AI in Physical Therapy: What’s Real, What’s Not, and What Actually Matters for Your Clinic

AI in physical therapy is moving fast, but what’s actually useful in a real clinic? Explore the AI tools delivering real ROI, where the evidence is still thin, and the critical mistakes PT clinics should avoid before jumping on the hype.

If you’ve been in practice for more than a few years, you’ve lived through several waves of technology that were supposed to change everything. Some did. Most didn’t, at least not in the ways they were promised. AI in physical therapy is arriving with the same fanfare, and the pressure on clinic owners and rehab directors to do something about it is real and growing.

So let’s skip the hype and have an honest conversation about what AI actually means in a clinical setting today: where it’s delivering genuine value, where the evidence is still thin, and how to think about it without either dismissing it or chasing every shiny new vendor pitch.

First, Let’s Define What We’re Actually Talking About

One of the most frustrating parts of this conversation is the language. “AI-powered” appears on every EMR pitch deck and RTM platform right now, and it doesn’t mean much without context. Here’s a working distinction worth keeping:

Artificial intelligence and machine learning refer to systems that learn from data. They identify patterns across large datasets and improve their outputs over time. Genuine machine learning in PT shows up in computer vision tools that analyze movement, natural language processing engines that power ambient documentation, and predictive analytics that flag patients at risk of dropping off.

Rule-based automation is something different. It follows fixed “if-then” logic: useful, often time-saving, but not intelligence. Many tools marketed as “AI-powered” are doing this kind of pattern matching or template logic. That doesn’t make them bad tools. It just means you should ask the right question before you buy: what is the actual AI component, and what is just automation?

That single question will save you a lot of wasted demos.

The Categories of AI Tools in PT Today

AI Documentation and Ambient Scribes

This is the area with the clearest, most immediate ROI right now. Ambient documentation tools use natural language processing to listen to a patient encounter and generate a structured clinical note in real time. For clinicians who have been spending 45–90 minutes after hours on documentation, this is meaningful. Less time on notes means more time with patients, and better skilled service documentation that holds up to audit scrutiny.

The downstream effect on staff retention is worth naming directly: documentation fatigue is one of the more underappreciated reasons clinicians leave. Tools that reduce that burden don’t just save time. They improve job satisfaction and reduce turnover, which has its own significant cost.

One important caveat: the clinician still needs to review and edit every note before it enters the chart. AI-generated documentation requires a human check. That’s not a limitation. It’s the appropriate workflow, and it’s non-negotiable from both a clinical and liability standpoint.

Computer Vision and Motion Analysis

AI-powered motion capture tools can analyze gait, functional movement patterns, and sport-specific mechanics, then flag deviations and surface treatment suggestions. The clinical value here is real, but so is the dependency on measurement quality. A tool analyzing movement is only as good as the setup, the camera position, and the standardization of the test. If you’re already practicing with rigorous, repeatable assessment protocols, these tools amplify what you’re doing. If you’re not, the outputs become difficult to contextualize and act on.

RTM Platforms with AI Components

Remote therapeutic monitoring (RTM) has been a legitimate billing category for a while, and the 2026 CMS update expanded the RTM CPT code set, including codes 98979, 98984, and 98985, to reflect more nuanced monitoring activities.¹ AI-enhanced RTM platforms now offer automated adherence tracking, symptom flagging, and engagement prompts between visits. For the right patient population, this is a genuine care extension tool. The remote care infrastructure you already have in place will determine how smoothly these tools integrate.

Generative AI for Clinical Reasoning and Research Access

Large language models (LLMs) and purpose-built clinical AI tools are increasingly being used as sounding boards for clinical reasoning. You describe a patient’s presentation and the tool helps you think through differentials, flag things you might have missed, or surface relevant literature quickly. Done well, this is like having a knowledgeable colleague available at any hour. It works especially well for experienced clinicians who want a second perspective, not a first opinion.

Purpose-built clinical tools that require an NPI to access are particularly strong here. They’re trained on peer-reviewed literature and return referenced responses rather than confident-sounding guesses. The distinction matters.

What’s critical to understand: generative AI is not ready to be used directly in patient education or clinical decision-making without careful expert oversight. Recent studies show LLMs reaching only ~70–74% agreement with expert consensus on knee osteoarthritis rehab program design.² For a patient with a complex presentation, a 26–30% disagreement rate isn’t acceptable. These tools support clinical reasoning. They don’t replace it.

