AI Use in Physical Therapy: Which Uses Need Clinician Sign-Off

In a physical therapy clinic, any AI output that enters the clinical record, advises a patient or sets exercise dosing needs a licensed clinician's review and sign-off. Admin tasks with no patient information, such as rotas, newsletters and policy drafts, can run with a manager's check. Clinical advice from a chat tool should not be used at all.

The trouble is that most uses sit between those ends: a reminder text, a draft note, a home programme. An owner needs a rule the front desk can apply in seconds on a busy Tuesday, and the simplest one is a ladder: the more an output can change what happens to a patient, the more senior the person who signs it. A reminder text sits near the bottom; a home programme sits near the top.

Follow me on Instagram@sagnikteaches

The five-level ladder

Five rungs cover almost every use a clinic owner will meet, from admin to clinical advice. Sorting by risk also keeps the burden sensible, because most AI use in a clinic is low risk and needs very little checking.

Connect on LinkedInSagnik Bhattacharya

Levels 1 and 2 can run with light oversight. Levels 3 and 4 never run without a clinician. Level 5 is a line the clinic draws: a model may be a source of questions to consider, but the clinician decides, documents the reasoning and owns the result. Tools marketed as clinical decision support are a separate category with their own regulatory and evidence questions, and are outside this tutorial.

Subscribe on YouTube@sagnikteaches

Level 1: admin work that needs only a manager's eye

This is where most owners should start. A chat assistant can draft the monthly newsletter, rewrite a cancellation policy in plainer words, outline a staff meeting, write a job advert, or propose a rota. None of this involves patient information, so there is no clinical sign-off, only a manager who reads it before it is used. A sensible control is a monthly check: the owner picks five pieces of work done with AI and asks whether each was reviewed. The patient retention messages tutorial covers marketing-side work at this level in depth.

What goes wrong. Invented facts about services (a technique the clinic does not offer), clinical claims in a newsletter ("this will cure back pain"), and tone that sounds sure when it should be careful. A manager catches these by reading with one question: "Is every sentence something we actually do or have approved?"

Level 2: patient messages from templates the owner approved

Reminders, rebooking nudges and review requests are the next rung. They touch a patient's name and appointment time, so they should run through your booking or messaging system, under a contract, rather than a pasted chat. The sign-off here is at the template: the owner or lead clinician approves wording once, and staff use it unchanged. A model can help write the template; it should not improvise each patient message.

An illustrative template, approved once: "Hi [first name], this is a reminder of your appointment at [clinic] on [day] at [time]. Reply C to confirm or call us to change it. Please do not reply with symptom questions; call the clinic." The last sentence matters, because a patient who texts "my knee swelled up overnight" to an automated number must not be left waiting. Set the rule that any reply containing a symptom word goes to a clinician queue the same day. For the front-desk side, the AI receptionist tutorial for referrals and insurance explains how to stop a phone assistant from drifting into clinical territory.

Level 3: notes and letters, signed by the clinician who treated

An AI scribe or dictation tool drafts the treatment note; a referral letter or progress report may be drafted by a model from your own notes. Both enter the record and may be read by a payer, a referring doctor or an auditor. The sign-off rule is blunt: the treating clinician reads the whole draft and signs as the author, every time.

The danger specific to physical therapy is the cloned note: the same assessment findings carried across visits with the dates changed. A tool that autofills from the last visit makes it easier and the record less honest. The tutorial on avoiding cloned notes shows the checks. On the payer side, the tutorial on whether payers accept AI-written documentation covers what claims reviewers look for.

A filled-in sign-off check for a draft note

Give clinicians a short check to run before signing. In an illustrative form, it reads:

1. Subjective: does it match what the patient actually said today, not last week?
2. Objective: are the measures (range, strength, pain score) the ones I took, with today's values?
3. Treatment: does it list what I did and the dose, not what is typical for this diagnosis?
4. Plan: does it reflect my decision about frequency and goals?
5. Anything invented: any test, finding or advice I did not perform or give?
6. Time: do the minutes and units match the visit?

If any answer is no, edit it. An error caught here is a minor annoyance; an error caught by a payer is a recoupment.

Before you pick a scribe, ask the vendor to show you on a demo note how it marks anything carried forward from the previous visit. If it cannot, your sign-off check has to catch every carried-over finding by hand, and the time saving shrinks.

Level 4: exercise programmes need the clinician's eyes on the plan

This rung catches owners out because the output looks harmless: a list of exercises with sets and repetitions. But a home programme is an instruction that a patient will follow alone, at home, possibly while in pain. A model has not examined the patient and does not know their precautions, their surgeon's protocol or their ability to understand the sheet.

