By Stuart Kerr, Technology Correspondent, LiveAIWire
The U.S. market for AI in physical therapy was valued at 178 million dollars in 2025 and is projected to reach 1.07 billion dollars by 2033, growing at a compound annual rate of over 25 percent a year, according to Grand View Research’s 2026 market analysis. That growth is being driven by a convergence of pressures that are not going away: an ageing population generating more rehabilitation demand, a chronic shortage of physiotherapy workforce capacity, and a persistent problem that has outlasted every previous generation of clinical technology, patients who are prescribed home exercise programmes and do not consistently do them.
The evidence base for specific AI physiotherapy applications is heterogeneous in ways that market growth figures alone do not capture. Some applications, AI motion analysis for exercise assessment, AI adherence monitoring, AI risk stratification for patient allocation, have genuine clinical evidence behind them. Others are entering the market with preliminary data that has not yet been tested in a rigorous trial. Understanding which is which matters for patients, clinicians and health systems making procurement decisions.
What Computer Vision Is Doing for Movement Analysis
The most technically mature application of AI in physiotherapy is computer vision based movement analysis. Systems trained on large datasets of annotated movement can now assess exercise form through a standard smartphone camera, identifying whether a squat is loading the knee correctly, whether a shoulder exercise is being compensated for with trunk movement, or whether a walking pattern shows an asymmetry worth flagging to a clinician, all without the patient attending a clinic.
A 2025 to 2026 systematic review of digital and intelligent rehabilitation technologies in stroke and neurological disorders, published on PubMed Central, examined 22 studies combining AI adaptive tools, wearable sensors, virtual reality and telerehabilitation platforms. Clinical studies in the review reported improvements in motor function, balance, gait and cognition, generally alongside high usability and adherence scores. The review’s authors were careful to note that many of the strongest effects came from technology-assisted training that increased task-specific practice, rather than from the AI component in isolation, a distinction that matters when evaluating specific product claims rather than the category as a whole.
The underlying mechanism is straightforward once the technology stack is unpacked. A camera captures a video stream, a pose-estimation model maps that stream onto a skeletal representation of the patient’s joints, and a classifier trained on thousands of labelled repetitions compares the patient’s movement against a reference pattern for the exercise being performed. The output is not a single pass or fail judgement. It is a set of measurements, joint angle at peak flexion, symmetry between left and right limb, tempo of the repetition, that a clinician can review alongside the patient’s self-reported pain and function.
Wearable inertial sensors, worn on the wrist, ankle or trunk, extend this capability outside the camera’s field of view, tracking gait and balance continuously through a patient’s day rather than during a scheduled assessment window. The combination of continuous wearable data and periodic camera-based assessment is where several of the more methodologically rigorous studies in the PMC review concentrated their strongest findings, because triangulating the two data sources helps rule out measurement artefacts that either sensor type would produce alone.
The Adherence Problem AI Is Trying to Solve
Non-adherence to prescribed home exercise programmes is the most persistent problem in outpatient rehabilitation. Patients who understand their exercises and intend to do them still routinely fail to perform them at the frequency and intensity a physiotherapist prescribes, and multiple studies place adherence to standard home exercise programmes as low as 35 percent. That gap between what is prescribed and what actually happens explains a meaningful share of the variation in rehabilitation outcomes, because an exercise programme that is not performed cannot produce the adaptation it was designed to achieve.
A rapid review of AI based digital rehabilitation, published on PubMed Central in 2025, found that AI systems can measure adherence more objectively than the self-report methods most clinics still rely on, tracking the frequency and duration of a patient’s actual engagement with prescribed exercises rather than depending on recall that is prone to social desirability bias. The review also found that personalisation, tailoring the AI intervention to an individual’s specific characteristics, coping style and behavioural profile, was one of the more consistent predictors of improved adherence across the included studies, more consistent than the presence of AI monitoring alone.
In other words, the design of the adherence system matters as much as the underlying AI capability, a pattern that shows up across most AI health applications currently entering clinical use, as our coverage of AI’s broader role in healthcare has found.
Stroke Rehabilitation and the Neurological Evidence
Stroke rehabilitation is the neurological application with the most active AI evidence base right now. The mechanisms of motor recovery after stroke, activity-dependent neuroplasticity, are sensitive to the intensity and repetition of practice in ways that often exceed what a human therapist can deliver within standard session frequency. Robotic and AI-adaptive systems that adjust exercise difficulty in real time based on a patient’s measured performance, rather than delivering a fixed protocol, are where some of the more rigorous stroke rehabilitation evidence has concentrated, because the optimal challenge level for neurological recovery changes with each patient at each stage of recovery.
The PMC systematic review referenced above found this real-time adaptation to be one of the more reproducible features associated with better outcomes across the studies it covered, though the review’s authors were explicit that certainty of evidence varies considerably by outcome measure and that firm superiority claims over well-matched conventional therapy are not yet established for most impairment-level outcomes.