Predictive Analytics in Practice Management

This is an underappreciated category. Practice management platforms with genuine machine learning can flag patients likely to disengage before they drop off, predict scheduling gaps, and model denial risk before claims go out. The ROI here is operational rather than clinical, but for a busy clinic, that matters. The catch: these tools require clean, consistent data to learn from. Fragmented documentation and inconsistent outcome data produce noise, not insight.

The Evidence Reality Check

Here’s where intellectual honesty matters most. AI in rehabilitation is moving fast, but the peer-reviewed evidence hasn’t kept pace with the vendor pitches.

The strongest evidence base right now sits in computer vision and ML applications for movement analysis in physiotherapy, where AI-assisted tools have shown effectiveness in detecting and classifying movement patterns across clinical and home settings.³ Robotic-assisted therapy for neurological rehabilitation also has a growing evidence base, particularly for stroke and spinal cord injury populations.⁴

Where the evidence is thinner: AI-generated patient education, LLM-driven clinical recommendations, and predictive outcome modeling in outpatient rehab. A 2023 systematic review noted that the clinical effects of AI in rehabilitation “have yet to be studied rigorously.”⁵ That’s not a reason to avoid these tools. It’s a reason to pilot them carefully and measure your own results rather than accepting vendor claims at face value.

The honest answer to “does AI improve patient outcomes?” is: in specific applications, yes, with growing support. Across the board, the jury is still out. Any vendor who tells you otherwise is overselling.

Will AI Replace Physical Therapists? Let’s Be Direct

This question is in the room at almost every clinic conversation about AI right now, and it deserves a straight answer rather than a diplomatic dodge.

Not in the way the fear suggests, but the honest version of that answer is more nuanced than the reassuring one. AI will not replace the skilled, relationship-driven, hands-on practice of physical therapy. The core of what an experienced clinician does: manual assessment, clinical reasoning built from years of pattern recognition, therapeutic rapport, motivational attunement. None of that can be automated. These are not just nice-to-haves. They are the mechanisms through which treatment actually works.

What AI will do is change which tasks consume clinical time. Documentation, routine outcome tracking, appointment reminders, basic patient education delivery, scheduling logic: these are already being automated. Clinicians who lean into that shift will find they have more time for the work only they can do. Clinicians who resist it may find themselves at a competitive disadvantage over time, not because AI replaced them, but because practices that adopted it can serve patients more efficiently.

The therapist who uses AI thoughtfully is not more replaceable. They’re more valuable, because they’re spending more time on the work that only a skilled human can do.

The second version of this fear is worth addressing separately: the concern that patients will turn to AI-powered home tools and stop booking appointments. This is a real shift happening at the edges of the market. For patients with straightforward, self-limiting conditions and high health literacy, some will. But for the patients who need what physical therapy actually provides, including skilled assessment, hands-on intervention, progressive loading protocols, and management of complex or chronic presentations, no app is a substitute. The practices most at risk are those already delivering low-differentiation care. The answer to that isn’t avoiding AI. It’s doubling down on the quality and depth of clinical care that can’t be replicated at home.

One more thing worth saying to clinic owners directly: if your response to AI efficiency gains is to immediately increase therapist caseloads without clinician input, you will erode exactly the goodwill those tools created. The time recovered from documentation belongs first to the clinician’s wellbeing and patient presence. That’s how you get sustained adoption and retained staff.

The Foundation Argument: AI Is Only as Good as Your Data

This is the part most AI discussions in PT skip entirely, and it’s the most important thing to understand before you spend a dollar on AI software.

AI tools learn from data and make inferences based on inputs. If your inputs are inconsistent, including manual muscle testing grades that vary by clinician, ROM measurements without standardized protocols, free-text narrative notes, and fragmented outcome tracking, the AI outputs will be unreliable at best and misleading at worst. Garbage in, garbage out, at scale.

Clinics that have already invested in objective, standardized measurement — consistent dynamometry, validated functional assessment tools, and structured outcome tracking — are already positioned to get real value from AI. Their data is clean, structured, and comparable across patients and over time. AI tools can actually learn from it.

Clinics without that foundation will find that AI tools produce outputs they can’t validate, contextualize, or act on. Before asking “which AI tool should I buy,” the better question is: “is our measurement and documentation foundation ready to make AI useful?”

That’s also the argument for why data-driven clinical practice matters beyond its direct clinical value. It’s the infrastructure that makes AI tools clinically useful rather than superficially impressive.

 

Are You AI-Ready? A Self-Audit for Clinic Leaders

Before evaluating any AI vendor, run through this checklist honestly. It’s not about having every box checked. It’s about knowing where your gaps are before a tool promises to fill them.

