AI Documentation in PT: What Medicare Requires of the Note
In July 2025, CMS updated MLN905364, “Complying with Medicare Signature Requirements,” to add a section on artificial intelligence. The answer is short: if you use a scribe, including artificial intelligence technology, you sign the entry, and “you don’t need to document who or what transcribed the entry.” The scribe, human or AI, does not sign or date the note.
What the fact sheet did not change is everything else. The therapy documentation requirements in Medicare Benefit Policy Manual §220.3 still apply to an AI-drafted note exactly as they apply to a hand-typed one, and the cloned-documentation problem Medicare’s contractors already police does not disappear because a machine wrote the paragraph. A well-formed AI draft that reads the same on every visit is not a compliance shortcut; it’s the exact pattern reviewers are trained to flag.
A signed business associate agreement (BAA) with an AI vendor is necessary, but it’s not the whole story. It says nothing about whether the patient was told the encounter would be recorded, and nothing about whether the clinician actually reviewed the draft before signing it. Both gaps matter, and both are already being tested, one in a pending lawsuit and the other in medical review. Current as of September 2026.
Does Medicare Allow AI-Generated Therapy Notes?
Yes, with one condition: a human signs. MLN905364 treats an AI scribe exactly like a human one: “If you use a scribe, including artificial intelligence technology, sign the entry to authenticate the documents and the care you provided or ordered. You don’t need to document who or what transcribed the entry,” and “we don’t require the scribe to sign or date the documentation.”
That signature is not a formality. Medicare Program Integrity Manual, Chapter 3, §3.3.2.4 explains why it exists: “to resolve authenticity concerns related to legitimacy or falsity of the documentation.” MLN905364 spells out the consequence: “Medicare claims reviewers look for signed and dated medical documentation from those responsible for providing care to the patient. If the entries don’t meet the signature requirements, we may deny the associated claims.” A missing signature, other than on an order, is fixed with an attestation statement, not a resubmission of the AI draft.
The manual’s section on templates, §3.3.2.1.1, is the closest thing Medicare has to a ruling on AI-generated content: “CMS does not prohibit the use of templates to facilitate record-keeping,” but discourages ones with “limited options and/or space for the collection of information,” because they “often fail to capture sufficient detailed clinical information to demonstrate that all coverage and coding requirements are met.” An AI draft is a template output in the manual’s sense: CMS neither bans it nor blesses it, and a reviewer judges what the note actually says.
What an AI-Drafted Note Still Has to Contain
Medicare Benefit Policy Manual, Chapter 15, §220.3 sets the bar for every treatment note, AI-assisted or not: documentation “must be legible, relevant and sufficient to justify the services billed.” Section 220.3.E lists the elements a treatment note requires, and each depends on a different kind of input. An ambient tool hears what was said in the room; it does not hear minutes per timed code, read a goniometer, or know what the billing system will submit.
| Documentation element | What an ambient/AI tool can capture | What the clinician must enter |
|---|---|---|
| Date of treatment (§220.3.E) | Whatever date the tool stamps on the session | Confirmation the date matches the encounter |
| Identification of each specific intervention/modality provided and billed, “in language that can be compared with the billing on the claim” (§220.3.E) | A spoken description of what was done during the session, if it was said aloud | The exact intervention language matched to the billed code |
| Total timed code treatment minutes and total treatment time in minutes, with “the billing and the total timed code treatment minutes” consistent (§220.3.E) | Nothing usable: how long a recording ran is not a count of timed minutes per code | The clinician’s own count of timed minutes per code, reconciled to units before signing |
| Signature and professional identification of the qualified professional who furnished or supervised the service, and a list of everyone who contributed (§220.3.E) | Nothing; CMS states “we don’t require the scribe to sign or date the documentation” | The clinician signs, dates, and lists contributors |
| Objective measurements showing progress toward goals (§220.3’s general documentation expectations, not a treatment-note element) | Only a value a clinician states aloud during the session, not a device reading | The measured value itself, entered by the clinician who took it |
The manual is explicit that “the treatment note is not required to document the medical necessity or appropriateness of the ongoing therapy services.” That burden sits with the progress report instead, due “at least once every 10 treatment days” under §220.3.D. An AI-drafted treatment note can skip the medical-necessity language, but it still needs the four §220.3.E elements above, and none of them come free from an audio feed. For the full list of Medicare’s five required therapy documents, see our Medicare PT documentation requirements guide.
