The Autonomous Clinic: Navigating AI, Sensors and the Equity Gap in Modern Physiotherapy

Musculoskeletal disorders account for a substantial share of global years lived with disability, and physiotherapy services in most health systems operate under demand that exceeds available capacity. Artificial intelligence has been proposed as a means of sharpening diagnostic judgement, structuring clinical reasoning and individualising exercise prescription. These three functions have developed along largely separate research trajectories, and they have rarely been appraised together against the specific epistemic and practical requirements of musculoskeletal physiotherapy. This critical narrative review evaluates the strength, consistency and translational maturity of the evidence across the three domains, drawing on literature retrieved from open scholarly databases and indexes and appraised for methodological adequacy rather than catalogued descriptively. The strongest evidence concerns image-based detection tasks, where deep learning systems achieve accuracy comparable with experienced readers and, when used as an assistive adjunct, improve human sensitivity without lengthening reading time. Evidence weakens considerably as the task moves towards the questions physiotherapists actually ask. Prognostic and triage models are typically developed in single cohorts, are seldom externally validated, and have almost never been tested for their effect on patient outcomes. Generative language models display case-dependent agreement with expert diagnoses under constrained conditions, alongside measurable rates of guideline omission and fabricated citation. Sensing technologies for exercise monitoring have advanced rapidly, yet joint-angle errors from markerless video remain large relative to the differences clinicians attempt to detect, and adaptive dosing algorithms have been evaluated predominantly in simulation or in healthy volunteers rather than in symptomatic populations. Across all three domains, the dominant limitation is not algorithmic performance but the absence of prospective, clinically anchored evaluation. Progress will depend less on further accuracy gains than on external validation, transparent reporting, and trials in which the outcome of interest is a change in patient function rather than a change in model discrimination.