 

AI Readiness Self-Audit
Measurement & Data Foundation
☐	We use standardized, objective measurement tools (dynamometry, instrumented ROM) rather than relying solely on manual testing
☐	Our assessment protocols are consistent across clinicians. The same test performed by two therapists produces comparable data
☐	We track validated outcome measures at intake, at discharge, and at defined intervals in between
☐	Our outcome data is structured and stored in a way that can be queried, not just filed as narrative notes
☐	We can currently answer: “what is the average functional improvement for our knee replacement patients at 6 weeks?”
Documentation & EMR Readiness
☐	Our clinical documentation is structured enough that key data points (diagnosis, intervention, outcome) are consistently captured in retrievable fields
☐	Our EMR vendor has a clear API or integration pathway for third-party AI tools
☐	We have a defined process for clinicians to review, edit, and sign off on any AI-generated documentation before it enters the chart
☐	Our documentation practices already support reimbursement compliance and would hold up to a payer audit
Clinical Culture & Workflow
☐	Our clinical team practices evidence-based care and is accustomed to using objective data to guide treatment decisions
☐	We have a culture where clinicians are willing to adopt new tools when there is a clear clinical or operational rationale
☐	We have identified a clinical champion who would lead an AI pilot and advocate for it with the team
☐	We have a process for running small pilots: defining success metrics upfront, measuring outcomes, and making go/no-go decisions based on data
Operational & Compliance Readiness
☐	We understand our current documentation time per patient and could measure a change after implementing an AI scribe
☐	We have a BAA process in place for third-party software vendors handling PHI
☐	We have reviewed or are prepared to review FDA SaMD classification for any AI tool making clinical recommendations
☐	We have a policy or are prepared to create one around informed patient consent for AI-assisted care tools

 

If you can check most of the boxes in the first two categories, you have a real foundation to build on. If those boxes are largely unchecked, the priority before any AI investment is shoring up the measurement and documentation infrastructure. That work will pay dividends regardless of which AI tools you eventually adopt.

 

A Framework for Evaluating AI Vendors

When a vendor comes in for a demo, here are the questions that separate signal from noise:

• What is the actual AI/ML component, and what is rule-based automation? Vague answers here are a red flag.
• What peer-reviewed evidence supports the tool? Testimonials and case studies are not clinical validation.
• Is it FDA-cleared as a Software as a Medical Device (SaMD)? If the tool is making clinical recommendations, ask about the regulatory pathway.
• What training data was used, and how representative is it of your patient population? A tool trained on hospital inpatients may not perform well in your outpatient sports rehab context.
• How does it integrate with your existing EMR, and who owns the data? Data portability matters if you ever switch platforms.
• Can the clinician review, edit, and override AI outputs before they enter the chart? This isn’t optional. It’s a clinical and liability requirement.
• What’s the realistic total cost of ownership? Include license fees, implementation, training time, and lost productivity during onboarding.

Walk away if you encounter:

• No clear answer on training data or clinical validation
• No FDA SaMD pathway for tools making clinical recommendations
• Pricing tied to efficiency gains you can’t independently verify
• No ability to demonstrate the workflow in a clinic similar to yours
• Resistance to a structured pilot before full commitment

 

Implementation: Start Small and Measure Both Sides

The clinic leaders who get the most out of AI adoption treat it like a clinical pilot, not a technology rollout. Pick one application. Documentation is the lowest-risk starting point for most practices. Define what success looks like before you start: documentation time per note, clinician satisfaction, note quality on audit review. Run it for 60–90 days. Measure it. Then decide.

Don’t roll out without training. The biggest implementation failures happen when tools are deployed without helping clinicians understand how to use them effectively, how to override AI outputs, and how to maintain their own clinical judgment as the primary filter. Evidence-based practice doesn’t change because a tool is involved. It’s still the clinician’s responsibility to evaluate the output critically.

 

The Human Element Isn’t Optional

There’s a legitimate concern in clinical practice that speed-at-all-costs AI adoption could erode the therapeutic relationship, the part of care that actually drives adherence, engagement, and outcomes. It’s worth taking seriously.

AI does not build rapport. It does not pick up on the nonverbal cues that tell an experienced clinician a patient is struggling with something they haven’t said out loud. It doesn’t adapt its approach based on a patient’s emotional state or motivational readiness. Those things belong to the clinician, and they are not being automated.

