Will Payers Accept AI-Written Physical Therapy Documentation?

Payers generally judge the note, not the tool that drafted it. An AI-drafted physical therapy note is acceptable when the therapist has reviewed it against what happened, signed it, and the finished record shows skilled care, medical necessity, timed minutes and a plan of care. If it fails any of those, it fails whatever wrote it.

I found no Medicare rule that bans software-assisted notes, and no general rule that requires disclosing a scribe on a claim. Commercial payers and state rules differ, and they change, so this is a position to confirm with each payer you bill, not a guarantee. What is certain is where an audit lands: on the signed record and whether it supports the claim. AI changes how you produce that record, and it changes where mistakes hide.

Follow me on Instagram@sagnikteaches

What an auditor actually looks for

Medicare contractors list the elements of a complete outpatient therapy record: the initial evaluation and any re-evaluations, a signed and dated plan of care with short and long-term goals, certification of the plan, progress reports, treatment encounter notes, and a discharge summary. For each claim period the treatment records should show total treatment time and identify each skilled modality provided. Progress reports must be written by the clinician, and are due on or before every tenth treatment day. Commercial payers usually layer their own medical-necessity criteria on top, often with visit limits and authorisation requirements.

Connect on LinkedInSagnik Bhattacharya

So a reviewer checks, in roughly this order:

Subscribe on YouTube@sagnikteaches

1. Is the plan of care current, signed and certified? No AI involvement; a missing certification sinks the claim before anyone reads the note. For Medicare dates of service from 1 January 2025, the contractor page says a signed and dated referral order plus a one-time transmission of the plan of care to the referring practitioner within 30 days of initial treatment can meet the initial certification signature requirement, so keep proof of that transmission.
2. Does each note justify skilled therapy today? What did the therapist do that a layperson could not?
3. Do the minutes add up? Timed codes need total timed minutes that support the units billed.
4. Does the note connect to goals and progress? Measurable change, or a reasoned explanation for the lack of it.
5. Is it signed by the treating clinician, in a way the payer accepts?
6. Does it look unique to this patient and visit? Identical wording across visits invites a "cloned documentation" finding.

Item 6 is where AI creates a new risk. Models reuse phrasing. A scribe configured with a fixed template can produce ten notes that differ only in a pain score, and that pattern is what reviewers read as copy-forward.

What the therapist still has to do, whichever tool is used

Notice that in every row AI produces words and the therapist produces the attestation. That is the whole acceptance question in one table.

Worked example: one visit, from AI draft to signed note

This is an invented illustration. A therapist treats a 68-year-old patient, six weeks after a total knee replacement, for 45 minutes. The scribe captures the conversation and produces this draft:

S: Pt reports knee is "much better", pain 2/10, walking without difficulty.
O: ROM 0-120. Strength 5/5. Gait normal.
A: Excellent progress. Pt doing very well.
P: Continue therapy 2x/week. Total time: 45 minutes.
CPT: 97110 x3, 97140 x1.

In the session, the patient actually said the knee was "better but still catches on stairs", pain was 4/10 after climbing, flexion measured 112 degrees, quadriceps strength was 4/5, and gait showed reduced stance-phase knee control on the left. The therapist spent 25 minutes on therapeutic exercise, 10 on manual therapy and 10 on gait training. Corrected, the note reads:

S: Pt reports improved pain, 4/10 after climbing one flight; knee "catches"
on descending stairs. Unable to carry shopping up stairs (goal 2).
O: L knee flexion 112 deg (was 98 at eval), extension 0. Quad 4/5 (3+/5 at eval).
Gait: reduced L knee control in stance, improves with verbal cueing.
Treatment: Therapeutic exercise 25 min (closed-chain loading, eccentric step-downs
with load progression); manual therapy 10 min (patellar mobilisation, soft tissue);
gait training 10 min (stance control cueing). Total timed minutes 45.
A: Progressing toward goals; stair descent remains limited by eccentric quad
weakness and needs skilled progression of loading and cueing to reach goal 2
safely. Risk of falls on stairs without it.
P: 2x/week for 4 weeks; progress step-down height; reassess flexion and quad
strength at the next progress report (due by the tenth treatment day).

