AI Agents for Physical TherapyClinics in 2026

Almost every piece of automation marketing aimed at outpatient rehab is built on a model borrowed from primary care: a patient books, a patient attends, a claim goes out, the money arrives. Fill more slots, lose fewer of them, and the practice does better. That model is wrong for physical therapy in a way that changes which software is worth buying.

A plan of care is a standing series — typically two to three visits a week for four to twelve weeks against a specific therapist, often in overlapping treatment slots where one therapist runs two or three patients at once. A cancellation does not cost you one visit. It costs a slot in a series, and it pushes the certification clock. The revenue event is not the booking. It is the completed episode, billed cleanly, inside three separate windows that expire independently of each other.

That is the whole opportunity, and it is unglamorous. The highest-yield thing an agent can do in an outpatient rehab clinic is not write notes, answer the phone or fill the schedule. It is to watch three counters that no human in the building is formally assigned to watch, and to surface them, per patient, every morning, before the visit happens rather than after the denial arrives.

The three depleting clocks in an outpatient rehab episode, plus the review threshold. Dollar figures are reported rather than primary — see the limitations section.

Read that table as a systems diagram rather than a compliance list. Two of the four rows are pure calendar arithmetic derivable from data your chart already holds. One is a counter whose decrement rule belongs to somebody else. One is an accumulator whose definition — which dollars count — is the single question most likely to be answered wrongly in an implementation. Everything in this guide follows from that shape.

The certification chase is the specialty's defining administrative workflow, and it is a strange one: the therapist writes the document, and then spends weeks trying to get a physician to sign it. The cadence is set out in the Medicare Benefit Policy Manual, Chapter 15, section 220.1.3, and it is worth knowing precisely, because the deadlines are what an agent computes.

• Initial certification duration: Covers the duration of the plan of care, or 90 calendar days from the date of the initial treatment, whichever is less.
• Timeliness — 30 days: Timely certification of the initial plan is met when physician or NPP certification is documented, by signature or verbal order, and dated in the 30 days following the first day of treatment, including the evaluation.
• Verbal orders — 14 days: If the order to certify is verbal, it must be followed within 14 days by a signature to be timely. That is a second, shorter clock nested inside the first, and it is the one most often missed.
• Recertification — every 90 days: At least every 90 days after initiation of treatment under that plan, or sooner if a significant modification becomes evident. A physician may certify for any duration up to 90 days.
• Who may not certify: Chiropractors may not certify or recertify therapy plans of care. Optometrists may certify only low vision services. A workflow that routes a certification request to the wrong provider type produces a document with no effect.

Then there is the change that most reshapes what software is worth building. Reported effective for dates of service on or after January 1, 2025: when a patient is referred by a physician or NPP, certification is satisfied by a signed and dated order or referral in the record, plus documentation that the plan of care was submitted to the referring provider within 30 days of the initial evaluation. The referring provider need not return a signed copy. Silence operates as assent. APTA cites 42 CFR Part 424 Subpart B and 89 Fed. Reg. 97710, 97912-97918.

If it holds as described, the operational consequence is large: the payable artifact stops being a returned signature and becomes proof of transmission. That is a document-rail problem rather than an API problem — fax confirmations, portal receipts, direct-message logs — and it is precisely the kind of evidence software produces reliably and humans lose. An agent that transmits the plan of care, timestamps the transmission, names the recipient and channel, and files the artifact immutably is doing something with direct revenue consequence and no clinical judgement in it at all.

What the agent must not do is decide that the plan of care is clinically appropriate, or that a modification is significant enough to trigger early recertification. Those are the therapist's judgements, and the physician's agreement is a clinical act that software cannot manufacture. The agent's job is to make sure nobody has to remember a date.

The second clock belongs to somebody else. Commercial payers and their delegated utilization-management vendors authorize outpatient rehab in blocks of visits, and the re-authorization has to land before the block runs out or the visits are simply unpaid. The clinic's exposure is a counting problem across dozens of concurrent episodes, each with a different payer, a different block size and a different submission channel.