The practices that benefit most from AI will be the ones that use it to recover time and cognitive bandwidth so clinicians can be more present, not more efficient in a way that makes care feel transactional. That distinction is everything.

 

Regulatory and Ethical Considerations

A few things every clinic leader needs to have on their radar:

HIPAA compliance is non-negotiable. Any AI tool handling patient data must have a signed BAA, clear data storage policies, and a transparent answer about where your data goes and who can access it.

FDA SaMD oversight applies to software that makes or supports clinical decisions. Tools with diagnostic or therapeutic outputs are subject to FDA review. Not all vendors are forthcoming about this, so ask directly.

Algorithmic bias is a real risk in tools trained on non-representative datasets. A motion analysis tool trained predominantly on younger, able-bodied populations may perform poorly with older adults or patients with significant comorbidities. Ask about training data demographics.

Patient consent and transparency. Patients should know when AI tools are being used in their care. Most will be comfortable with it, but they deserve to be informed.

 

What This Means for Your Clinic Right Now

Here’s the competitive reality: clinics that build a strong measurement and documentation foundation now are positioning themselves to benefit from AI tools as they mature. Clinics that skip that foundation will find AI outputs increasingly difficult to validate or act on. The gap between data-ready practices and data-poor ones will widen, not close, as AI becomes more embedded in clinical workflows.

You don’t need to be the first adopter. But sitting entirely on the sidelines for the next 12–18 months carries its own risk, particularly in documentation efficiency, where early-adopting practices are already recovering meaningful clinical hours per week. That compounds over time.

BTE Technologies has been building the data foundation that makes all of this possible for over 25 years, through precision functional testing equipment, strength testing and dynamometry systems, and evidence-based rehabilitation protocols that generate the kind of objective, structured data that AI tools actually need to perform. That’s not a coincidence. It’s a design philosophy built on the conviction that good outcomes require good data, and that conviction is more relevant now than it’s ever been.

Start with the readiness audit above. Address the gaps that matter most. Then approach AI adoption the same way you approach clinical decision-making: with a clear outcome in mind, a defined measurement plan, and a healthy skepticism toward anything that can’t show its evidence.

The clinics that benefit most from AI won’t be the first to adopt every new tool. They’ll be the ones with the clinical and operational foundation to evaluate, implement, and measure AI honestly, the same way they evaluate everything else.

 

Want to learn more about how BTE’s measurement and rehabilitation systems support data-driven clinical practice?

Explore our full range of functional testing and rehabilitation equipment built to give clinicians the objective data infrastructure that makes every tool in your clinic, AI or otherwise, perform at its best.

 

Morgan Hopkins, DPT, CMTPT is a Physical Therapist and freelance healthcare writer. She spent over eight years treating patients in outpatient orthopedics before transitioning to medical writing. Her clinical specialties include intramuscular dry needling, dance medicine, and sports medicine. Morgan is extremely passionate about holistic wellness, preventative care and functional fitness and uses writing to educate and inspire others.

 

References:

1. Centers for Medicare & Medicaid Services. (2025, November 24). 2026 annual update to the therapy code list (CR 14250). U.S. Department of Health and Human Services. https://www.cms.gov/files/document/mm14250-therapy-code-list-2026-annual-update.pdf
2. Gürses, Ö. A., Özüdoğru, A., Tuncay, F., & Kararti, C. (2025). The role of artificial intelligence large language models in personalized rehabilitation programs for knee osteoarthritis: An observational study. Journal of Medical Systems, 49(1), 73. https://doi.org/10.1007/s10916-025-02207-x
3. Zhou, Y., Nor Rashid, F. A., Mat Daud, M., Hasan, M. K., & Chen, W. (2025). Machine learning-based computer vision for depth camera-based physiotherapy movement assessment: A systematic review. Sensors, 25(5), 1586.https://doi.org/10.3390/s25051586
4. Bin, L., Wang, X., Jiatong, H., Donghua, F., Qiang, W., Yingchao, S., Yiming, M., & Yong, M. (2023). The effect of robot-assisted gait training for patients with spinal cord injury: A systematic review and meta-analysis. Frontiers in Neuroscience, 17, 1252651. https://doi.org/10.3389/fnins.2023.1252651
5. Sumner, J., Lim, H. W., Chong, L. S., Bundele, A., Mukhopadhyay, A., & Kayambu, G. (2023). Artificial intelligence in physical rehabilitation: A systematic review. Artificial Intelligence in Medicine, 146, 102693. https://doi.org/10.1016/j.artmed.2023.102693