Best AI to Help With PT Documentation: Avoiding Cloned Notes

A PT clinic can use AI for documentation without cloned notes by making each note start from what happened in that session (what the patient did, how they responded, what changed) and by having the treating therapist review and sign every note. AI drafts from session content; it must never copy the last visit forward.

Auditors and payers read cloned notes as a sign that the record was not written for this visit. Copy-forward was a problem long before AI, and a badly set up AI tool can make it worse by producing fluent, near-identical paragraphs at speed. The fix is a handful of per-visit fields the AI cannot invent, a similarity check across a patient's notes and a review step that cannot be skipped.

Follow me on Instagram@sagnikteaches

What auditors mean by a cloned note

Cloned documentation means notes worded exactly like, or very like, earlier entries, either across visits for one patient or from patient to patient. That is close to the wording one Medicare contractor, National Government Services, used in a documentation note (read here in an archived copy): it asked for individualised notes for each encounter showing the condition needing treatment, the treatment given and, where it applies, the patient's progress, and said cloned documentation would be treated as misrepresenting medical necessity, whether it came from an electronic record, a preprinted template or handwriting. The note applied this to any discipline that has to show progress towards treatment goals, which includes therapy.

Connect on LinkedInSagnik Bhattacharya

The manuals frame the same risk through medical necessity. CMS's Program Integrity Manual (chapter 3, section 3.3.2.1.1, read in November 2026) says CMS does not prohibit templates, discourages limited-option ones built on check boxes and predefined answers, and warns that templates designed mainly for reimbursement often fail to show that coverage and medical-necessity requirements are met. An AI drafting tool is, in effect, a very fast template, so the same warning applies to it.

Subscribe on YouTube@sagnikteaches

Two honest caveats. First, the manual chapter I read does not use the word "cloned" at all; the risk reaches you through medical-necessity review and contractor guidance rather than one named rule. Second, a template is not itself a problem. A template is a structure for capturing a note, and it can be specific to the patient and the visit. Default text left untouched is what gets flagged. Other payers have their own manuals, so treat what follows as a clinic process and confirm the details with each payer you bill.

Where AI adds cloning risk and where it removes it

AI changes the cloning picture in both directions.

• It adds risk when a tool is asked to "write a note like last visit", when a scribe fills a missing section with typical language, or when a therapist accepts a draft in twenty seconds. Typical language is the definition of cloned text.
• It removes risk when it drafts from the session itself (spoken or dictated content), when it flags sections with no session content, and when it compares a new note to previous ones and warns about overlap.

The deciding factor is the input. An AI that works from what was said and done produces notes that differ because the sessions differ. An AI that works from the last note produces clones. You can see which one you have by running two consecutive visits for the same patient and comparing outputs.

The per-visit details an AI cannot make up

List the items a note must contain that exist only because this session happened. Make the therapist supply them (by voice or a quick entry) before the AI drafts anything.

If a field is empty, the draft should say so, not fill it. That single rule removes the most common way AI notes slide into boilerplate.

The plan-change row matters more than it looks. Medicare's Benefit Policy Manual (chapter 15, section 220.3) says a treatment added or changed between progress reports must be recorded and justified, and that new exercises or changes to the programme help show the services are skilled. A note that records "added single-leg balance because stairs pain is now 4/10" is doing two jobs at once: it is specific to the visit, and it is evidence that a therapist's judgement was needed.

A prompt that drafts from session content only

Whether you use an AI scribe or a general assistant on a plan covered by your data agreement, set instructions that stop it from borrowing language. A prompt you can adapt:

Draft a daily PT note from the session content below. Use only what is in the content. Headings: Subjective, Objective measures, Interventions (with sets, reps, load, time), Response, Assessment, Plan.
Rules: do not copy wording from any earlier note. If a heading has no content, write "not recorded" and list what the therapist should add. Do not write generic phrases such as "patient tolerated treatment well" unless the therapist said it. Keep measurements and units exactly as given. Do not add diagnoses or goals.
Session content: [therapist's dictation or typed points]

Sample output for the knee example above, illustrative only:

Subjective: Pain 4/10 after stairs, down from 6/10. Slept through twice this week.
Objective: R knee flexion 112 degrees; extension lag 3 degrees.
Interventions: Step-ups 3 x 10 at 15 cm with 2 kg added; manual patellar glides 8 minutes.
Response: Mild swelling after step-ups; settled after 5 minutes of elevation.
Assessment: not recorded. Therapist to add clinical reasoning and progress against goals.
Plan: Progress to single-leg balance next visit. Hold stairs training until swelling stays down.

