Why the work stays close to the patient
Physical therapy is measured in hands and in minutes. A therapist watches a knee bend, feels how the tissue resists, and changes the plan in the same session. That loop of observe, touch, adjust is the center of the job. Software can watch. It cannot feel resistance through a joint, or ease off when a patient flinches.
Two tasks make the point. Examining a patient’s strength, balance and range of motion depends on physical contact and on judgment about pain the person may not describe well. Manual therapy, including joint mobilization and soft tissue work, needs trained hands and second-by-second feedback. Neither task has a clean digital version.
There is a third piece that gets less attention: getting the plan done. A home program only works if the patient repeats it through a painful first week after surgery. Coaxing, pacing and renegotiating goals with a worried family is part of the treatment, not paperwork around it. Our Still needs a human score for this job is 75 out of 100 (higher is safer), and how we score occupations explains where that figure comes from.
What AI does, what it assists, and what it leaves alone
The clearest wins are clerical. Tools already draft progress notes from dictation and clean up visit summaries, and they can pull billing codes and prior-authorization language out of a chart. The share of task time our model puts in the fully automatable group is 4%, and it is mostly records work rather than treatment.
A bigger slice is assisted work. Software can propose a starting home exercise program from a diagnosis and then let the therapist edit it. Camera and sensor systems can score gait or squat mechanics and flag changes between visits, which speeds up what a clinician would otherwise estimate by eye. Assisted tasks come to 32% of task time, and they still end with a person signing off.
The rest sits with people. Hands-on treatment of a post-surgical or neurological patient, in-room evaluation and reassessment, and supervising assistants and aides all stay human in our task split, at 64%. Coverage, our estimate of the task time AI can handle today, reads 23 out of 100, and the coverage method sets out how that share is built.
What the evidence shows so far
Not much has been tested head to head. Our quality-parity grade for this job is D, which means there is no published comparison of an AI system against a licensed physical therapist across the full job: examination, manual treatment, progression decisions and discharge. Because of that, we publish no parity number here. A grade with no number is a statement about missing measurement, not about AI being weak.
What would settle it is specific. A controlled trial in which sensor-guided or app-guided rehab is compared with therapist-led care, in matched patients, measuring function, pain, adherence and re-injury at set intervals. Studies that look only at exercise recall, chatbot answers or motion tracking accuracy sit next to the job rather than on it. The sources listed further down this page are graded on that basis.
When this could change
Most likely after 2042 (8 in 10 of our scenarios). The replacement-year method explains exactly what that window measures and how the spread is produced.
Two things could pull it earlier. Cheap, accurate motion capture in ordinary clinics and living rooms would make remote monitoring routine, which can stretch the gaps between in-person visits. And if payers accept monitored home programs in place of some scheduled sessions, caseloads get reshaped even though the profession stays.
Two things hold it back. A large part of this job is physical, and our robotics read puts the hardware needed at dexterous humanoid level, which is not in clinics today; the guide to humanoid robots and physical jobs covers why that bar is high. State licensure and supervision rules are the other brake: treatment plans and reassessments carry a licensed name, and no model holds a license. Demand is also still climbing. BLS counts about 267,330 physical therapists in the United States, with median pay of $102,760 and projected employment growth of 11.9% from 2025 to 2035 (BLS).
How to stay needed
Lean into the work the task split leaves with people. Complex evaluation, where the presenting complaint is not the real problem, is the clearest example. Hands-on care for post-surgical, neurological and older patients is the second. Supervising assistants and aides, and holding the quality bar across a caseload, is the third, and it grows as more of the routine work gets delegated.
Two skills are worth building now. First, reading and auditing AI-drafted documentation fast, so notes stay accurate and defensible rather than plausible. Second, working with motion and outcome data: knowing what a sensor score actually measures, and when to overrule it. Both make you the person who signs the plan, not the person who types it.
What to do: pick one documentation task this month, run it through an assistive tool, and track how much time you keep and how often you correct it.
Nearby work moves at a different pace. Compare this job with occupational therapists, physical therapist assistants and athletic trainers, or put any two side by side in the job comparison tool. You can also read the wider picture for diagnosing and treating practitioners, the healthcare sector, or the list of jobs that mostly need a person.