AI Triage and Clinical Decision Support
Beyond exercise and monitoring, AI is being used in physiotherapy for patient triage, identifying which patients need urgent individual assessment, which can be safely managed through group or self-directed pathways, and which are at higher risk of a poor outcome without early intervention. This category carries a different evidence standard than adherence monitoring, because the consequences of a triage error are different in kind from the consequences of a missed monitoring alert.
Several health systems have piloted AI-supported physiotherapy triage with results suggesting it can allocate the majority of patients appropriately, but the evidence here is earlier stage and more setting-specific than the evidence for computer vision motion analysis or adherence monitoring, and it should be treated accordingly by anyone evaluating a specific triage product.
What the Evidence Does Not Yet Support
The AI physiotherapy market includes products considerably ahead of their evidence base. Tools claiming to diagnose specific musculoskeletal pathology from movement analysis alone, without clinical examination, are operating in a regulatory grey area in most jurisdictions and have not been shown to match clinical assessment in head-to-head testing for most conditions. Tools marketed around identifying chronic pain mechanisms from movement patterns are, in several cases, making claims that the published evidence does not yet support.
The heterogeneity across this sector is exactly why clinicians and health systems need to evaluate specific products against specific claims rather than deferring to the category’s overall growth trajectory, a caution that applies as much to the wearable and smartphone health tools covered in our look at what your smartphone actually knows about your health as it does to physiotherapy specifically.
The Workforce and Capacity Case
The AI physiotherapy market’s growth is happening against a backdrop of physiotherapy workforce shortages across most high-income health systems, with vacancy rates that are expected to worsen as populations age faster than training pipelines can produce qualified physiotherapists. AI monitoring, triage and home exercise support tools are explicitly positioned as extensions of a limited workforce, allowing each physiotherapist to support more patients by delegating routine monitoring to AI while retaining clinical judgement for assessment and treatment planning.
This is the strongest practical case for AI physiotherapy, and it is where the economic incentive and the clinical incentive line up most cleanly. A department using AI to monitor home exercise compliance across 200 patients, with the AI flagging the minority who are not adhering for priority follow-up, is providing better targeted care than the same department attempting to monitor the same caseload through periodic phone calls. The AI is not replacing clinical judgement. It is making that judgement better targeted, which is the implementation model the evidence most clearly supports, and it echoes the workforce extension argument being made for AI in adjacent areas of care, including in elderly care, where the same demographic and staffing pressures are driving similar tools.
What Patients Should Know Before Using AI Rehab Tools
For patients considering or already using AI physiotherapy tools, the practical guidance follows fairly directly from the evidence. Applications with the strongest support, AI-guided exercise programmes with motion feedback for musculoskeletal conditions and AI adherence monitoring for post-surgical rehabilitation, are most valuable when they sit inside a care pathway that includes human physiotherapist oversight, not as standalone substitutes for professional care.
Tools marketed directly to consumers as alternatives to physiotherapy consultation, for pain diagnosis, for autonomous treatment planning, for managing complex neurological rehabilitation without professional involvement, are making claims that exceed the current evidence and carry a real risk for patients who delay appropriate assessment while relying on them. The most useful question to ask before adopting an AI physiotherapy tool is whether it has been validated in a population similar to your own condition and age, and whether it is designed to integrate with professional care or to substitute for it.
This is the same question worth asking of any AI health tool making a clinical claim, including the AI mental health chatbots examined in our analysis of whether AI therapy actually works, where the same distinction between integration and substitution turns out to be the one that predicts good outcomes.
Cost and reimbursement are worth understanding too, because they shape which tools are actually built for integration rather than for standalone consumer use. In several health systems, insurers and employers are now paying for AI physiotherapy platforms directly, often through remote therapeutic monitoring billing codes that reimburse clinicians for reviewing AI-collected exercise data between in-person visits.
Tools built around that reimbursement model have a structural incentive to connect their data back to a supervising clinician, because that connection is what the billing code requires. Tools sold as a direct consumer subscription with no clinician in the loop do not have that same incentive, and patients should treat the absence of clinician integration as a meaningful signal about how the product is likely to have been evaluated and validated, not simply a difference in business model.
Integration With Existing Healthcare Systems
The practical challenge of deploying AI physiotherapy tools at scale is not primarily technical, it is integration. Most healthcare systems run electronic health record and referral management platforms that were not designed with AI physiotherapy tools in mind, and many AI physiotherapy products are not designed to integrate with them either. The result is that a lot of deployments operate as isolated apps rather than as a connected part of a patient’s clinical pathway, which limits both their clinical usefulness and their ability to generate the outcome data that would strengthen the evidence base going forward.
There is a commercial case study worth noting here: a randomised controlled trial of one AI-guided musculoskeletal programme, evaluated across more than 1,200 low back pain patients over twelve months and reported in Grand View Research’s market analysis, found claims-verified medical costs 80 percent lower for patients using the app compared with standard care, a result that reflects what integrated, clinically supervised deployment can achieve when the tool is built to work alongside a care pathway rather than around it. Standalone deployments that generate engagement data without generating clinical outcome data may satisfy a product roadmap without answering the question that actually matters for patients and health systems, whether the tool is working.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.