Compare the two. The draft said "excellent progress" and recorded measurements that matched no one. It had no breakdown by treatment, so the billed units had nothing to support them, and it contained no reason skilled care was still needed. A payer reviewing the draft would see a patient who needed no more therapy. The edited note shows what the therapist did, why only a therapist could do it, and what is left to achieve. The corrections took about four minutes. That four minutes is the product you are actually buying.

Checking the units against the minutes

For timed codes, Medicare counts the total minutes of timed services and converts them into units: 8 to 22 minutes is 1 unit, 23 to 37 is 2, 38 to 52 is 3 and 53 to 67 is 4. In the example, all 45 timed minutes were timed codes, so three units are supportable in total. How they are split between codes depends on the minutes per code, which is why minute-by-code matters. The draft billed 97110 three times and 97140 once, which is four units and would require at least 53 minutes. That would be an overbilling error generated from a template, not by anyone intending it. Other payers may use different rounding rules, so check each contract.

Ask your scribe or practice software to produce minutes per service as a separate field that you complete, not one it infers. A model estimating times from a recording will round and smooth. For more on claim-stage checks, see AI checks for claims and time entries.

Medical-necessity language: before and after

AI drafts drift towards flattering, generic statements. Reviewers want specifics that show a skilled need.

Instruct the tool to quote your measurements and never to write a conclusion you have not stated. A prompt line such as "Use only the objective values I dictated; if a value is missing, write MISSING" gives you a visible gap instead of an invented number.

The audit trail: what to keep and why

If a payer asks how a note was produced, you should be able to answer in minutes. Keep a simple trail:

• The tool and version in use, with the date it was switched on and the vendor's data terms on file.
• Whether the draft is retained. Some systems keep the original AI output and the edit history; others overwrite. Decide which you want, and know which you have.
• Who edited and signed, and when. The signature timestamp should be close to the visit; sign-off weeks later looks like a backlog, not a review.
• Your written review rule, e.g. "every AI-drafted note is read in full and signed within 24 hours", plus the monthly audit result.
• Patient notification. How you told patients about recording, and what happens when one declines.

Electronic signatures are accepted when the system protects them against alteration, and the contractor page asks providers who use an electronic record to send their process for creating the signature plus an example of how it displays once signed, so find that explanation now and file it. Stamped signatures are a different matter: the Medicare Program Integrity Manual says they are not typically acceptable, with a narrow exception for an author with a documented physical disability. Check your contractor's signature page before relying on any shortcut.

Commercial payers: ask provider relations in writing

Commercial plans publish their own clinical policies and sometimes ask for notes through a third-party review company. Most focus on the same questions as above: functional goals, objective measures, progress and a reason for continued care. I did not find a payer policy that treats AI-drafted notes differently, but I did not review every payer, so ask. A short email to provider relations works:

Subject: Documentation prepared with a scribe tool

Hello, we are an outpatient physical therapy practice (provider ID [ID]).
Our therapists use a scribe tool that drafts treatment notes, which the
treating therapist reviews, edits and signs before they are final.
Please confirm whether you have any policy on this, and whether you
require anything additional (disclosure, attestation, or retention of
drafts) when records are requested for review.
Thank you, [your name]

Keep the reply with your compliance file. A "no specific policy" answer in writing is useful evidence of reasonable steps.

Three clinics, three different risk profiles

The same scribe behaves differently depending on who uses it and what they bill. Here are three invented illustrations.