A lot of 2026 vendor content implies this problem is about to be standardised away. It is worth being precise about what is actually happening, because the distinction between what binds payers and what helps practices is where budgets get wasted.

The 2027 API is a payer-side obligation. A practice benefits only if its EHR or its vendor builds a client against it. Payer readiness is not practice capability.Ask any vendor which specific payers' FHIR prior-authorization APIs it will consume, and get the answer in writing with a date attached.

The adoption data agrees. The 2025 CAQH Index, published by DataSpring (formerly CAQH) and released February 19, 2026, puts fully electronic prior-authorization adoption at 40%, up from 31% in the 2023 Index, built on data from more than 600 provider organizations and health plans representing 63% of insured lives. The corollary is the number that should shape your plan: roughly 60% of prior authorizations are still not fully electronic. Four different things are sold as "ePA" — the X12 278 transaction, HL7 FHIR and Da Vinci, portal automation, and fax or phone. Make a vendor tell you which one it means, per payer.

One more thing that touches outpatient rehab directly. WISeR reaches the office setting — place of service 11 — not just hospitals, and its selected services include epidural steroid injections for pain management and percutaneous vertebral augmentation, which is the referral population many rehab clinics share with pain management and physiatry. WISeR is live: implemented January 1, 2026, it survived a GAO rule determination and a Congressional Review Act repeal resolution that the Senate rejected in late July 2026, with a House appropriations rider still pending. It is politically unstable and legally operative. It also runs a real federal gold-card equivalent: from July 2026, CMS and the participants automatically exempt providers from prior authorization and pre-payment review on demonstrated compliance — at least ten prior-authorization requests in an assessment period, plus a minimum affirmation rate that each participant sets and posts publicly. The exemption is granted at the individual NPI level, added quarterly, and held for at least a year. CMS does not publish the threshold number.

Note the counterweight before you build a strategy on gold carding. Medscape's reporting on state-mandated gold-card programs is headlined on the finding that experts say they offer little relief. Exemption programs are worth tracking; they are not a plan.

The third clock is an accumulator. Reported figures for CY2026: a $2,480 KX modifier threshold for physical therapy and speech-language pathology combined, and $2,480 for occupational therapy separately, up from $2,410. Claims above the threshold submitted without the KX modifier are denied. A separate $3,000 targeted medical review threshold is not indexed until after 2028. We mark these as reported: the CMS therapy-services page returns an error to automated retrieval, so the numbers come from APTA's payment-thresholds page, which cites the Medicare Claims Processing Manual Chapter 5, section 10.3.2.

The rest of the billing mechanics are, by contrast, exactly the kind of deterministic arithmetic that software should have been doing for a decade. From the Medicare Claims Processing Manual, Chapter 5:

1. 1.The 8-minute rule: 1 unit for 8 through 22 minutes; 2 units for 23–37; 3 units for 38–52; 4 units for 53–67; 5 units for 68–82; 6 units for 83–97. Units billed must reconcile to timed minutes documented. This is a table lookup, and an agent that runs it before the claim drops catches the mismatch while the therapist can still correct the note.
2. 2.Discipline modifiers GP, GO, GN: Required on therapy services — GP for a PT plan of care, GO for OT, GN for SLP — and a claim line may carry only one. A presence-and-uniqueness check, trivially automatable, and a recurring source of avoidable rejections.
3. 3.CQ and CO assistant modifiers: Required since January 1, 2020 for services furnished in whole or in part by a PTA or OTA, paired to GP and GO respectively. Since January 1, 2022 those lines are paid at 85 percent of the otherwise applicable Part B amount, with the 15 percent reduction taken last, right before sequestration. Unpaired CQ or CO claims are rejected and returned as unprocessable — which means an unpaired modifier is not a denial you appeal, it is a claim that never landed.
4. 4.Multiple procedure payment reduction (MPPR): Reduces the practice-expense component when more than one unit or procedure is furnished to the same patient on the same day. It applies to multiple units as well as multiple procedures, and it applies across disciplines — PT plus OT plus SLP on the same day. Full payment goes to the unit with the highest practice expense. An MPPR-aware expected-payment estimate is one of the few genuinely useful forecasting outputs available to a rehab clinic.