What you would fix: the assessment is blank on purpose. A therapist who writes "ROM improving, consistent with healing; still limited on stairs, which needs skilled cueing" has just added the sentence that makes the note defensible. That sentence is the therapist's job. The AI's job was to get everything else onto the page.

Running a similarity check across a patient's notes

You do not need special software to spot cloning. Once a month, for each therapist, pull three consecutive notes for five patients and look for these signs.

1. Whole sentences repeated word for word across visits.
2. The same numbers (pain score, range) with no change over several weeks despite different interventions.
3. Identical response lines such as "tolerated well" every visit.
4. Different patients with matching paragraphs, which suggests a shared default.

You can also ask an AI to do the comparison on de-identified text. Remove names, dates of birth and anything identifying, then use a prompt like this:

Below are three consecutive daily PT notes for one patient, de-identified. List any sentences that appear in more than one note unchanged or with only minor changes. List fields where the values are the same in all three. Do not rewrite the notes. Output a table: sentence or field, which notes, what is missing that a reviewer would expect to differ between visits.

Treat the answer as a pointer, not a verdict. A static finding, such as an unchanged diagnosis, is legitimate. Repeated treatment-response wording is not. If you use patient text at all, confirm your agreement with the AI provider covers it, and prefer to review notes in the clinic's own system where the data already lives.

A worked month in a four-therapist clinic

Say a clinic with four therapists sees 480 visits a month. Notes took about nine minutes each when typed, and several therapists leaned on copy-forward at the end of a long day. The clinic introduces a scribe that drafts from session dictation, with the "not recorded" rule above.

• Week 1: each therapist dictates three lines after each visit (subjective, measures, interventions). Draft notes appear within minutes. Review takes about three minutes per note instead of nine to write.
• Week 2: the clinic lead pulls five patients per therapist, three notes each: 60 notes. She finds seven with repeated response wording ("tolerated well, no adverse reaction") from one therapist, whose dictation had become habitual.
• Week 3: the therapist agrees to dictate one response sentence that names what actually happened. Repeats drop in the next audit sample from seven to two.
• Week 4: the clinic records its numbers: average review time, notes with "not recorded" left unfilled at signing (target zero), repeats found.

The saving here, 6 minutes a note across 480 visits, is 48 hours a month. These are illustrative figures. Time your own clinic for two weeks before and after, because the result depends on how well your therapists dictate.

Same condition, different notes: three patients with one diagnosis

The hardest cloning risk is in a busy week of near-identical patients: three people after knee replacement, all in week four. Their notes will legitimately share vocabulary. What must differ is what happened. A before and after, using fictional patients, shows how a therapist makes that visible.

Cloned version, patient one and patient two: "Patient tolerated exercises well. Range of motion improving. Continue plan of care."

Specific version, patient one: "Flexion 98 degrees, up 8. Completed bike 8 minutes at low resistance; stopped early on the second try at minute 6 because of pulling behind the knee. Needs verbal cue to keep the hip level on step-ups. Add heel slides at home, 2 x 15 daily."

Specific version, patient two: "Flexion 84 degrees, no change in two visits. Swelling visible at the joint line, so held step-ups and used 10 minutes of elevation with ankle pumps. Messaged the surgeon's office about the plateau. Re-measure Thursday."

Each of the second pair could only have been written about that person on that day. The therapist reading the draft should be able to tell, from a note alone, which patient it concerns without seeing the name. That is a handy test: cover the name and see if the content points to one person.

