Sword Health

Phoenix performs real work rather than acting as an interface: motion tracking with real time biofeedback during exercise, and around the clock natural language support between sessions. That combination is what allows a supervised program to run without a clinician present. Held back from A because a licensed clinician remains in the loop for every member by design, and the product delivered is a care program rather than software. The company describes this explicitly, which is the right disclosure.

One of the clearest oversight structures in the index, and unusually it is described per product rather than in general. A matched licensed specialist remains in the loop for every interaction, advising on complex questions and sensitive conversations, with the specialist type adapting to the program: Doctors of Physical Therapy for musculoskeletal and lifestyle support. The AI is explicitly positioned as continuous support between clinician touchpoints rather than as a substitute for them, and the company states where human guidance takes over.

The components are named and none of them is characterised.

What is described: motion sensors combined with computer vision and artificial intelligence to guide exercises to completion and deliver real time feedback, with session detail passed to the supervising clinician so the programme can be adjusted. The AI care specialist adds around the clock natural language support. As an account of what the system does for a member, that is clear.

As an account of how well it does it, nothing is published. No accuracy figure for motion tracking or exercise form assessment, no false correction rate, no description of the training data, no validation study of the biofeedback itself as distinct from the programme's clinical outcomes, and no model update or versioning policy.

The distinction between those last two matters here. The company publishes substantial outcome evidence about the care programme, which is graded elsewhere on this record. Programme outcomes tell a buyer whether the overall intervention worked. They do not tell a buyer whether the motion analysis correctly identifies a poorly executed movement, which is the specific function the technology performs and the point at which an error reaches the member as incorrect guidance.

Ask for validation of the motion analysis component specifically, and what happens when the system cannot reliably read a member's movement.

Nothing identifies any party in the chain: no model or model family, no foundation model provider for the natural language care specialist, no hosting arrangement and no sub processor list was located in two passes. The holding is unusual enough to describe precisely.

Motion sensor data and computer vision capture a member performing exercises inside their own home, session by session across a course of treatment, alongside natural language interaction with an artificial intelligence care specialist, clinician notes and outcome measures.

Longitudinal visual and motion data of a person's body moving in their residence is an uncommon category of health information and it is generated continuously rather than at visits, and nothing states whether video is stored at all or only derived positional data, which is the first question because it determines what a breach or a subpoena would expose.

The training question is pointed rather than routine at the reported scale of more than eleven million care sessions: a company operating computer vision motion analysis at that volume has an obvious interest in the corpus for model development and no published statement addresses whether member sessions feed it. One boundary changes the answer entirely and should be established first.

Where the same capability is sold directly to a consumer rather than through an employer or plan, the federal health privacy rule may not attach at all. Ask what is stored from a session, for how long, whether it trains models, and which regime governs your channel.

Scale is substantial and specific: more than 800,000 members, 11 million AI care sessions, over 1,000 clients, and reported per member savings of $3,177 a year for the musculoskeletal program with customer avoided costs exceeding $1 billion. The company also cites peer reviewed research finding high intensity digital physical therapy as effective as clinic based care, which is the right comparison to make.

Held back from A because the economic figures are vendor calculated without published methodology, and the cited peer reviewed work supports the modality rather than this vendor's specific implementation.

A distinctive holding and no published terms governing it.

What is collected goes beyond a clinical record. Motion sensor data and computer vision capture a member performing exercises inside their own home, session by session, over a course of treatment. Alongside it sits natural language interaction with an AI care specialist, clinician notes, and outcome measures. Longitudinal video and motion data of a person's body moving in their residence is an unusual category of health information and it is generated continuously rather than at visits.

Nothing was retrieved describing retention periods for motion or visual data, whether video is stored or only derived skeletal or positional data, whether recordings are used to train or improve models, whether a member can decline that, or what is deleted at the end of a programme.

That last question is the one to press. A company operating computer vision motion analysis at the scale reported, more than eleven million care sessions, has an obvious interest in that corpus for model development, and no published statement addresses whether member sessions feed it.