Everything in that list is computable from data the chart already holds, requires no write access, and produces a result you can count inside one billing cycle. That combination is rare, and it is why we push clinics toward pre-submission checking before anything more ambitious.

One structural note for practices whose referral base includes physiatry. Orthopedic surgery, physical medicine and rehabilitation, pain management, anesthesiology, interventional pain management and neurosurgery are the six Part B specialty codes in the low back pain cohort of the Ambulatory Specialty Model — a mandatory Innovation Center model running January 1, 2027 through December 31, 2031, with an eligibility threshold of 20 or more attributed episodes, two-sided risk, and scoring that includes meaningful use of certified EHR technology. It applies to the attributed clinicians, not to the therapy practice directly, but it changes the interoperability conversation with your referring physicians, and it is the most defensible reason we know to bring them into an integration project rather than build one in isolation.

Because a PT episode is a series, the economics are about completion rather than capture. A patient who attends four visits of a twelve-visit plan has consumed an evaluation, a certification cycle and an authorization, and produced a fraction of the episode's value — while occupying a recurring slot that could have carried someone through to discharge. Attendance management, cancel and no-show backfill, and the distinction between a clinical discharge and a silent dropout are the operating levers that matter most.

And here is where we have to stop and refuse to give you a number.

Naming the absence is more useful to you than repeating the number, for a concrete reason: if there is no credible industry benchmark, then your own baseline is the only number a business case can rest on. Measure it before you automate anything. Visits scheduled versus attended, by therapist and by day of week. Episodes that reach documented discharge versus episodes that simply stop. Days from initial evaluation to certification. Re-authorizations submitted before the block ran out versus after. Four weeks of that data costs nothing but attention, and it converts every later ROI conversation from an argument into arithmetic.

On the automation itself: reminders, waitlist backfill and re-engagement outreach are real and useful, and they are also where a rehab clinic most easily walks into state disclosure law. Keep the line clean. Confirming an appointment, telling a patient how many visits remain on an authorization, or asking them to call the front desk is administrative — California's AB 3030 expressly excludes appointment scheduling, billing and other clerical or business matters from its definition of patient clinical information. A generative message that discusses the patient's pain, progress or what they should do about a flare-up is clinical, and in California it needs a generative-AI disclaimer and instructions for reaching a human provider unless a licensed provider read and reviewed it first. In Texas, HB 149 §552.051(f) reaches further: any AI use in relation to a health care service or treatment must be disclosed to the patient no later than the date the service is first provided, with no human-review exemption on its face.

And a naming rule that catches people out: in California, holding an AI out with a title implying a health care licence is enforceable against the entity deploying it, with each use a separate violation. Calling your reminder agent "Nurse Ava" is not a branding decision.

This is the section most likely to be wrong in every other article you read on this topic, for two reasons: it usually mixes payer-side and provider-side products into one list, and it usually ranks on accuracy figures that nobody has audited. We are going to do neither, and explain why before we rank anything.

Now the distinction that saves money.The most expensive common mistake in this category is buying a payer-side tool. Several of the best-funded, best-marketed "AI prior authorization" companies do not sell to practices at all — they sell to health plans, and their product is the thing on the other side of your request. A practice cannot deploy them. Some of them are, quite literally, adjudicating your requests.

Payer-side vendors that appear in provider-facing "top prior-auth software" lists. Checked 2026-08-12.