Progress notes and re-evaluations need the most care

Daily notes are where cloning creeps in, but the summaries that justify continued care are where it hurts most. A progress note or re-evaluation should show change against the goals set at the start. An AI can pull the numbers together: initial measures, last three sets of measures and the stated goals, laid out for the therapist. It should not write the clinical conclusion.

From the measures below, build a table of goal, baseline, current value, change and visits since baseline. Do not write an assessment. Flag any goal with no new measure in the last three visits. Measures: [de-identified list]

The table saves ten minutes of hunting through the record. The therapist then writes why the patient still needs skilled care, in their own words, and states the revised goals. If you bill Medicare, the same manual section sets the minimum progress report period at once every 10 treatment days, and a report written by an assistant does not count as a complete progress report, so the therapist must write one in each period. Other payers set their own intervals, so take them from each payer's current manual and put them into your reminders. A ten-day counter in the record, or a weekly report written by habit, stops the AI-assembled table from turning into a late, rushed summary.

Plateaus and stable patients: when unchanged is correct

Some patients really do stay the same, such as a person with a chronic condition on a maintenance plan. Their notes will look similar, and that is not automatically cloning, but it needs more care, not less. A note for a stable patient should say what was assessed to confirm stability (which measures, compared with when), why skilled therapy is still needed, and what would change the plan. Tell the AI to leave the "why skilled care" line for the therapist and to flag when three notes in a row have identical measures.

Training therapists to dictate for the draft

Most weak AI notes trace back to weak input. A two-minute habit gives better drafts: say the patient's report, the numbers, what you did with dosage, how they responded, and what changes next time, in that order. Practise with a colleague listening.

• Say numbers with units ("flexion one-twelve degrees"), since a missing unit becomes a guess.
• Name the exercise and the dose in one breath ("step-ups, three sets of ten, fifteen centimetres, two kilos").
• Say one sentence about response that you would not say about the next patient.
• Say the plan change, even if it is only "no change, same programme, here is why".

After two weeks, compare drafts from the same therapist for the same condition. If the response lines have started to match, that is the moment for a gentle conversation, not a policy memo. A therapist who is tired at 5 pm needs a quick way to give a specific sentence, such as a voice note on the way out, rather than a lecture about cloning.

An audit-day folder you can prepare now

If a payer asks for records, you will want to show the process as well as the notes. Keep one folder, updated each quarter, holding four things: the current templates (with any default text removed), the monthly similarity-review results with dates, the written review checklist therapists use, and a record of any corrective action taken. It is also worth recording which AI tool is used, from what date, and who reviews its output. Notes remain the therapist's clinical record regardless of who or what drafted them, and the folder helps you show that.

The therapist review step, in writing

The review step is what turns a draft into a record, and it needs to be as concrete as the prompt. Give therapists a six-point read before they sign.

1. Does the subjective match what the patient said today, not last week?
2. Are the measures and units right, and does any number look copied?
3. Do the interventions list dosage, and did I actually do each one?
4. Is the response specific to this visit?
5. Is the assessment my own reasoning, with a skilled need stated?
6. Does the plan follow from today, not from a template?

Rule: no note is signed with a "not recorded" marker still in it, and no note is signed within a minute of the draft appearing unless it is a very short visit. A time-stamp check on sign-off lets the clinic lead spot rubber-stamping. For the broader question of which AI uses need a clinician's signature, AI use in physical therapy: which uses need clinician sign-off sets out the line.

Which tools to look at, and what to ask them

I have not ranked products here, because the fit depends on your record system and payer mix. When you compare tools, ask these questions:

• Does the draft come from this session's content, or from a template plus last visit's note?
• Can I set a rule that empty sections stay empty?
• Does it flag overlap with earlier notes?
• Does the note go into the clinic's record system as a draft awaiting sign-off, with the therapist's name and time-stamp on the final?
• Where is the audio kept, and for how long?

Two sibling tutorials cover tools directly: four PT scribe tools compared and physiotherapy clinic software compared for built-in AI. A rolling summary of how payers view AI-written notes is in will payers accept AI-written physical therapy documentation.