An illustrative prompt and sample output (labelled as such), then what a clinician fixes:

Write a one-page home exercise sheet in plain English for an adult
five weeks after a knee replacement who has been cleared to progress.
Include six exercises, sets and repetitions, a warm-up, and a line on when
to stop and call the clinic. Do not include medical claims.

The sample output lists seated knee extensions, heel slides, straight-leg raises, mini squats, step-ups and calf raises with "3 sets of 10". It reads well and is wrong for several patients. The clinician fixes: the squat depth that exceeds this surgeon's limit, the step-ups before the patient can manage stairs, the repetitions that ignore this patient's swelling, and the stop-and-call line, which must use the clinic's own wording and number. Only then does it go out. The tutorial on home exercise programmes with AI shows a fuller drafting workflow where the clinician stays in charge of the content.

The rule at Level 4: the clinician who knows the patient reads every line of any AI-drafted programme or instruction, edits it for that patient, and is the named author on the sheet. A programme that a patient receives without that step is a clinical decision made by software.

Level 5: the decisions that stay with the clinician

Some work should not be delegated to a chat tool whatever the sign-off: diagnosis, deciding whether a symptom is a red flag, progressing load, deciding when to refer on or discharge, and anything involving a patient who is deteriorating. A clinician may use a model like a colleague to talk through options after the assessment, using no identifying detail, but the reasoning and the decision are theirs, and the record shows that.

It helps to write this down for staff, because patients will also ask AI tools about their own symptoms and arrive with printouts. A short line in your clinic policy covers it: "We welcome questions about information you have found; your clinician will explain how it applies to you."

A six-clinician clinic over one illustrative week

To see how sign-off works day to day, picture a clinic with six clinicians and two front-desk staff. One week, the owner logs every AI-assisted item and where it sits on the ladder (all figures illustrative).

The review time is 545 minutes, about 9 hours across six clinicians, around 1.5 hours each. Without the tools, writing 110 notes from scratch at, say, 7 minutes each would be 770 minutes. Reviewing at 3 minutes saves 4 minutes a note, or 440 minutes, but only if the review is real. The sum also shows where sign-off time goes: 93 per cent of it is on notes and exercise sheets. That is the right place for it. If clinicians complain that review takes too long, the answer is a better tool or better prompts at Levels 3 and 4, not a lighter sign-off.

A referral letter, before and after sign-off

Letters to referring doctors sit at Level 3 and show the sign-off at its most useful. Suppose a clinician dictates three lines about a patient and asks the approved tool to turn them into a progress letter. The illustrative draft says the patient "has made excellent progress and has met all goals" and "no further treatment is required". The clinician had said the patient met two of three goals and still had a limited overhead reach.

The draft is fluent, positive and wrong in two places. The clinician corrects "all goals" to the two actually met, restores the reach limitation, and changes "no further treatment is required" to the plan she had decided: two more visits and a review. She signs. Had the letter gone unread, the referring doctor would have closed the case and the patient's shoulder would have been left to chance. The pattern is general: models smooth over the awkward parts of a clinical story, and the awkward parts are the ones that matter.

What front-desk staff may say about AI, and what they may not

Patients ask. Sooner or later someone asks the desk whether an AI wrote their note, or whether the clinic's chat assistant can tell them what is wrong with their back. Give staff two short scripts so they are not improvising a clinical answer.

• "Did a computer write my notes?" "Your therapist may use a tool to help draft notes, and they read and sign every one. If you would rather we did not use it for your visit, tell us and we will note that."
• "Can your assistant tell me what's wrong?" "No, our assistant helps with bookings and practical questions. Your therapist assesses you and explains what they find. I can book you in or pass a message to them now."

Make it equally clear that patients can decline AI-assisted notes. Consent rules vary, so ask your adviser what applies to you; the habit of offering an alternative costs little and builds trust.

A patient arrives with a chatbot's advice

An illustrative scene. A patient with shoulder pain turns up with a printed plan from a free chat tool: a list of stretches and a note that "this will fix it in two weeks". The clinician's job is not to dismiss it but to treat it as information from the patient. She reads it, flags two stretches that would aggravate the patient's current irritability, explains why, and records in the note what the patient had been doing and what she advised instead. The clinic policy is the same as for any other source: the clinician decides, and documents. Patients who feel heard are more likely to keep telling you what they are doing at home.

Moving a use up or down the ladder

The ladder should change as evidence comes in. Here are two examples with numbers (illustrative). Reminder wording starts at Level 2 with the owner approving every template change. After three months with no complaints and no clinical replies mishandled, the owner reduces the check to a monthly sample of ten messages. That is a sensible move down. In contrast, exercise sheets start at Level 4. During a month's audit the owner finds that in 6 of 20 sampled sheets the clinician changed the dose or removed an exercise. A 30 per cent edit rate is not a reason to stop, but it is a reason to improve the prompt, add the clinic's standard precautions to it, and keep every sheet at full clinician review. Evidence moves uses; hope does not.