That pattern is stacked most deeply at Optum, whose provider-facing Digital Auth Complete went live in January 2026 and is powered by Humata Health — the WISeR participant adjudicating Medicare prior authorizations in Oklahoma — inside UnitedHealth Group. A practice buying prior-authorization software there is buying from the parent of its largest payer, running on the engine of a Medicare prior-authorization adjudicator. None of this is illegal, all of it is publicly disclosed, and we found it disclosed in no vendor comparison article. Ask about it directly, in the room, and write the answer down.

The provider-side ranking. It is nine, not ten. After removing payer-side products, autonomous-coding products that are a different category, and one enterprise platform with no evidence it sells to independent practices, nine credible provider-side candidates remain. We are not backfilling a tenth to reach a round number, because the only way to do that would be to include a vendor we could not verify is still actively selling — and that is exactly the failure this section exists to avoid.

Provider-side prior-authorization and RCM automation platforms, ranked on verifiable attributes only. Checked 2026-08-12. Rankings are opinion; the inputs are public.

The finding that matters most to a rehab clinic is what is missing from that table. Sixteen vendors were examined and one — Candid Health — publishes a fully verifiable named integration list. No rehabilitation-specific EHR appears anywhere in that one verifiable list, and for the other fifteen we could not verify a named EHR integration at all, so whether any of them connects to your rehab EMR is simply unestablished on the public record. Not one publicly states SOC 2 Type II or HITRUST status on the pages we fetched. And exactly one published price fact exists across the entire roster: Availity Essentials is free to providers for sponsoring payers. Sixteen vendors, zero published dollar prices otherwise. Treat every claim about fit with your specific rehab EMR as unverified until the vendor demonstrates it in your instance, with your data.

Four categories were deliberately excluded rather than ranked. Payer-side products (above). Autonomous coding platforms such as Nym Health and CodaMetrix — healthy companies, different category; CodaMetrix has the most credible provenance in the roster, having spun out of Mass General Brigham, but that provenance is not a published study and we found no peer-reviewed evaluation. Enterprise health-system RCM such as AKASA, where we found no evidence of sales to independent practices. And 2026 entrants— R1 RCM's Phare Access (launched January 15, 2026; R1 was taken private in November 2024 by TowerBrook and Clayton, Dubilier & Rice, so no public financials), Optum's Digital Auth Complete, UiPath's healthcare agentic push announced at ViVE in February 2026, and PrescriberPoint (April 2026) — all new, none independently evaluated. New is not disqualifying. Unproven and unverifiable together are.

Finally, the figures we will not print, and the reason for each. This table is the part of the section we would most like other publishers to copy.

Figures excluded from this article, with the reason. Checked 2026-08-12.

Against all of that, here is a statistic that does survive sourcing, because it comes from a professional association surveying its own members with a stated methodology. The 2025 AMA Prior Authorization Physician Survey — a nationwide survey of 1,000 practicing physicians, 400 primary care and 600 specialists, fielded in December 2025 and released May 2026 — reports 40 prior authorizations per physician per week and 13 hours of physician and staff time weekly, with 40% of practices having staff working exclusively on prior authorization. And the number that should sit at the front of every vendor meeting: 60% of physicians are concerned that AI increases or will increase prior-authorization denial rates, while the entire vendor category sells AI as the fix. That tension is real, it is sourced, and it belongs in your diligence.

This is the single most important architectural fact in the whole cluster, and it is the one most likely to be glossed over in a demo. The ONC/ASTP Certification Companion Guide for §170.315(g)(10), last updated May 15, 2026, states it directly:

The § 170.315(g)(10) certification criterion requires Health IT Modules to support API-enabled "read" services for single and multiple patients. … These services specifically exclude "write" capabilities, where authenticated and authorized third-party applications would be able to create or modify EHI through a secure API.

