Why this work stays in the room
Occupational therapy assistants spend the day within arm’s reach of a patient. You guide someone through therapeutic exercises, steady a transfer from bed to chair, and read a wince before deciding whether to push on or stop. Ask whether AI will replace occupational therapy assistants, and the task split above gives the plain answer: 78% of task time still needs a person in the room.
The teaching side is just as physical. Helping a stroke patient relearn dressing, bathing or cooking means watching a hundred small adjustments and changing your cue mid-task. A model can suggest an activity. It cannot place a hand under an elbow, judge how much weight a shoulder will take today, or calm someone who is frightened of falling.
Where change is already landing is the paperwork wrapped around the session. Progress notes, billing codes, scheduling and supply logs are text and numbers, and text and numbers are what current tools handle best. That is task erosion, not a job disappearing. It shows up as fewer unpaid minutes after a shift, and sometimes as thinner entry-level rosters where a clinic uses the time saved instead of hiring.
What AI does, what it assists with, what remains yours
The tasks our data puts in the “AI does it” group are documentation-shaped: drafting progress notes from dictation, and handling insurance and scheduling paperwork. That slice is 0% of task time. The clinical content still has to be checked and signed by a person, because the note has to match what actually happened in the session.
A larger set of tasks sits in the assisted group, at 22% of task time. Think of building a home exercise handout from the therapist’s plan, or tracking repetitions and progress between visits. The tool speeds up the writing and the tallying. The decision about what a patient is ready for stays with the therapist and the assistant who watched them move. Overall, this job scores 16 for coverage, our estimate of the share of task time AI can handle today; how coverage is measured explains what that number counts.
Everything hands-on lands in the needs-a-human group: safe transfers and positioning, physical cueing through an activity, adjusting and fitting adaptive or orthotic equipment, and reporting what you observed back to the supervising therapist. Nearly half this job’s content is physical work, and the robot class that could attempt it is a dexterous humanoid, which is not something clinics can buy, insure and supervise at a sensible price. Our guide to humanoid robots and physical jobs walks through why that gap is wide.
What the evidence actually supports
Honest answer first: nobody has tested AI head to head against occupational therapy assistants on their own work. Our quality-parity grade for this job is D, which is the grade we use when there is no direct measurement. So we publish no parity number here, and you should be wary of anyone who does.
What exists is adjacent. Documentation tools have been measured on drafting speed and note quality in clinical settings generally, not on rehab sessions led by assistants. Published work on AI in occupational therapy has focused on how the profession should adopt these tools, not on whether a model can substitute for a treating clinician. That is a real difference, and the evidence list on this page is graded accordingly.
What would settle it is specific: a trial that compares patient function scores and goal attainment between assistant-led sessions and sessions where an AI system plans and delivers the activity, with the same supervision and the same patients. Until that study exists, the parity question stays open. Our quality-parity method sets out the grades, and the full scoring method shows how the three questions fit together.
When the picture could shift
Most likely after 2045 (8 in 10 of our scenarios). The replacement-year method explains what that window is based on and how wide the uncertainty is.
Two things could pull it earlier. First, documentation automation built straight into electronic health records, which keeps shaving minutes off the admin half of the role. Second, falling costs for capable hardware: the cost panel above already shows an AI tool running for a fraction of a year of human labor on the paperwork tasks, which is exactly where substitution pressure starts.
Two things hold it back. Licensure and supervision rules tie treatment to a credentialed person working under an occupational therapist, and payers generally require that documented human contact. And the physical work needs reliable, gentle, insured dexterity around frail bodies, which no deployed robot offers today. Demand is the other side of this: BLS projects employment for this job to grow about 21.5% between 2025 and 2035, from roughly 51,290 jobs, with median pay near $72,300 (BLS). Growth that fast does not sit well with a story about the role vanishing.
How to stay needed
Lean into the parts of the day that stay with people. Three worth building a reputation on: safe transfers and positioning, hands-on cueing through daily-living tasks like dressing and meal prep, and fitting or adjusting adaptive equipment so it actually gets used at home.
Two skills raise your floor. One is clinical observation you can put into words, because a therapist changing a plan needs your description of what the patient did, not a generated summary. The other is editing AI drafts well: knowing what a note must contain for the record and the claim, and catching what a model invented or smoothed over.
What to do: keep a short log of the judgment calls you made in sessions this month, and use it in your next review or interview.
Nearby roles are worth a look if you are weighing a move. Compare this job with occupational therapy aides, physical therapist assistants and occupational therapists, which share much of the same work. You can see the whole group on the therapy assistants and aides family page, or how the wider healthcare sector scores. Put any two of them side by side with the job comparison tool, or see where hands-on roles land in our list of jobs least exposed to AI.