Last updated: June 1, 2026
It’s 6:15 on a Tuesday. You’ve got four unfinished SOAP notes, a new post-op rotator cuff patient who needs a home program before they leave, and a waiting room that isn’t waiting. So you open ChatGPT, type “home exercise program for rotator cuff repair,” and ten seconds later you have a clean, formatted list of exercises. It looks great. It’s also, in this case, a little too eager — it included an overhead movement your surgeon’s protocol explicitly forbids for another three weeks.
That’s the whole story of AI in physical therapy right now, in one moment. The tool is fast, fluent, and genuinely useful for the parts of your day that eat your evenings. It’s also a confident generalist that has never read your patient’s op report, doesn’t know their weight-bearing status, and will happily hand you a textbook protocol that’s a year out of date. PTs who use it well have a phrase for what they’re working with: it’s a B+ DPT — a talented intern who drafts fast and needs everything checked.
This is how to get the speed without the risk. We’ll build a home exercise program (HEP) the safe way, turn shorthand into a clean note, and stay on the right side of HIPAA — because the part nobody mentions in the “10 ChatGPT prompts for PTs” posts is the part that protects your license.
What’s actually happening in clinics
Adoption among PTs is real but early, and it’s lopsided: far more therapists use AI for documentation than for exercise programs. The reason is obvious once you’re in the chair — notes are the thing stealing 45 to 60 minutes a day, so that’s where the relief is most wanted. On the community side, the conversation has moved past “should we?” to “here’s my setup”: therapists trading prompt structures, comparing ambient scribes, and even building their own tools. One PT taught himself enough to build a HEP generator with Claude in a few months; another runs a HIPAA-compliant pipeline through AWS so, in his words, “the PHI never leaves” his own cloud.
The market is following the demand. The U.S. AI-in-physical-therapy space was worth roughly $178 million in 2025 and is projected to pass $1 billion by 2033. That growth isn’t about robots replacing therapists — it’s about software absorbing the documentation and handout-formatting work that has nothing to do with hands-on care. The therapists getting value aren’t the ones who trust the output. They’re the ones who treat it as a first draft and keep their clinical judgment firmly in the loop.
Building a HEP the safe way
Here’s the workflow that keeps the speed and removes most of the danger. The danger, almost always, is what the AI doesn’t know — so the fix is telling it, explicitly, before it writes a word.
The difference between a generic exercise list and a clinically useful draft is entirely in the prompt. The PTs who do this well front-load the safety constraints. A prompt that works looks like this:
“You are helping me draft a home exercise program. Patient: 58-year-old, 4 weeks post right rotator cuff repair, following a standard protected protocol. PRECAUTIONS (put these in bold at the top): no active external rotation, no shoulder elevation above 90°, no lifting. Goal: protect the repair while maintaining elbow/wrist mobility. Build a phased program — Phase 1 now, with criteria for advancing to Phase 2. For each exercise give sets, reps, and frequency, plus a position or range limit. Add an ‘Activities to Avoid’ section and a ‘When to Call Your Therapist or Surgeon’ section. Write at a 6th-grade reading level. Do not include any exercise that violates the precautions above.”
Notice what that prompt is doing. It states the precautions and asks for them in bold at the top of the handout. It asks for phases with advancement criteria, because left alone the model defaults to a single generic protocol. It demands an “Activities to Avoid” and a “When to Call” section, which turn a list of exercises into actual patient education. And it sets a reading level, because the best exercise in the world does nothing if the patient can’t follow it.
Then comes the step the viral prompt threads skip entirely: you read it against your own clinical judgment before it touches a patient. Does any exercise contradict the precautions you just typed? Is the dosage appropriate for this stage and irritability? Did it quietly include something that doesn’t fit the diagnosis? This is a 60-second read for an experienced clinician, and it’s the difference between a tool and a liability.
Turning shorthand into a SOAP note
The same pattern saves even more time on documentation. Type your shorthand right after the session — “Pt c/o 4/10 R knee pain on stair descent, AROM flex 115° (up 10°), quad sets 3×15, gait training 10 min, tolerated well, plan step-ups next visit” — and ask the AI to expand it into a clean SOAP note.