— ONC/ASTP Certification Companion Guide, §170.315(g)(10), updated 2026-05-15

So an agent can read the chart through a certified API in any certified EHR. Every write — booking the next visit in the series, posting the charge, filing the note, updating the plan of care, marking the certification received — happens one of three ways, and you must know which one your vendor means:

1. 1.A vendor-discretionary write API: A commercial arrangement with your EHR vendor, governed by that vendor's app review and API terms, frequently with per-transaction fees. There is no certification requirement compelling any EHR vendor to offer this. It is a negotiation, and it has a price and a lead time.
2. 2.Robotic automation driving the user interface: A bot typing into your EHR or a payer portal. Ask directly whether it runs under a named human's credentials, because if it does, your audit trail says your employee did it. Ask how many portals the vendor maintains and what its mean time to repair is after a portal redesign. No vendor publishes that number.
3. 3.A human: Slower, and frequently the correct answer for anything with billing, legal or clinical consequence. The agent assembles and proposes; a named person commits.

There is a security dividend hiding in this constraint that is worth taking deliberately. An agent that cannot write cannot be injected into writing. Prompt injection is unsolved — a controlled simulation published in JAMA Network Open in December 2025 found attacks succeeded in 94.4% of 108 evaluations at turn four and persisted in 69.4% of follow-ups, with flagship models showing high susceptibility; and work published in Nature Communications found that sub-visual prompts embedded in medical imaging data can cause harmful output and are non-obvious to human observers. Those are simulations rather than field data, and we found no published study of injection through referral faxes, portal messages or payer portals in a live practice — the absence of that evidence is itself worth stating. But the mitigation that survives all of it is architectural: least privilege at the API boundary, read-only by default, a human signature as the only path from draft to record, unique agent identity so a successful attack is reconstructable, and out-of-band confirmation for anything irreversible. Note the uncomfortable corollary of the "non-obvious to human observers" finding — a human in the loop who cannot see the injection cannot review it away. Frame this as blast-radius reduction, never as solved.

Two HIPAA mechanics follow from giving an agent its own access. 45 CFR §164.312(a)(1) already frames access rights as attaching to persons or software programs, and §164.312(a)(2)(i) makes unique user identification a required, not addressable, implementation specification. A shared service account that makes agent actions indistinguishable from a human's is the compliance failure to name and avoid. And your model provider, your agent-framework vendor and that provider's own cloud host are all business associates or subcontractor business associates under 45 CFR §160.103 — one BAA with the app vendor does not close the chain. Whether a given model provider will sign a BAA, and on which tier, changes without notice; re-check the provider's own current documentation on the day you sign, not on the day you read an article.

Write this down before anything is built, get your compliance lead to sign it, and put it in the vendor contract. The boundary is not a philosophical position; most of it is set by statute and manual, and the parts that are not are set by what you can afford to have go wrong unattended.

The autonomy boundary for an outpatient rehab clinic. Authorities cited in the sources list.

The organising principle behind the whole table is a single statutory sentence. 21 U.S.C. §360j(o)(1)(A) excludes from the device definition software intended for administrative support of a health care facility, and it names the things: processing and maintenance of financial records, claims or billing information, appointment schedules, business analytics, practice and inventory management, determination of health benefit eligibility. An agent doing those things is outside the device definition by statute, not by FDA grace. The moment it starts producing clinical recommendations, a different and much harder four-criterion test applies under §360j(o)(1)(E), and FDA's Clinical Decision Support Software guidance reissued January 29, 2026 does not make that easier so much as it announces an enforcement-discretion policyover one criterion failure. Write it that way, because discretion is revocable without notice-and-comment and an exclusion is not. FDA said nothing new about generative AI specifically; do not let a vendor tell you otherwise, and treat any claim that FDA "approved" or "cleared" an administrative AI product as a red flag on its face.

One more duty that is live today and routinely missed: 45 CFR §92.210 imposes an ongoing duty on covered entities to make reasonable efforts to identify uses of patient care decision support tools that employ input variables measuring race, colour, national origin, sex, age or disability, and to mitigate the resulting discrimination risk. It applies whether or not the tool is AI and whether or not FDA calls it a device. If your agent prioritises a worklist, that prioritisation is in scope for the inventory.

