee. But after more than a decade treating patients and running clinics, I’ve come to think we’ve aimed AI at the easiest part of the problem. Whilst notes are frustrating, the bigger admin problem is everything required to deliver good care between appointments. The invisible workload of good care Think about what happens after a fairly normal MSK consultation. You might need to update the patient’s exercise program, film a new exercise, find the right video, and write down the sets, reps and loading parameters. You might want to send the patient a summary of what you think is going on, remind them of their goals and explain what progress should look like over the next few weeks. Perhaps there’s an insurer form that needs completing before another block of treatment is approved. Then there’s the patient you haven’t seen for three weeks, who you were supposed to book again in seven days. Should you message them? Has reception already tried? What were their goals again? Did they stop coming because they were better? Because they weren’t improving? Because life got busy? Or because they forgot why the next appointment mattered? None of these jobs are particularly difficult, but multiply them by twelve patients a day, five days a week, and suddenly ‘good patient care’ has generated a second job. That’s the opportunity in AI that interests me most. Not simply making the documentation of care faster, but making the delivery of good care easier. We started with clinician-centred AI The first wave of healthcare AI has understandably been clinician-centred. Documentation is repetitive, expensive and unpopular. So we built machines that could do more of it for us. Whilst that’s a win, faster notes don’t necessarily mean better care. If an AI scribe saves me 30 minutes at the end of the day, I’ve improved my working life. If AI helps my patient understand their diagnosis, remember their exercises, see their progress and stay engaged with their rehabilitation, we may have also improved their life too. All practitioners want to deliver optimal patient care, but we’re often limited in what’s achievable by the number of hours in the day. This is where AI for physiotherapy starts to become much more interesting. Exercise adherence in MSK care is hardly a solved problem. Recent reviews continue to describe uptake and adherence to exercise-based rehabilitation as suboptimal [1]. We also know that adherence isn’t simply a matter of telling someone to ‘do their exercises’. Self-efficacy, social support, goal setting, instruction and demonstration are all relevant [2]. Digital interventions aren’t a magic wand either. A systematic review and meta-analysis found that digital rehabilitation improved therapeutic exercise adherence at intermediate follow-up, but not consistently at short- or long-term follow-up [3]. So the lesson clearly isn’t to throw more technology at patients, but rather to make the care we’re already trying to provide more individualised, consistent and easier to act on. The gap between knowing and doing Most clinicians already know what good care looks like. We know patients need clarity about their condition, we understand the importance of regular goal setting, and we know progress should be measured and visible. We’re also clear on the benefits of inter-appointment check-ins and the importance of following up with patients who have dropped out of care prematurely. But there is often a gap between knowledge and execution. A busy clinician can genuinely believe in all of those things and still finish a Tuesday afternoon with three exercise programs to update, two insurer forms to complete and a patient from last week they meant to follow up. What does patient-centred AI actually look like? Imagine finishing a consultation and, before your patient has walked out of the clinic, they receive a message with an up-to-date treatment plan. Not a generic post-appointment email, but their plan in a living portal. It explains the working diagnosis in understandable language. Their goals and objective measures are there. Progress since their initial assessment is visible. Their exercise program reflects what you discussed five minutes earlier and their recommended appointment schedule is clear. Then you see them again next week and rather than recreating the document, the new consultation updates it. Their shoulder flexion has moved from 120 to 150 degrees? Progress measure automatically updates. They’re back to swimming twice a week? The goal updates. You’ve progressed their external rotation exercise? The rehab program automatically progresses with their newly prescribed exercise. The treatment plan becomes a living representation of their rehabilitation rather than a PDF that was accurate for approximately seven minutes after their initial assessment. Another example is exercise prescription. Historically, creating a genuinely individualised program has been surprisingly fiddly. Search a library, find something close enough, add instructions, change dosage and maybe even film the patient on their phone if you can’t find what you’re looking for (with no record of the video for the practitioner). Now imagine the exercise content being created from what the practitioner actually says during the appointment. A suitable video can be automatically matched. An AI-generated image can demonstrate an unusual exercise that doesn’t exist in the library. Or, better still, you can film the patient performing their own exercise correctly and add it to their program in real time. The clinician still makes the clinical decisions, but AI removes the effort required to turn that decision into something useful and accessible for the patient. What about when the patient disappears? This might be the part I’m most interested in and excited by. We talk a lot about exercise adherence in physiotherapy, but we talk less about treatment-plan adherence. Imagine a patient presents with a problem that we reasonably expect will require a period of rehabilitation. Together we establish some goals and start treatment. Then, somewhere between ‘feeling a bit better’ and actually rebuilding the capacity required for their goal, they disappear. Historically, our systems haven’t been particularly sophisticated at handling this. Maybe reception runs a recall list, or the practitioner notices an empty space in the diary and remembers them. Maybe nobody does. What if the software already understood the plan? It knows the patient was recommended to return in seven days and it knows they haven’t booked. It knows their goal was to get back to running 10 kilometres, but only progressed to three kilometres at their last review. That creates the possibility of a very different automated message to patients: Not, “Hi John, you’re due for an appointment.“ But something closer to, “Hi John. At your last appointment you’d built your running back to 3km and we’re working towards your goal of 10km. Your next review was planned for this week so we can reassess your progress and adjust your loading. Tom has a free spot tomorrow at 3pm, would you like to book that?“ Add a one-click booking option based on their usual clinic, practitioner and appointment preferences, and suddenly following up patients doesn’t require the clinician to spend their lunch break trawling through a recall list. And importantly, the purpose isn’t to squeeze another appointment out of someone who doesn’t need one. If the patient has reached their goal, brilliant. The purpose is to reduce the number who accidentally fall off a plan they haven’t finished. Horizontal AI versus vertical AI This is where I believe healthcare AI will increasingly diverge. A general medical scribe has an extraordinarily broad job, such as understanding conversations across different professions, specialties, conditions and workflows. But an MSK clinician doesn’t just need software that understands general medicine. We need software that specifically understands MSK care, both in terms of content and practitioner workflow. That’s the difference between horizontal and vertical AI. A horizontal AI system might understand that I said “three sets of eight split squats” and document it correctly. Whereas AI built for the vertical of physiotherapy would automatically convert that into rehab videos or images and deliver it to a patient’s phone and track daily progress. It would understand that today’s objective measures may represent progress against measures taken four weeks ago, and track those improvements over time in a way that’s visible for both the practitioner and the patient. And it understands that the patient’s goals, treatment plan, exercises, outcome measures, appointment recommendations and clinical record are a continuum or relevant context across appointments rather than six unrelated pieces of information. This has been a major lesson for me while building Preve. Full disclosure: my interest in this isn’t entirely academic. I’m a physiotherapist and clinic owner, and these frustrations were a large part of why we built Preve in the first place. From day one, it was about building a tool that was centred around better patient care, whilst also doing the heavy lifting for the practitioner. What happens in the real world? We’re now starting to see what happens when AI is applied to the patient journey rather than just the clinical record. In an internal study, across hundreds o
