Physical therapy is a hands-on profession, which is exactly why the noise around "AI in physical therapy" is so hard to read. Vendors demo motion-capture that grades a squat; clinicians go back to a caseload where the real constraint is documentation, not squat grading. Both things are true at once, and the gap between them is where most of the disappointment lives.

This is a practical look at what AI is genuinely doing in physical therapy today, where it stops, and how to evaluate a tool before it costs you a month.

Where AI actually helps today

Documentation and note-drafting. This is the least glamorous use and the one with the clearest return. A structured note drafted from a session summary, then corrected by the clinician, removes a large share of after-hours charting. The clinician still owns the note — the model just gets it to 80% so the remaining 20% is editing rather than composition.

Movement analysis from ordinary video. Pose estimation from a phone camera has become genuinely usable for gross measures: joint angles, symmetry, tempo, rep counting. It is good at consistency — the same measurement, taken the same way, week over week. That consistency is the value, not the precision of any single reading.

Home-exercise adherence. Most home programmes fail quietly. A model can nudge, count, and flag when a patient stops without needing the clinician to chase. Adherence tracking is a scheduling and messaging problem more than an AI problem, but the pattern-spotting on top of it is real.

Triage and intake. Structured intake questions, sorted and summarised before the first appointment, means the first ten minutes are not spent gathering history that a form could have collected.

Where it stops

Be clear-eyed about the limits, because this is where vendor demos get quiet.

It does not palpate. A large part of assessment is tactile and contextual. No camera captures tissue quality, guarding, or the patient's face when they shift position.

It generalises from populations you may not treat. A model trained largely on healthy adults performing clean movements will behave unpredictably on post-surgical, paediatric, geriatric, or neurological presentations. Ask any vendor what their training population looked like. A vendor who cannot answer that is telling you something.

It cannot carry clinical liability. A suggestion is not a decision. Every regulator that has addressed this treats the clinician as the decision-maker, and the documentation needs to reflect genuine clinical reasoning, not a pasted model output.

Patient data is the hard part. Health data carries real obligations — HIPAA, GDPR, or your local equivalent. "The vendor says it is secure" is not a compliance position. Where is the data processed, who can see it, what is retained, and can you get it deleted?

How to evaluate a tool without losing a month

A few questions cut through most sales conversations:

1. What does it remove from my week? If the answer is not measured in hours, it is a toy. Adding a tool that produces work is a net loss.
2. What happens when it is wrong? A good tool makes its errors obvious and cheap to correct. A bad one buries them in a plausible-looking note you sign.
3. Does it fit the workflow I already have, or does it demand a new one? Tools that require the whole clinic to change behaviour rarely survive contact with a busy Tuesday.
4. Can I export everything? If your notes and measurements are trapped, you are renting your own records.
5. Who is accountable? Name the person who reviews model output before it reaches a patient.

Run any tool on a narrow slice first — one clinician, one caseload, four weeks — and measure the thing you said you would measure. Most failed rollouts are failed evaluations.

A note on "specialist" AI

General-purpose chatbots are broad and shallow. They will answer a physical therapy question confidently and sometimes wrongly, because nothing in them distinguishes a well-supported protocol from something that appeared once on a forum. The useful pattern is narrower: an assistant scoped to a domain, built with input from people who actually practise it, and honest about the boundary of what it knows.

That is the idea behind Misar.AI — specialised assistants built with domain experts rather than one model asked to be everything. For a clinician, the practical difference is what happens at the edge of competence: a scoped assistant should say it does not know and point you back to the literature, instead of producing something fluent and unsupported. It is free to start, which is the right way to test whether a claim like that survives your own questions.

The honest summary

AI in physical therapy is currently strongest at the work around treatment — notes, intake, adherence, tracking — and weakest at treatment itself. That is not a disappointing conclusion. Administrative load is a leading contributor to clinician burnout, and a tool that gives back several hours a week is doing real work even if it never touches a patient.

Be suspicious of anything that promises to assess. Be interested in anything that promises to get you out of the clinic on time.

Integrating AI into Existing Documentation Workflows

AI‑generated note drafts can be slotted directly into your current electronic health record (EHR) by exporting structured templates in HL7 or FHIR formats. When the system populates key sections—chief complaint, assessment, plan—clinicians only need to review and correct, which keeps the clinician in the decision loop and satisfies liability requirements.
Clinicians should treat the AI output as a first‑draft, not a final product. By setting up a version‑control log that flags when the AI model was last updated, the practice can audit the source of any discrepancy and ensure that the documentation reflects genuine clinical reasoning.
Change management is critical: a brief onboarding session that walks staff through the “edit‑then‑sign” workflow, coupled with a quick reference guide, will reduce resistance. A small group of champions can gather feedback on the AI’s accuracy, allowing the vendor to fine‑tune prompts or train a custom model for your patient population.

Leveraging AI for Remote Monitoring and Telehealth

Phone‑camera pose estimation is now robust enough to provide real‑time joint‑angle feedback during a virtual session. By overlaying a live angle readout on the patient’s video feed, the clinician can correct movement patterns instantly, creating a loop of immediate reinforcement that is hard to achieve in a purely asynchronous video review.
For home‑exercise adherence, a lightweight mobile app can push daily reminders, record video snippets, and use machine learning to count repetitions and flag deviations. The app can then sync with the EHR, ensuring that adherence data is part of the patient’s longitudinal record without manual entry.
Security is non‑negotiable. All data—video, sensor readings, and text—must be encrypted in transit (TLS) and at rest. Consent forms should explicitly state that the data will be processed by a third‑party AI provider, and the practice must have a clear data‑deletion protocol that can be invoked at the patient’s request.

Future‑Proofing Your Practice with Modular AI Platforms

A modular architecture—where AI services are exposed via RESTful APIs—lets a practice swap or upgrade components without overhauling the entire system. For example, a new gait‑analysis module can be added without disrupting the existing note‑drafting service. Open standards such as HL7 FHIR and OpenAPI specifications make integration painless.
Continuous learning is essential. The vendor should provide a transparent model‑update schedule, and the practice should retain the right to audit the model’s decision logic. If a model is fine‑tuned on your own data, you maintain ownership of that data and can enforce stricter privacy controls.
When scaling across multiple clinics, a cost‑benefit model that tracks time saved per clinician, reduced documentation errors, and improved patient engagement can demonstrate ROI. A phased rollout—starting with a single clinic, measuring key metrics, and then expanding—minimizes risk while building a data‑driven case for broader adoption.

Key Takeaways

• Use AI to draft structured notes and cut charting time by roughly 80% of composition effort.
• Apply phone‑camera pose estimation for consistent joint‑angle and symmetry tracking; focus on repeatability rather than single‑read precision.
• Employ AI‑driven adherence nudges and rep counters to flag missed home exercises, treating it as a smart scheduling tool.
• Verify vendor training data and exclude solutions built only on healthy adult populations when treating post‑surgical, paediatric, geriatric, or neurological patients.
• Demand transparent data‑processing policies—know where data is stored, who can access it, and how it can be deleted—to meet HIPAA, GDPR, or local regulations.
• Pilot any AI tool on a single clinician’s caseload for four weeks, measure time saved and error visibility before full clinic rollout.

Originally published at https://www.misar.blog/@misar-ai/articles/ai-physical-therapy-what-it-can-actually-do