Seven phases, in order, with an owner, an entry criterion, an exit criterion and a defined response when the phase fails. The sequencing is deliberate: everything read-only comes before anything that writes, and everything with a countable result comes before anything with a narrative one. Week ranges assume a single-site clinic of five to fifteen therapists and overlap by design.

Two notes on economics. The integration layer is largely a fixed cost paid once and reused: the first agent pays for the connection, the identity, the audit trail and the review queue; the third inherits all of it. And the sequencing above deliberately front-loads the workflows whose results are countable in a single billing cycle, because a clinic that cannot show a result in a quarter will not fund a second phase.

These are the failure modes specific to outpatient rehab, in roughly the order we see them appear. Note that almost none of them is a model failure. They are integration failures, definition failures and trust failures — which is where these projects actually die.

Failure modes, detection signals and rollbacks for an outpatient rehab agent deployment.

Twelve things that should slow a purchase down. None is automatically disqualifying on its own; three together usually are.

• An accuracy or auto-approval percentage on the homepage: There is no independent benchmark in this category. Every such figure is vendor-published with no denominator. Ask what the denominator is, which payers, and which service lines. The answer is more informative than the number.
• Two-decimal precision on an internal metric: A '99.24% clinical accuracy' claim signals an unaudited internal measurement dressed as science. So does a headline metric that contradicts itself inside a single press release.
• Cannot say which of the four ePA channels it means, per payer: X12 278, FHIR/Da Vinci, portal automation, or fax and phone are four different products. A vendor that answers 'we handle prior auth' has not answered the question.
• 'Writes back into your EHR' without naming how: Certified API access is read-only. Make them say whether it is a commercial write integration, UI automation, or a human. If it is UI automation, ask whose credentials it runs under.
• 'Supports all EHRs' with no named systems: An unfalsifiable claim on the exact axis that determines whether the product works for you. One vendor in the roster we examined makes this claim on the page whose entire purpose is to list integrations.
• An AICPA logo or a 'HIPAA compliant' badge offered as a security certification: An AICPA mark is not a SOC 2 Type II report, and HIPAA compliance is a property of a covered entity's practices and contracts, not of software. Ask for the SOC 2 Type II report and the BAA before the demo, not after the contract.
• No disclosure of payer ownership or adjudication relationships: Ask directly: do you or your parent adjudicate prior authorizations for any payer we bill, including under the CMS WISeR model? Do payers hold equity in you? These are answerable questions with public answers.
• A single BAA offered as covering the whole chain: Your model provider and its cloud host are subcontractor business associates. A BAA with the app vendor alone does not close the chain, and 'we train on your data' terms collide with the rule that a business associate may not use PHI in a way the covered entity could not.
• Reluctance to run a retrospective on your own historical data: The only meaningful evaluation available is against your last few hundred claims, your worst faxes and your real payer mix. A vendor unwilling to be measured that way is telling you something.
• Claims that FDA 'approved' or 'cleared' an administrative AI product: FDA's 2026 CDS guidance grants enforcement discretion over a criterion failure. Enforcement discretion is not clearance and is revocable. The word 'approved' in this context is a competence signal.
• A prompt-injection 'guardrail' effectiveness rate: No independent benchmark exists for clinical prompt-injection defence. Any number here is marketing. Ask instead about least privilege, agent identity, audit logging and what the agent physically cannot do.
• Pressure to decommission the manual process at go-live: Run both until the detection signals hold. A vendor whose commercial model depends on you switching off the fallback immediately has misaligned incentives with your revenue cycle.

These are the bands Frenchy Digital uses to scope healthcare operations AI work in 2026. They assume integration feasibility and baseline measurement are in scope from the start, because retrofitting either one is what turns a fixed-price project into an argument.

Frenchy Digital cost bands for physical therapy and outpatient rehab AI agent engagements, 2026.