One boundary worth naming because it changes the answer. Where the same capability is sold directly to a consumer rather than through an employer or plan, the federal health privacy rule may not attach at all, and the protections would then be whatever the consumer terms say.

Ask what is stored from a session, for how long, and whether it trains models.

No published position was located, and the company's status under the rule is genuinely dual in a way a buyer should resolve.

No business associate agreement, addendum, role statement, subcontractor flow down, breach notification timetable or review cadence was retrieved.

The dual status is the substantive point. Sold to self insured employers and health plans, the purchasing plan is a covered entity and this vendor processes member health information on its behalf, which makes it a business associate. But the company also directly employs the licensed clinicians who deliver the care, assigning each member a Doctor of Physical Therapy who assesses, prescribes and adjusts a treatment programme. An organisation providing health care through licensed clinicians and transmitting health information in standard electronic transactions is a covered entity in its own right, not merely a business associate. This index has drawn the same distinction for a clinical laboratory that bills for testing it performs.

Which applies determines real obligations, including whether a notice of privacy practices is owed directly to members rather than handled through the plan. Nothing published addresses it.

The adjacent direct to consumer product launched in 2026 sits outside the scope of this record and outside the rule entirely where a consumer buys it, which is a third status again.

Ask which capacity the company contracts in, and whether members receive a notice of privacy practices.

Two genuine credentials, both correctly named, with a specific audit result stated.

HITRUST CSF certified status is held for the digital therapy platform and several of its applications. The SOC 2 Type 2 examination is reported as completed with zero exceptions, which is more disclosure than the usual bare claim: the type is named, and stating that the examination produced no exceptions describes the outcome rather than merely the existence of a report. Very few vendors in this index publish the result.

That combination is strong. HITRUST is the healthcare specific framework and a Type 2 examination tests whether controls operated effectively over a period rather than whether they were designed at a point in time. Holding both, and naming both precisely, puts this record in the top tier of this index on security terminology alone.

Two qualifications. The original certification announcement dates from 2020, and both credentials run on renewal cycles, so a buyer must confirm current status rather than rely on the announcement; later company statements indicate the standards are maintained across its newer product lines but do not give dates. And no public trust centre was found, so documents come on request.

One caution that belongs on the regulatory axis rather than here: the marketing bundles these two real credentials with an FDA listing under a single claim about being certified. The security credentials are real; the packaging is not.

Ask for current certification dates and the report period.

The regulatory claim is stated accurately in body text and overstated in framing, and the difference matters.

What is true: the company holds an FDA listed class II medical device. Listing means notifying the agency that a device exists and is being marketed. It is a registration step. It is not clearance, and no submission has been reviewed on its merits. This index has graded a vendor down for exactly this conflation before, and applies the same reading here.

What is not true is the packaging. Company announcements have headlined HITRUST, SOC 2 and FDA certifications together, and described the result as making it the most certified provider in its category. There is no such thing as an FDA certification. Bundling a genuine information security certification, a genuine accounting attestation and a regulatory listing under one word invites a buyer to read all three as equivalent achievements when one of them is an administrative filing.

Credit where due: customer facing pages say FDA listed rather than cleared, which is correct.

The regime that actually governs the clinical service is state licensure. Care is delivered by Doctors of Physical Therapy to members wherever they live, which requires licensure in the member's state and raises the usual multi state telehealth questions. Nothing published describes the licensure model, which is the substantive regulatory question for a service of this shape.

Ask what the device listing covers, and how clinician licensure is maintained across states.

No subgroup data, in a technology stack where this index has already recorded the relevant variation twice.

No bias evaluation, governance framework, model update policy or third party review was retrieved.

Two documented failure modes apply directly. Computer vision pose estimation carries measurable performance variation across skin tone and body size, a finding this index has already recorded for a direct competitor in the same category. And conversational systems carry variation across accent, dialect and language, which reaches the natural language support layer.

The population makes both consequential rather than academic. Musculoskeletal programmes are bought by large employers and health plans covering demographically broad workforces. The members most likely to have limited access to conventional physical therapy, and therefore most likely to be steered to a digital alternative, are disproportionately those the literature identifies as least well served by pose estimation and speech recognition. A system that reads movement less reliably for larger bodies is a system that gives worse guidance to a population with higher musculoskeletal burden.