One instruction is non-negotiable here: “Use only the information I provided. Do not invent any measurements, ranges, or scores. If something is missing, leave it blank or write ’not assessed.’” Without that line, the model fills gaps with plausible-sounding numbers — a goniometry reading you never took, a pain score the patient never gave. And a fabricated objective measure isn’t just sloppy; if that note becomes the basis for a billing claim, you’ve crossed from documentation error into Medicare/Medicaid fraud territory, with audit, recoupment, and licensure exposure attached. The note has to reflect what you actually did. AI can format it; it can’t witness the session.
The HIPAA layer the prompt threads skip
This is where a clinician’s guide has to part ways from a consumer’s. The free, fast workflow everyone posts about has a problem: consumer ChatGPT, Claude, and Gemini do not sign a Business Associate Agreement. Under HIPAA, any vendor that handles protected health information on your behalf needs one. Paste a patient’s name, date of birth, or any combination of details that could identify them into a public chatbot, and you may have just created a reportable breach.
There are two clean ways around it:
- De-identify before you paste. Replace the name with “Patient A,” strip the date of birth, MRN, address, dates of service, and referring provider. Use “58-year-old male” instead of anything that points to one person. This is the minimum, not the nicety.
- Use a tool covered by a BAA. That means an enterprise tier with a signed agreement, a health-specific ambient scribe that offers one, or a self-managed setup (some PTs route Claude through AWS Bedrock with a signed BAA so patient data stays inside their own cloud). For anything with real identifiers, this is the only safe path.
What this means for you
If you’re an outpatient ortho PT: The phased-HEP prompt is your biggest win — your caseload is full of the post-op and overuse cases where a clean, leveled handout matters. Build two or three reusable prompt templates with the precaution scaffolding baked in.
If you’re a home-health PT: Lean on AI for the plain-language patient and caregiver handouts, where reading level and a clear “when to call” section genuinely improve carryover. Keep all identifiers out unless you’re on a BAA-covered tool.
If you’re a PTA: Drafting is your sweet spot, and the review step is your protection. Generate the note or handout, then verify it against what actually happened in the session — and remember the documentation is still authored and owned by the supervising structure you practice under.
If you’re a new grad or student: Use AI to learn faster, not to skip the reps. Ask it to explain why a progression criterion is what it is, then check it against a real protocol. The therapists who worry about “losing the skill to do it manually” aren’t wrong — build the judgment first, then let the tool save you time.
If you own the clinic: Decide on one approach before your therapists each build a shadow workflow on personal ChatGPT accounts. Pick a BAA-covered tool, write a one-page policy, and resist the temptation to turn time saved into higher productivity quotas — the therapists already feel that one coming.
What AI can’t do here
- It doesn’t know your patient. No imaging, no op report, no precaution flags unless you type them. The contraindications it misses are the ones you didn’t mention.
- It will hand you outdated protocols. Its training has a cutoff; your clinical guidelines don’t. One PT had to catch it recommending contrast baths for CRPS — a treatment that fell out of favor years ago.
- It invents data when you let it. A made-up range of motion or pain score is the single most dangerous output in a clinical note. The “do not invent data” instruction and your own eyes are the only safeguards.
- It defaults to generic. Without condition-specific prompting, you get a population-level protocol, not a program for this tendon at this stage.
- It can’t carry the responsibility. The licensed clinician owns every exercise prescription and every word of documentation, no matter what drafted it. As one therapist put it, watching people grab prompts online: the worrying part is “not one question about how accurate it is, not one about how safe it is — just give me the prompt because it’s free.”
The bottom line
The time AI can give a physical therapist back is real — it’s the 45 minutes of charting and the handout-formatting, not the clinical reasoning. The therapists who win with it aren’t the ones who trust it or the ones who refuse it. They’re the ones who learned to brief it like a sharp intern: state the precautions, demand the phases, keep the PII out, and read every line before it reaches a patient. Do that, and you get your evenings back without ever handing over the judgment that makes you the therapist.
Want to build the documentation side without the HIPAA guesswork? Our AI Therapy Notes: The HIPAA-Safe Workflow course walks through exactly how to use AI for clinical notes while keeping patient data protected and your documentation defensible.
Sources
- APTA — Artificial Intelligence in Physical Therapy
- HHS — HIPAA & De-Identification of PHI (Safe Harbor)
- PMC — Large language models in physical therapy practice (peer-reviewed review, 2024)
- HHS — Business Associate Contracts (BAA requirements)
- CMS — Complying With Medical Record Documentation Requirements