How AI Drafts Turn Into Cloned Documentation
The failure mode MACs already watch for has a name: cloned documentation. CGS, the Jurisdiction 15 Part B contractor, defines it plainly in guidance dated September 6, 2012 and still posted: “This ‘cloned documentation’ does not meet medical necessity requirements for coverage of services rendered due to the lack of specific, individual information. All documentation in the medical record must be specific to the patient and her/his situation at the time of the encounter,” and warns that “Identification of this type of documentation will lead to denial of services for lack of medical necessity and recoupment of all overpayments made.” The same guidance treats the mechanism as beside the point, listing “copy and paste, cloning, and carry forward, among other terms” as names for the same behavior. The same holds for an AI draft: a model that produces the same well-formed paragraph for every visit is, to a MAC, indistinguishable from copy-forwarded text.
This is not theoretical. First Coast Service Options’s TPE results for outpatient rehabilitation services list top denial reasons as documentation that “does not support medical necessity as listed in coverage requirements,” notes that “did not support the number of therapy minutes/units billed,” missing signed certification, and missing initial evaluation, across review rounds for 97110, 97112, and 97140 that run as recently as 2026. Fluent, consistent AI language makes this pattern easier to produce and harder to catch. See our Medicare PT claim denials guide.
A second, related risk is invention, not repetition. The AMA’s November 2024 principles on augmented intelligence state it without hedging: “the risks of these tools fabricating content are well known.” The same passage notes that “these LLMs are constantly evolving,” so the output for the same fact pattern can change over time. A cloned note repeats what was true before; a fabricated note states something never true at all. Either one fails §220.3’s requirement that documentation be “relevant and sufficient to justify the services billed.”
Is an AI Scribe Vendor a HIPAA Business Associate?
Yes, by the plain text of the definition. Under 45 CFR 160.103, a business associate is “a person who: (i) On behalf of such covered entity … but other than in the capacity of a member of the workforce of such covered entity or arrangement, creates, receives, maintains, or transmits protected health information for a function or activity regulated by this subchapter.” A vendor whose software records the encounter audio, transcribes it, and drafts the note does exactly that, which triggers 45 CFR 164.502(e)(1)(i): a covered entity may let a business associate handle PHI “if the covered entity obtains satisfactory assurance that the business associate will appropriately safeguard the information,” documented in a contract under 45 CFR 164.504(e). No BAA, no lawful basis to touch PHI, regardless of how good the transcription is. Our HIPAA-compliant form builder guide covers what a compliant BAA requires; the same groundwork applies to an AI scribe vendor.
Recording Consent: The Risk a BAA Doesn’t Cover
A BAA covers what happens to PHI once it exists. It says nothing about whether the patient agreed to the recording that created it, and that gap is now the subject of active litigation. On April 8, 2026, a class action was filed in the U.S. District Court for the Northern District of California against Sutter Health and MemorialCare over their use of an AI scribe. TechTarget’s reporting names the vendor, Abridge, and describes claims under the California Confidentiality of Medical Information Act, the California Invasion of Privacy Act, and the federal Wiretap Act: plaintiffs allegedly “did not receive clear notice that their medical conversations would be recorded by an artificial intelligence platform, transmitted outside the clinical setting, or processed through third-party systems.” As of September 2026, the case is pending; these are allegations, and no court has ruled on their merits.
The industry’s own liability carriers have already flagged this risk. Texas Medical Liability Trust, a physician liability carrier, wrote in July 2025 that “several states now require explicit consent from patients for the recording and processing of health care encounters using AI,” and recommended practices “add language to your patient consent forms that addresses AI scribe use” and “explain the technology and its purpose in your informed consent discussions with patients using clear, non-technical language.” TMLT also warns that “uncritical acceptance of AI suggestions carries the risk of errors in the chart,” since AI systems may “hallucinate” or “create information about a patient simply to provide an answer in an EHR template or fill in an information gap,” and recommends practices “set aside dedicated time each day to carefully review documentation” and execute “a Business Associates Agreement (BAA)” before choosing the software.