The clinician in the loop is a genuine mitigation and it is graded highly elsewhere on this record. It is not a substitute for knowing whether the automated layer performs evenly, since the clinician sees the system's output rather than the raw session.

Ask for motion analysis performance by body size and skin tone, and for language coverage.

Substantial published outcome evidence exists for the care programme and it does not validate the component that can cause harm, which is the distinction this record turns on and one this index has now recorded in several forms. Programme outcomes tell a buyer whether the overall intervention worked for a population.

They do not tell a buyer whether the motion analysis correctly identifies a poorly executed movement, and that is the specific function the technology performs and the point at which an error reaches the member directly as incorrect guidance. A person recovering from injury being told their form is correct when it is not, or being corrected toward a movement that aggravates their condition, is a harm the programme level outcome measure would absorb rather than surface.

Nothing published gives an accuracy figure for motion tracking or form assessment, a false correction rate, a description of training data, a validation of the biofeedback as distinct from the programme, or a model update policy. The mitigating structure is real and worth stating: session detail passes to a supervising clinician who can adjust the programme, so a human is in the loop across sessions even though not within one. Ask for validation of the motion analysis component specifically, what happens when the system cannot reliably read a member's movement, and what the member is told about the limits of the feedback.

The axis needs rescoping for this business model, and little is published against the rescoped version.

This is not sold to providers, so there is usually no electronic health record on the buyer's side at all. The purchaser is a self insured employer or a health plan, and the systems that matter are eligibility files, benefits administration platforms, claims feeds for outcome measurement, and whatever the plan uses for care management. Grading against clinical record integration would penalise the business model rather than describe it, which this index has ruled against elsewhere.

Against the correct surface, nothing specific was retrieved. No named benefits platform or eligibility integration, no claims data exchange description, no standards support and no interface documentation.

One development points the other way and is worth noting. A separate artificial intelligence division launched in 2025 sells care management agents to payers and providers, described as modular and built to integrate into existing systems and workflows. That is an explicit interoperability claim aimed at exactly the buyers this axis contemplates, and it is asserted without naming a single system, standard or integration.

The clinical continuity question sits underneath all of it. A member's programme, progress and outcomes live with this vendor. Whether any of it reaches the member's own physician or their record is not described anywhere.

Ask how eligibility and outcomes flow, and whether anything reaches the member's treating clinician.

Hardware to the member's home, processing somewhere unstated.

The member facing deployment is well described. A tablet is shipped free of charge, paired with motion sensors, and the mobile application acts as the hub for the programme. There is no clinic visit and no equipment for the employer or plan to host, which is the point of the model.

Everything behind that is unpublished. No cloud provider, no region, no tenancy model, no statement of whether motion and visual data is processed on the device or transmitted for analysis, no subprocessor list, no backup posture and no retention schedule.

The on device question is the one that matters most and it is the same one this index asks of every sensing product. If pose estimation runs locally on the tablet and only derived measurements leave the home, the exposure is narrow. If visual data is transmitted for processing, video of a person exercising in their residence is crossing a network to a vendor environment. Those are materially different and nothing distinguishes them.

Jurisdiction adds a further gap. The company operates in the United States, Europe and Australia, which are three materially different data protection regimes, and no regional handling or residency commitment is published for any of them.

Ask where motion analysis runs, what leaves the tablet, and in which region member data rests.

Structure is disclosed even though amounts are not, and the structure is the interesting part: the company sells as an employer and health plan benefit and states that Pulse carries outcomes based pricing, meaning the vendor takes financial risk against results. Third party coverage indicates risk sharing arrangements more broadly. A buyer can understand how they would be charged and how risk is allocated without a sales call, which is more than most vendors here offer, but not the rate.

Coverage is stated per program and has broadened quickly: musculoskeletal care through Thrive, women's health through Bloom, and cardiometabolic care through Pulse, announced for later 2026. Held back from A because the expansion is recent and Pulse is not yet generally available, so validated coverage is narrower than the marketed range. Buyers should confirm which programs are live for their population.