Sword

Phoenix, which the company calls its AI Care Specialist, is the engine of the member experience: it gives real time movement guidance during exercise sessions from motion sensors and a computer vision feature that derives movement data from the member's device camera, and it supports and personalizes the program between clinician touchpoints.

The company sells the whole as AI Care, with a licensed clinician overseeing each member's care plan, and its US member privacy notice states that Phoenix makes no autonomous clinical decisions. What is delivered is a care program with clinicians inside it, across several lines (Thrive for musculoskeletal pain, Bloom for pelvic health, Pulse for cardiometabolic care, Mind for behavioral health, plus Predict and the Sword Intelligence agents sold to payers and providers), and the clinician service carries value of its own. The model is the engine of the core module rather than the whole product.

The oversight structure is described and one part is missing. The home page says a human clinical specialist oversees each member's care plan, with the specialist type set by the program (Doctors of Physical Therapy for musculoskeletal care, clinician led care for Bloom, a psychologist led model for Mind), and the US member privacy notice states that Phoenix makes no autonomous clinical decisions: all clinical judgment, treatment planning and escalation decisions are made by licensed clinicians.

What is not published is where the system stops inside a session: the conditions under which Phoenix hands a conversation or an exercise to the clinician, what happens when the camera or sensors cannot read a movement reliably, and how quickly a clinician responds to a member's message. The home page line "Everyone gets the care they deserve" is a mission statement, not a claim that the system does not err.

The approach is described in general terms and nothing is identified. The company describes Phoenix as an AI care specialist giving real time movement guidance, built on motion sensors and a computer vision feature that derives movement tracking vectors and positional data from the device camera feed, together with conversational support between sessions.

No model, model family or foundation model provider is named for either the pose estimation or the conversational layer, no training data is characterized, and no versioning or update practice is published, so a buyer cannot tell when the system changed. The company's outcome evidence describes the care program, not how accurately the motion analysis reads a movement.

The architecture is described and no model provider is named. The US member privacy notice says member information is shared with service providers under contract, including cloud services infrastructure providers, fraud prevention, security and communications providers, and research partners, and it names Google Analytics, Meta and LinkedIn for site analytics and advertising; it links no subprocessor list and names no host.

For the clinical data path it says the computer vision feature derives movement vectors from the camera feed in real time and does not store raw images, and that session video or audio is not recorded without consent. No provider is named for the conversational care specialist, and the trust center (SafeBase) lists frameworks without a subprocessor list. A buyer can see which categories of party receive data but not which companies, and not whether a third party language model sits in the conversation path.

The home page states three headline figures, each with its source: $3,177 in annual member savings (Risk Strategies Consulting, June 2024, an outside actuarial consultancy), a 3:1 average client gross ROI (the company's own trailing twelve month reconciliation of its contractual ROI guarantees) and 55 percent of members surgery free (the company's member base, October 2024).

In August 2026 the company published a clinical and economic evidence report on its musculoskeletal program reporting $3,177 per member per year, a 3.2x return and 54 percent of members with moderate to severe surgery intent remaining surgery free, which is its own analysis. The company also cites peer reviewed research comparing its digital physical therapy with conventional clinic based care.

These figures describe the care program as a whole, clinicians included, rather than the motion analysis alone, and none of the economic figures publishes its method, comparison group and denominator. The company reports more than 800,000 members and 11 million AI care sessions, which is scale of use rather than evidence of benefit.

Sword's US member privacy notice (Sword Health, Inc., updated 10 July 2026) publishes categorical commitments. The computer vision feature processes the device camera feed in real time to derive movement tracking vectors and positional data, and raw images are not stored. Video or audio of sessions is not recorded without explicit prior consent.

Identifiable personal data is not used to train large language models or other general purpose AI systems, while data de-identified to the HIPAA standard (45 CFR 164.514(b)) may be used to improve service quality and product functionality and for research and analytics, with no opt out stated. Retention is stated only in outline: account information until the account is deleted, and health records for no less than the period the law requires, with no maximum.

The schedule is the main gap: no window is given for derived motion data, conversations with the care specialist or deidentified sets. No safety engineering is published either: nothing on guardrails for the care specialist's conversations, how an unsafe answer is detected, or how a safety event is handled.

The HIPAA position is stated through documents rather than a business associate statement. The trust center (SafeBase) lists HIPAA and HITECH, and the home page carries a HIPAA mark. The US member privacy notice (Sword Health, Inc., New York, updated 10 July 2026) reaches the service: it says the company does not directly provide health services and contracts licensed professionals to do so, that information collected in the course of those services may be protected health information, and that such information is handled under a separate HIPAA Notice of Privacy Practices, which the site footer links.

That is a covered entity posture for the clinical service, supported by a substantive privacy document. What is not published is the company's position toward the employers and health plans that buy the program: no business associate status, agreement or subcontractor terms are stated for plan data.