Senior-led delivery runs $150–$225 per hour, and ongoing retainers run $2,500–$9,500 per month covering model and dependency upgrades, evaluation expansion, incident response and a quarterly technical review. Every engagement carries a 30-day post-launch warranty, and you receive a written scope with a fixed-price phased proposal within 5 business days of the discovery call.

A budgeting note specific to rehab. The three-clock watcher is the cheapest thing on this list and the only one whose payback is countable in a single billing cycle, because a denial avoided is a denial you can name. Sequence it first, prove it, and fund the rest out of what it returns. Clinics that pilot three disconnected point products in parallel get materially worse economics than clinics that sequence two or three workflows through one integration.

A short, unflattering list. If you are building a business case, build it on this rather than on the confident parts.

• The 2026 threshold dollar figures are reported, not primary: CMS's therapy-services page returns an error to automated retrieval. The $2,480 KX threshold and the $3,000 targeted-review threshold come from APTA, which cites the Claims Processing Manual. High confidence, not primary. Confirm with your MAC before configuring an accumulator.
• The 2025 certification exception is not yet in the manual: The Benefit Policy Manual PDF still carries older §220.1.3 text. The Federal Register pages are the controlling authority. We could not confirm the exception against a current manual revision, so treat it as reported and cite the Federal Register.
• Which dollars count toward the threshold: We did not verify from a primary source whether the accumulation runs on submitted charges or allowed amounts, or exactly how a practice should treat mid-year payer changes. This is the definition most likely to be implemented wrongly, and it is a question for your MAC and your billing lead, not for a vendor.
• No exact Medicare payment amounts appear in this article: We did not fetch the Physician Fee Schedule Addendum B, so we print no per-code dollar figures. Any vendor ROI model built on per-code payment assumptions should show you its source.
• Security certifications for every vendor in the ranked table: Not one of the sixteen vendors examined publicly states SOC 2 Type II or HITRUST status on the pages we fetched, and none advertises a HIPAA BAA — although, handling PHI, they must execute them. We could not turn this into a table column with real values, so we turned it into buyer guidance instead.
• Named EHR integrations for most of the roster: Only Candid Health publishes a fully verifiable named list. Adonis's integrations URL returns a 404 and only 'Epic' is verifiable as page text. Every other integration claim in this category is the vendor's own assertion, and we found no EHR-side listing corroborating any of them.
• The read-only certification fact was not re-fetched this session by every source: The §170.315(g)(10) read-only characterisation rests on the ONC/ASTP certification companion guide updated May 15, 2026. It is settled and uncontroversial, but note that a proposed ASTP/ONC rule published December 29, 2025 would remove the clinical decision support certification criterion and reserve §170.315(a)(9). It remains proposed; we located no final rule.
• Colorado's AI law has a contested effective date in circulation: SB 26-189, signed May 14, 2026, repeals and reenacts the Colorado AI Act. Its enrolled text sets an effective date of January 1, 2027 and applies to consequential decisions made on or after that date, with a healthcare carve-out for HIPAA covered entities plus a general patient notice of use of advanced technologies. An August 12, 2026 date circulates in secondary trackers, derived from a site-wide banner about bills passed without a safety clause; this bill has one. We use the enrolled text. We also found no verified litigation against it and do not repeat that claim.
• State utilization-review AI statutes could not be cited: A growing number of states, including Texas, Arizona and Maryland, restrict payers from using AI as the sole basis for a medical-necessity denial. We could not verify the individual statutes, citations or effective dates, so we name no bill numbers. The same applies to the gold-carding statutes outside Texas.
• No PT-specific completion or dropout benchmark exists that we could source: See the attendance section. Your own baseline is the only defensible number.
• Prompt injection through the channels a clinic actually uses is unstudied: The published work covers patient dialogue and medical imaging. We found no published study of injection via referral faxes, portal messages or payer portals in a live practice. Extrapolating is analysis, not evidence — and the absence of evidence is not reassurance.
• One classic productivity figure, correctly caveated: The often-quoted finding that for every hour of direct clinical face time there are nearly two additional hours on EHR and desk work comes from Sinsky et al., Annals of Internal Medicine, 2016 — 57 physicians, 430 observed hours, four specialties, four states, with the authors noting the data came from self-selected, high-performing practices and may not be generalizable. It is a decade old and it did not study physical therapists. We cite it only as context, never as a current measurement of your clinic.