Writing the policy so it is one page

An owner's policy does not need legal language. Put the ladder on a page, add four rules, and put it where staff will find it.

1. Which tools. Name the approved tools and the level each is approved for. Anything at Level 3 or 4 uses a tool under a business associate agreement or equivalent contract for that product and plan. Vendors differ on this: OpenAI's help centre says it does not offer a BAA for ChatGPT Business, Anthropic's BAA covers specific services such as its API and Enterprise plans after the Primary Owner activates it, and Google Workspace requires an administrator to accept a BAA first. OpenAI's separate ChatGPT for Clinicians, which lets an individual sign a BAA, currently admits only verified physicians, nurse practitioners, physician assistants and pharmacists in a single market, so physical therapists are not on its list. Read each vendor's current page.
2. Who signs. Level 3 and 4: the treating clinician. Level 1 and 2: the manager or owner.
3. What is never done. No patient identifiers in a general chat tool, no AI clinical advice to patients, no unsigned AI draft in the record.
4. What happens on a mistake. Record it, correct the record, tell the owner, and change the instruction that produced it.

The wider question of choosing a compliant vendor is covered in the tutorial on which vendors sign a BAA, which is worth reading before you promote any tool from Level 1 to Level 3.

Checking the sign-off is real, not ceremonial

A sign-off that takes five seconds is not a review. Three light checks tell you whether yours works. First, sample ten signed notes each month and compare each with what you know of the visit; look for invented content and carried-over findings. Second, look at timestamps: if a clinician signs 12 notes in 90 seconds, nobody read them. Third, ask clinicians what the tool got wrong this month. If the answer is always "nothing", they are either not reading closely or not telling you, and both are a problem.

Track one number for each level: errors found at sign-off. A rising number at Level 3 is useful information about the tool, and a number that stays at zero for months should make you more suspicious, not less. When the evidence says a tool is good, you can move a use down the ladder, for example from a manager reading every reminder template to a monthly sample. When it says otherwise, move it up or switch the tool off. Reviewing the ladder every quarter keeps it honest as tools and your clinic change.

Questions clinic owners ask about sign-off

Does every AI output need a clinician to read it?

No. Admin tasks with no patient detail, such as a rota or a newsletter, need only a manager's glance. Anything that enters the clinical record, advises a patient or sets an exercise dose needs a licensed clinician to read and approve it before it is used.

Can an AI-drafted home exercise programme go straight to the patient?

Not without review. A clinician must check it against the patient's diagnosis, stage of healing, precautions and ability. Generic programmes written by a model can include movements that are wrong for a particular person.

Who signs off an AI-written treatment note?

The treating clinician, who reads the whole note, corrects it and signs as the author. The sign-off must be real, not a click-through, because the clinician is responsible for what the record says.

What does sign-off look like in practice?

A named clinician reviews the output, edits it, and records that they did so, either by signing the note in the system or by initialling the draft. For lower-risk items a short weekly sample check by the clinic owner is enough.

Can front-desk staff use a chat assistant?

Yes, for tasks with no patient information: drafting a newsletter, rewording a policy, planning a rota. For patient messages, give them approved templates and a rule that anything describing symptoms goes to a clinician.

Further reads

• Physical Therapy Prior Authorisation: Track Visits With AI — Tracking visit counts and authorisations without a clinical decision.
• Physical Therapy Billing: AI Checks for Claims and Time Entries — Claim and time-entry checks that sit beside note sign-off.
• Using ChatGPT for Physical Therapy: Referring-Doctor Updates and Outreach — Referring-doctor updates and outreach with ChatGPT.
• AI Tools for Physiotherapy Clinics: What Each One Actually Does — What each category of tool does in a physio clinic.
• Physiotherapy Clinic Software Compared: Which Has Built-In AI? — Which practice software has AI features built in.
• How Physiotherapists Use AI to Write Treatment Notes Faster — How clinicians use AI for treatment notes, step by step.
• Best AI Scribe for Physical Therapy: Four Tools Compared — Four scribes compared on the features a PT clinic should weigh.
• Physical Therapy Policies and Procedures: A Clinic Manual Drafted With AI — Where an AI policy line fits in the clinic manual.
• Use AI to Write a Care Plan Review From Daily Notes, Safely — Draft a care plan review summary from a month of daily notes without exposing identities, then trace every line back before sign-off.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Center articles on business associate agreements and ChatGPT for Clinicians (checked November 2026); Anthropic privacy centre on BAAs for commercial customers; Google Workspace health-data compliance help. Sign-off rules here are operational suggestions, not licensing advice; check your own licensing board and payer rules.