What APTA and the AMA Have Said
APTA has already treated ambient AI documentation as mainstream enough to warrant formal guidance twice. In September 2025, APTA published a practice advisory on AI-enabled ambient scribe technology, defining such tools as “systems that operate discreetly in the background and use artificial intelligence to automatically capture, transcribe, and summarize patient-provider interactions into structured clinical notes,” and covering “documentation responsibilities and regulatory considerations.” It’s a member-only benefit, so its content beyond that public description isn’t summarized here. It builds on an earlier position: in September 2024, APTA adopted policy P07-24-10-09 on ethical AI integration, supporting “the ethical development and integration of artificial intelligence that reduces administrative burden and enhances physical therapist practice, education, and research for the best interest of patients, clients, the profession, and society.”
Support and caution aren’t in tension. The AMA’s AI principles document, cited above for the fabrication risk, adds: “at the present time, these technologies are largely unregulated, as there is no current regulatory structure for generative AI clinical decision support tools unless they meet the definition of a medical device regulated by the FDA.” No regulator checks the AI’s work before it reaches the chart. The clinician’s review is the only check that exists.
A Pre-Signature Checklist for AI-Drafted Notes
Everything above points to one conclusion: the note is only as defensible as the clinician’s review before signing. Run every AI draft against this list first.
| Check | Why it matters |
|---|---|
| Every intervention listed was actually delivered today, in language that matches the claim | §220.3.E requires intervention language “that can be compared with the billing on the claim” |
| Timed minutes per code and total treatment time are the clinician’s own numbers, not the AI’s estimate, and units follow from them | Billing and timed minutes must be consistent; TPE denials cite unsupported minutes/units |
| Objective measures are today’s actual measured values, not carried-forward language | Cloned values are exactly what MAC medical review targets |
| The note shows what changed and why, and boilerplate that would read the same on any visit is removed | Skilled-care documentation must reflect the day’s clinical reasoning, not a template |
| Nothing in the note describes a test, measurement, or statement that did not happen | TMLT’s hallucination warning and the AMA’s fabrication risk both point here |
| The signing clinician is the one who furnished or supervised the service, and contributors are listed | §220.3.E’s signature and identification requirement |
| The clinician signs and dates the entry; the scribe, human or AI, is not required to | MLN905364: “we don’t require the scribe to sign or date the documentation” |
Questions to Ask Before You Buy
Get answers to these in writing before an AI documentation tool touches a patient encounter:
- Is a BAA executed before any PHI flows through the tool, and does the vendor have its own agreements in place with any AI model provider it relies on?
- Does the practice have written patient notice and consent language for recording, and a real way for a patient to decline?
- Where are audio and transcripts stored, for how long, and are they retained after the note is signed? (Get a straight answer; this is where PHI exposure concentrates.)
- Can the clinician edit the draft field by field before signing, and does the system record that the review actually happened?
- Does the tool prefill minutes, units, or objective values? If so, treat every one of those fields as unverified until the clinician enters it directly.
- Does the practice have a written policy that AI drafts are reviewed before signature, with staff trained on what that review has to catch?
None of these are answered by a vendor’s marketing page; they’re answered by the contract, the consent form, and the practice’s own review policy.
None of this is a case against structured intake easing documentation burden; it’s a case for keeping the clinician in the loop on what the record says. CuratoForm, Rehabilitation Health’s form builder, runs standalone or as part of PrismEHR. Integrated with PrismEHR, raw response answers and AI-generated content, such as subjective documentation drafted from intake responses, can populate clinical notes and power chart alerts, and that AI-generated content is reviewed by a clinician before it enters the record. Rehabilitation Health signs a BAA with every customer and has BAAs in place with its AI vendors. See how this fits physical therapy practices, and read our companion guide on what Medicare’s therapy documentation requires for the elements every note still has to contain.