Billing links: time, codes and the claim

Documentation and billing travel together. A note that says "therapeutic exercise" with no minutes, or a time entry that does not match the note, is a second reason for a payer to look closer. For Medicare, the treatment note for each day must give the date, each intervention provided and billed in language that can be compared with the claim, the total timed-code minutes and total treatment time, and the signature of the professional who furnished or supervised it; the manual says the billing and the total timed minutes must be consistent. Before a claim goes out, a billing check should compare minutes in the note with minutes on the charge, and units billed with the interventions listed. AI can run that comparison on a day's notes and flag mismatches for a human. The method is in physical therapy billing: AI checks for claims and time entries. Treat the flags as questions, not corrections; the therapist decides what happened.

A realistic failure and how it showed up

A clinic set up a drafting tool with a saved template for post-operative knee patients. The template contained a default "Response" line and a default "Plan" line. Therapists added measures and interventions, then accepted the rest. Three months later an internal audit found that 41 of 55 sampled notes for the same patient group had the identical response and plan lines. Nobody had cloned a note on purpose; the template did it for them.

The fix had three parts: delete the default response and plan text so the fields start blank, require a free-text sentence in each, and add the monthly similarity review. A first audit after the change found three notes with repeated phrasing, which the lead raised in a team meeting. This is an illustration of the mechanism, so check your own templates for default text first. It is the cheapest fix in the whole process.

What to do if you find cloned notes already

Stop adding to the problem first: change the template and the process now. Do not alter or backdate old entries to make them look different, because alterations to a record can be worse than the repeated wording. Take advice from your compliance adviser or a billing specialist on what to do about past claims, and keep a short written record of the corrective steps and the date you made them. That record shows an auditor that the clinic saw the problem and fixed it.

If you would like help mapping where your documentation process creates repeated text and choosing one AI step worth adding, that is what my AI implementation consultation covers, at the contact page.

Cloned notes and AI in a PT clinic

Is using a note template the same as cloning?

No. A template that gives structure and prompts is normal. The problem is identical or near-identical wording carried from visit to visit or patient to patient, which makes the record unable to show what happened in this session. Templates need patient-specific content added.

Does an AI scribe prevent cloned notes?

It can help, because it works from what was said in the session, but only if the therapist reviews the output. A scribe fed by a fixed template, or one that fills gaps with typical wording, can still produce notes that read alike. Compare notes across visits to find out.

How many notes should I audit each month?

A small clinic can start with five notes per therapist per month, picked at random across different patients, plus any note that took under a minute to review. Raise the sample if you find a pattern. The point is to catch drift early.

What should I do if cloned notes already exist?

Do not edit old records to disguise them. Stop the practice going forward, take advice from your compliance adviser or billing specialist on past notes, and document your corrective process. Alterations to a record after the fact create their own risk.

Further reads

• Physical Therapy Policies and Procedures: A Clinic Manual Drafted With AI — Put the documentation rules into your clinic manual.
• Physical Therapy Patient Retention: AI Messages Between Visits — Between-visit messages that also need to avoid sameness.
• Creating Home Exercise Programmes With AI for Physio Patients — Individual home programmes, the same principle.
• AI Scribes for Doctors: Six Note Errors to Catch Before You Sign — Six note errors to catch before signing.
• AI Scribe for Occupational Therapists: A Buyer's Checklist — Buyer questions that carry over from OT.
• Using ChatGPT for Physical Therapy: Referring-Doctor Updates and Outreach — Referring-doctor updates that draw on accurate notes.
• AI for Group Therapy Practices: Notes, Consent and Review Steps — Write one note per member from a group session without leaking other members' details: the rule, the consent wording and the review steps.
• AI for Occupational Therapy Documentation: A Monthly Audit Checklist — A monthly audit workflow for OT practice owners: sample de-identified notes, score them with a rubric, list gaps by clinician, re-audit.
• 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: CMS Medicare Program Integrity Manual chapter 3 (section 3.3.2.1.1) and Medicare Benefit Policy Manual chapter 15 (section 220.3), read November 2026; National Government Services documentation note on cloned documentation (archived copy). Confirm against each payer's current manual.