None of this argues against building. It argues for building one read-only workflow with a measured baseline, a named write path, and a human boundary written down before the first line of code. The clinics that get value from agents are the ones that instrumented the before-state and picked a workflow where the result is countable in a quarter — which, in outpatient rehab, is almost always the three clocks.

A physical therapy clinic is not a booking business with a clinical layer on top. It is an episode business running against three independent expiry clocks, in which the most expensive events are silent: a certification that lapsed, a block that ran out two visits ago, a threshold crossed without a modifier. None of those is a hard problem for software. All of them are unowned.

So the sequence is: measure the before-state, get a written answer on the write path, build the read-only clock watcher, make proof of transmission a retained artifact, then work outward to authorization packets and pre-submission checking. Keep the KX attestation, the signature and every clinical judgement with a licensed human, because statute and manual put them there and no vendor configuration moves them. And when a vendor hands you a percentage, ask for the denominator — in a category with no independent benchmark, that question is worth more than the number.

A one-page diligence checklist.Take this into the vendor meeting. Every question has a verifiable answer, and a vendor's willingness to answer is itself data.

1. 1.Who do you sell to?: Payers, providers, or both. If both, ask whether the payer business touches any plan you bill, and whether the company or its parent participates in the CMS WISeR model.
2. 2.Name your integrations, and show me the page: Not logos. Named systems, on your own site, that I can read as text. Then show me a corroborating listing on the EHR vendor's marketplace.
3. 3.Which of the four ePA channels do you use, per payer?: X12 278, FHIR/Da Vinci, portal automation, or fax and phone. Give me the list by payer for my top ten by volume.
4. 4.How do you write into my EHR?: Commercial write API, UI automation, or a human. If UI automation, whose credentials? How many portals do you maintain and what is your mean time to repair after a portal redesign?
5. 5.Send me the SOC 2 Type II report and the BAA before the demo: And the subcontractor list, including your model provider and its cloud host.
6. 6.Will you run a retrospective on my last 500 claims?: On my payer mix, my worst faxes and my rehab EMR. Before contract, not after.
7. 7.What is the denominator on every percentage in your deck?: Which payers, which service lines, what period, who audited it.
8. 8.What can your agent physically not do?: Give me the tool list and the permission scope. I want the answer to include the KX modifier, the signature, and submitting to a payer as the submitting party.
9. 9.Which specific payers' FHIR prior-authorization APIs will you consume in 2027?: Names and dates, in writing. Payer readiness is not your capability.
10. 10.What is your rollback?: When your integration breaks, what happens to my worklist, how fast do I know, and who tells me?

If a vendor answers all ten without hedging, you are dealing with a serious company. If it answers three and reframes the rest as roadmap, you have learned something worth more than the demo.

One last word on where the liability sits. Nothing in federal law lets software be the licensed professional. Medicare's "incident to" rules define auxiliary personnel as an individual who meets state licensure requirements — software is not an individual, cannot be excluded by the OIG, and cannot hold a licence, so there is no incident-to pathway for AI-performed work. Corporate-practice statutes say artificial legal entities have no professional rights, privileges or powers. And Texas put the whole doctrine in five words when it permitted practitioners to use AI for diagnostic purposes only while acting within the scope of their licence, "regardless of the use of artificial intelligence," and only if the practitioner reviews all records created with it.

The licence does not move. Neither does the liability. Build accordingly, and the rest of this is just engineering.