Two recognized audits of how the company operates security are named, HITRUST certification and a SOC 2 Type 2 examination, and listed on a trust center (SafeBase) alongside HIPAA, HITECH, GDPR, PIPEDA, Cyber Essentials and ISO 13485, the last a quality management standard for medical devices rather than a security audit. The trust center gives no dates, versions or report periods, and its documents sit behind a request for access, so the artifacts are not retrievable without asking.

The company's 2020 certification announcement reported the SOC 2 Type 2 examination as completed with no exceptions; current report periods are not published. The site also publishes a responsible disclosure page for security researchers.

The current pages do not state the US regulatory position. The home page shows a CE mark beside HIPAA, HITRUST and SOC 2 marks under the heading "Certified for security, privacy, and clinical excellence", and the instructions for use library lists versioned documents for the Digital Therapist (version 9, DT-IFUPEU-09) and Sword Bloom for the United Kingdom, Portugal and Ireland, which shows a device regime engaged in Europe.

No US instructions for use, registration, listing or clearance number appears. Earlier company announcements described the Digital Therapist as an FDA listed class II device; a listing notifies FDA that a device is marketed and is not a review of it, and grouping a device mark with security attestations under one certified heading invites reading them as equivalent. The clinical service runs under the state licensure of the physical therapists and other clinicians who deliver care, and nothing published describes how licensure is maintained across states.

Sword publishes policy language about how its AI is used and no evaluation of how it behaves across groups. Its US member privacy notice (updated 10 July 2026) states that Phoenix does not make autonomous clinical decisions, that all clinical judgment, treatment planning and escalation decisions are made by licensed clinicians, and that identifiable personal data is not used to train large language models or other general purpose AI systems.

The trust center (SafeBase) lists security and privacy frameworks and no AI governance framework or AI management certification. No bias evaluation, subgroup performance, model update policy or outside review is published. Two variations documented in the wider literature bear on this product: camera based pose estimation can perform differently across skin tone and body size, and conversational systems across accent, dialect and language. Musculoskeletal benefits are bought for broad workforces, so an unmeasured gap would reach many members, and the supervising clinician sees the system's output rather than the raw session.

Sword publishes no warranty, indemnity or remediation commitment attached to what Phoenix tells a member, and no error rate for the motion analysis that produces the feedback. What exists is a review structure: the US member privacy notice says Phoenix makes no autonomous clinical decisions and that licensed clinicians make all clinical judgment, treatment planning and escalation decisions, and the home page says a human clinical specialist oversees the care plan, so session detail reaches a clinician who can change the program.

That is a check across sessions, not within one: a movement wrongly scored as correct during a session reaches the member before anyone reviews it. The home page's financial commitments are about cost, not correctness: outcome based pricing and contractual ROI performance guarantees, reported as a 3:1 average gross return from the company's own trailing twelve month reconciliation. They answer for savings to the employer, not for the accuracy of the guidance, and are described under Commercial Transparency.

The program is bought by employers, health plans and unions rather than by providers, so the systems that matter are eligibility files, claims feeds for outcome measurement and a plan's care management tools, alongside the member's own clinicians and record. Nothing published names a benefits platform, eligibility or claims integration, record system, standard or interface.

The Sword Intelligence line of care management agents, sold to payers and providers, is described by the company as modular and built to fit existing systems and workflows, which is an integration claim with no system named. Whether any part of a member's program or outcomes reaches the member's treating physician is not described.

The member side is described: the company ships a tablet and motion sensors to the member's home for the musculoskeletal program, and the US member privacy notice says the computer vision feature derives movement vectors from the device camera feed in real time without storing raw images. Nothing is published about where member data rests or who hosts it.

The notice refers to cloud services infrastructure providers without naming them, gives no region, and points members in the EU, the United Kingdom and Canada to separate notices, which suggests regional handling without committing to it. Whether the movement vectors are derived on the device or after transmission is not stated.

No price is published, and how the program is sold is discoverable. The home page sells to employers, health plans, unions, and consultants and brokers under Outcome Pricing, described as transparent pricing based on real health outcomes, and cites contractual ROI performance guarantees reconciled over a trailing twelve months, with a 3:1 average client gross return.

That tells a buyer that part of the charge depends on measured results, but not the unit of charge (per eligible employee, per enrolled member or per episode), the base fee, the share of the fee at risk, or the measure and baseline the guarantee is tested against. Purchase runs through a demo or a sales contact.

Coverage is named per program: Thrive for musculoskeletal pain and recovery, Bloom for pelvic health, Pulse for cardiometabolic care, Mind for behavioral health on a psychologist led model, and Predict, sold to employers, health plans and unions in the United States, with separate privacy notices and instructions for use for members in the United Kingdom, the EU and Canada.

Validation stands behind part of it: the musculoskeletal program carries the company's outcome figures and peer reviewed comparisons with clinic based care, while no validation is published for the newer lines. The coverage described here leaves out the direct to consumer channel.