A solo therapist with mostly commercial patients. Visit limits and authorisations dominate. The documentation risk is a note that fails to show progress when the payer reviews the request for further visits. The therapist's best use of AI is to draft the subjective and treatment sections quickly and then spend the saved minutes writing the assessment line in her own words, with measurements. She keeps a one-page record of what her payers require for extended authorisation, and the scribe prompt asks for the same items each time.

A four-therapist practice with a large Medicare caseload. Here the risk is volume: 80 visits a week, each needing timed minutes, signed and dated notes and progress reports on or before the tenth treatment day. The practice adds a rule that no AI-drafted note can be signed unless the minutes-by-code field is complete, and a tenth-visit flag in the schedule prompts the treating therapist, not an assistant, to write the progress report. The monthly audit samples two notes per therapist. Over a quarter they find the scribe regularly writes "patient tolerated treatment well" in the assessment, so they change the template to leave the assessment blank for the therapist to complete.

A clinic that employs assistants and aides. Assistants treat patients under the therapist's supervision and document their own interventions. An AI scribe running in an assistant's session can draft their entries, but the progress report and any assessment of change must come from the therapist. The practice sets scribe accounts by user so the audit trail shows which person reviewed which entry. A therapist never co-signs a block of notes at the end of the week without reading them.

A realistic failure and how it showed up

Say a practice switched on a scribe in March. Three months later a payer requested records for 15 claims for a single patient. The notes looked polished, but all 15 had the same first two sentences, the same assessment paragraph and the same "pt demonstrates good rehab potential" line. Pain scores changed by one point between visits with no explanation, and eight notes listed four units of timed codes for visits that the schedule showed as 40 minutes long. The therapist had signed each within a day, so the signature trail was fine; the failure was in what he had signed without reading closely.

What would have prevented it: a template that left assessment and plan blank, a minutes-by-code field that had to be filled in manually, and a monthly sample that would have caught the unit mismatch in week two. What the practice did afterwards: it paused the scribe for assessments, kept it for subjective and exercise lists, and retrained staff on the review step. It also stopped using the scribe's suggested codes. None of this involved the payer rejecting AI as such; the payer rejected unsupported units and notes that did not show skilled care.

What to do when a payer requests records

Most requests are routine, but treat each as an audit that might escalate. First, send exactly what is asked, complete and in order, including the plan of care, certification, evaluation, progress reports, treatment notes and any minute logs. Second, check each note before it goes: does it still match the schedule, and are signatures and dates intact? Third, if you find an error, do not edit the original record. Add a dated late-entry correction according to your policy and your payer's rules, and keep the original visible. Fourth, if a note's wording looks like a template artefact, resist the urge to rewrite it. Altering records after a request looks worse than imperfect notes.

The audit trail you built earlier is what you send alongside if asked about your process. A one-page description of your review rule, the vendor terms and the monthly audit log answers most process questions without more.

Questions to put to a scribe vendor before you buy

• Is the AI output stored separately from the signed note, and can I see an edit history?
• Can I make minutes by service a required field the AI never fills in?
• Can I switch off code suggestions, or have them show as a separate prompt?
• Does the tool let me leave sections such as assessment blank by default?
• Where is audio stored, for how long, and is it used for training? Some vendors have changed defaults after launch, so ask how you will be told.
• What happens to my data if I cancel?

Vendor pages for scribes mostly promise time saved and compliance support. Treat the compliance wording as marketing until you have matched it to your payers' documentation requirements and your contract. The vendor cannot sign your notes for you, and in an audit its brochure is not evidence.

One habit helps every clinic: read the note the way a reviewer would, with only the note and the plan of care in front of you, and ask whether a stranger could tell why this visit needed a therapist. If the answer depends on what you remember rather than what is written, the note is not finished, however clean the AI draft looked. Doing this on the last note of the day, before signing, takes under a minute and catches most of the problems described above.

