Cadence

AI agents are the engine of a chronic care service that also includes an embedded clinical team. The home page says the agents carry out medication titration, lifestyle coaching, alert triage, care coordination and symptom assessment, automating routine care and protocol driven decisions, and that 99 percent of alerts are resolved without physician escalation.

HypertensionOS, the company's prescription software for Stage 2 hypertension, incorporates AI assisted functionality that carries out routine steps in medication initiation and titration within predefined eligibility criteria and deterministic safety checks, under a licensed clinician with prescriptive authority.

The service sold to health systems is the platform and the care team together, and the clinical staffing carries value of its own, so the model is the engine of the core module rather than the whole product.

The oversight structure is described and its thresholds are not. The home page calls the service physician led with safety guardrails, says care team calls are answered within five minutes, and says 99 percent of alerts are resolved without physician escalation; agents handle titration, coaching, triage, coordination and symptom assessment.

The August 2026 HypertensionOS announcement adds that medication steps run within predefined eligibility criteria and deterministic safety checks, under a licensed clinician with prescriptive authority. What is missing is the rule that decides when an alert or a medication step goes to a person: the criteria, the checks and the escalation thresholds are named but not published, and nothing says what happens after an automated decision proves wrong.

The resolution figure has also moved: earlier company material reported 55 percent of vitals alerts resolved autonomously, and the current home page reports 99 percent resolved without physician escalation, which may count nurse and care team resolution.

The work the agents do is described and the technology behind it is not. The home page lists medication titration, lifestyle coaching, alert triage, care coordination and symptom assessment, and the August 2026 HypertensionOS announcement says the software incorporates AI assisted functionality and runs within predefined eligibility criteria and deterministic safety checks, which suggests a rule layer around a learned component without saying which part is which.

No model, model family, foundation model provider, training data or version is named, nothing states what triggers an escalation or how thresholds are set or personalized, and no update practice is published, so a partner cannot tell when the triage logic changed.

Who else sits between a patient's readings or conversations and the agents' decisions is not established. No model or foundation model provider is named for the conversational agents that reach patients by text and phone, no hosting provider or subprocessor list is published, and the device suppliers and telecommunications carriers in the data path are not named. The August 2026 HypertensionOS announcement names no technology supplier. The privacy policy linked in the site footer may name categories of recipients.

The program's headline cost and utilization claim rests on a peer reviewed, multi site study with a comparison group. "The Impact of a Remote Patient Care Program on Healthcare Costs and Utilization Among Medicare Patients With Chronic Disease" (Mayo Clinic Proceedings, February 2026), led by an author at Massachusetts General Hospital with coauthors from Rush, Providence, Baptist Health and Duke and four Cadence staff, is a retrospective difference in difference analysis of Medicare claims for 5,872 Cadence patients against 11,449 propensity matched controls across 15 states: total cost of care fell by $1,302 per patient per year, inpatient spend by $1,428, and hospitalizations by 27 percent (64 admissions per 1,000 patients a year), with larger reductions reported for heart failure, sepsis and stroke.

Further peer reviewed work covers hypertension (JACC: Advances, July 2025: 23,638 patients, 43 percent increase in blood pressure control, 7/5 mmHg average reduction), heart failure medication use (Journal of Cardiac Failure, September 2024) and program mechanics (NEJM Catalyst, October 2025).

The design is observational, so the effect rests on the matching, and every study evaluates the care program with its clinicians rather than the accuracy of the agents' triage, for which no figure is published.

What is published is a general commitment: the patient terms of use say Cadence must protect protected health information in accordance with HIPAA and describe permitted uses for treatment, payment and health care operations. Nothing answers the questions AI raises.

Three data streams need answers: daily readings from connected devices in patients' homes, which accumulate over years in a chronic care program; text and phone conversations between patients and automated agents, including any call audio; and clinical staff access to partner health system records. No retention window is stated for any of them, no statement says whether patient data trains or tunes the agents, and no de-identification position is published.

The service runs through Cadence Solutions, Inc. and affiliated professional entities, so which entity holds which data is also unstated. No safety engineering is published either: nothing on guardrails for patient facing agent conversations or how a safety event is handled. The privacy policy and Notice of Privacy Practices linked in the site footer may answer part of this.

Cadence states its HIPAA position to patients and supports it with documents that reach the service. The patient terms of use say Cadence must protect protected health information in accordance with HIPAA and describe the permitted uses for treatment, payment and health care operations, and the site footer links a Notice of Privacy Practices and a privacy policy.

That is a covered entity posture for the clinical service delivered through the company and its affiliated professional entities, supported by a substantive document. What is not published is the position toward the health systems that partner with it: Cadence staff work inside partner records, which is ordinarily a business associate relationship, and no business associate status, agreement or subcontractor terms are stated for that side.

Controls are asserted with nothing independent behind them. No SOC 2, HITRUST, ISO/IEC 27001 or other attestation is named and no trust center is linked. The footer's Security disclosures page is a Responsible Disclosure Policy (Cadence Solutions, Inc., 1 October 2025): it invites researchers to report vulnerabilities in the Cadence platform at cadence.care and its subdomains to a security mailbox, promises acknowledgment within seven business days and triage of critical issues within ten, and asks for coordinated disclosure. That is a reporting route, not an independent check: no penetration testing, testing cadence or audit is described. The website privacy policy (1 October 2025) adds only that reasonable security measures are used.