A monthly self-audit you can run in an hour

1. Pull 10 AI-assisted notes across therapists, plus 3 non-AI notes for comparison.
2. For each, compare the note with the schedule: visit length, treatments done, units billed.
3. Check each AI note for a changed measurement or an unexplained "no change". If a statement can't be traced to the session, mark it.
4. Compare notes for the same patient across visits and measure similarity. Three identical paragraphs in a row is a flag.
5. Check signature timestamps: how many were signed within 24 hours?
6. Record the findings and any retraining. A dated log shows a functioning compliance programme.

Illustratively, if 3 of 10 notes needed a corrected measurement and 2 had unsupported units, the right response is not abandoning the tool but tightening the review step and the template: remove default text, make minutes a required field, and keep the scribe to draft only subjective and treatment description until measurements are entered manually.

Where AI drafts help most and least

Strongest use: turning a spoken visit into an organised subjective and treatment description. This saves typing and costs little risk, because you are checking facts you remember. The scribe comparison covers the tools.

Middling use: assembling progress reports from charted data. The tool can pull the measurements; you write the assessment.

Weak use: evaluations and plans of care, where clinical reasoning is the product. Draft from your own dictated findings and read as if you were the reviewer.

Avoid: letting a model choose codes or units from the transcript. Your software or biller should check it, and you decide. The related risk, notes that read the same every time, is covered in avoiding cloned notes.

The short version for payers: they accept the note you can defend. Write the policy that says every AI draft is read, corrected and signed, keep the trail, and audit it monthly. If you want help building that routine around your own software, that is what my AI implementation consultation covers; see the contact page. Related tutorials are collected in the AI guide.

Questions therapists ask about AI notes and audits

Does Medicare ban AI-drafted therapy notes?

I found no Medicare rule that bans a note because software drafted it. The documentation requirements apply to the finished record: it must be complete, show skilled care and medical necessity, record timed minutes, and carry the treating clinician's valid signature. A tool that helps produce that record is a drafting aid, not a substitute author.

Do I have to tell payers I use an AI scribe?

I could not find a general rule requiring disclosure on each claim. Contracts, state rules and payer manuals differ, so read your provider agreements and ask the payer's provider relations team in writing. Tell patients plainly, because consent for recording is a separate duty.

Who is liable if the AI writes something false?

The therapist who signs the note. Signing attests the record is accurate and that you provided the service as described. A vendor's accuracy claims do not move that responsibility, so review every draft against what you did.

Can an assistant or aide use the scribe for me?

Assistants can document the treatments they provide and may write elements of a progress report between clinician reports, but Medicare's manual says those are not complete progress reports: the therapist must write one in each reporting period. Any AI draft of a progress report needs the therapist's own review, edit and signature, and a draft prepared for one clinician should not be signed by another.

Further reads

• AI Receptionist for Physical Therapy: Referral and Insurance Questions — Front-desk AI for referral and insurance questions.
• Physical Therapy Patient Retention: AI Messages Between Visits — Between-visit messages that keep patients attending.
• Using ChatGPT for Physical Therapy: Referring-Doctor Updates and Outreach — Referring-doctor updates that fit the documentation trail.
• Physical Therapy Prior Authorisation: Track Visits With AI — Tracking authorised visits alongside documentation.
• How Physiotherapists Use AI to Write Treatment Notes Faster — The day-to-day note-writing workflow, minus the payer angle.
• Physical Therapy Policies and Procedures: A Clinic Manual Drafted With AI — Turn the review rules below into a written clinic policy.
• AI Use in Physical Therapy: Which Uses Need Clinician Sign-Off — A clinic owner's ladder for AI in physical therapy: which uses run alone, which need a clinician to sign, and which stay human.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Noridian JF Part B outpatient therapy documentation requirements; Medicare Claims Processing Manual chapter 5, section 20.2 (timed-code units); Medicare Benefit Policy Manual chapter 15, section 220.3 (progress reports); Medicare Program Integrity Manual chapter 3 (signatures); all read November 2026. Payer rules vary: check your own contracts.