The exposure has three parts that a platform attestation alone would not cover: the connected device fleet in patients' homes (authentication, firmware updates and decommissioning when a patient leaves), clinical staff access to partner health system records (provisioning, logging and revocation), and the platform itself. The company operates through Cadence Solutions, Inc. and affiliated professional entities in Florida, New Jersey, California and New York, so any attestation's scope across those entities also needs stating.

Controls asserted, with a published vulnerability disclosure route and nothing independent behind it.

Cadence states a device position for its medication software and it is a pathway in progress, not an authorization. The company's 13 August 2026 announcement says HypertensionOS is prescription software as a medical device incorporating AI assisted functionality that supports clinician supervised medication initiation and titration for patients with Stage 2 hypertension, that FDA announced it as the second participant in its TEMPO pilot for digital health devices (tied to the CMS ACCESS Model), that it is intended only for use in the pilot at this time, and that participation is not FDA approval or clearance or a determination of safety or effectiveness.

That is a position stated plainly with its limits. What a buyer still has to reconcile is scope: the home page lists medication titration among the agents' current work without saying whether titration outside the pilot runs on HypertensionOS or on clinician executed protocols, and the rest of the service (alert triage, coaching, coordination) states no position of its own. The clinical service runs through affiliated professional entities under state medical practice rules, and remote monitoring billing conditions apply where automation handles alerts.

Controls are described in design terms, and no evaluation of how the agents behave across patient groups is published. The home page calls the service physician led with safety guardrails, and the August 2026 HypertensionOS announcement says the software acts within predefined eligibility criteria and deterministic safety checks.

Neither the criteria nor the checks are published, and nothing states how the triage logic that resolves most alerts without a physician was validated, who reviews it, or how changes are governed. The published studies describe the populations served (the JACC: Advances hypertension paper reports 23,638 patients, 57 percent in rural or underserved areas) without results compared across groups.

Two input problems documented in the wider literature bear on this service: home blood pressure accuracy depends on cuff size and placement, and pulse oximetry can read differently across skin pigmentation, so an alert triaged on those readings inherits their error. Enrollment that depends on a smartphone and connectivity also shapes who is served.

Cadence publishes a review structure for its medication management and nothing that stands behind an output. The August 2026 HypertensionOS announcement describes clinician supervised medication management: the software carries out routine titration steps within predefined eligibility criteria and deterministic safety checks and is intended for use by licensed clinicians with prescriptive authority, so a clinician can challenge a proposed change.

The alert side has no published check: the home page says 99 percent of alerts are resolved without physician escalation, and nothing says what decides that, how often an alert is closed that should have been escalated, or whether a partner can audit closed alerts. A wrongly closed alert leaves no artifact and surfaces later, if at all, as a deterioration nobody traces back.

No warranty, indemnity, remediation commitment or error rate for the triage or titration logic is published, and the company's participation in outcome based payment models is about cost rather than the correctness of an output.

Integration is claimed and nothing names the systems. Cadence clinical staff work inside partner health system workflows and the company describes deep record system integration, and the home page names partner health systems (Montefiore Einstein, Providence, Duke Health, Yale New Haven Health, Rush, Sharp and UC Health among more than 20).

No record system, marketplace listing, standard, interface or integration document is named, and nothing says whether readings, alerts and care notes are written back to the chart automatically, by staff working in the partner's record, or not at all.

The service shape is described and its hosting is not. Connected devices in patients' homes transmit daily readings to Cadence's platform, automated agents reach patients by text and phone, and Cadence clinical staff work inside partner health system workflows across more than 20 health systems in the United States. The privacy policy linked in the footer (Cadence Solutions, Inc. and its affiliated professional entities, updated 1 October 2025) covers the website only; it says information collected there is transferred to and processed in the United States, names Google Analytics and an unnamed AI voice agent vendor for call routing, and refers member data to the healthcare provider's Notice of Privacy Practices and device data to the manufacturers' policies. For the service itself no hosting provider, region, tenancy model, retention schedule or subprocessor list is published, so the location is implied by a United States service for United States health systems rather than committed.

The corporate structure adds a question: patient data is generated in a clinical relationship with affiliated professional entities and processed on infrastructure run by Cadence Solutions, Inc., so which entity holds the clinical record, and what happens to it when a partnership ends, needs settling before patients are enrolled.

No price is published, and how the service is sold is discoverable. Cadence is sold to health systems as a partnership in which the company runs the program, and it describes its economics in terms of outcome based payment: it has announced participation in the CMS ACCESS Model, a Medicare program that pays participants for outcomes in chronic care, and its patient terms make the patient financially responsible for the services, which for a Medicare beneficiary means cost sharing on a monthly service.

What drives the cost is visible, the patients enrolled and served, but no fee, share or rate is published, and a payer program's payment is what Medicare pays rather than what a health system pays Cadence.

The population and conditions are enumerated: older adults at home with hypertension, type 2 diabetes, heart failure and COPD, served through more than 20 partner health systems, with HypertensionOS scoped to Stage 2 hypertension in the TEMPO pilot.

Validation stands behind part of it: peer reviewed studies cover hypertension (JACC: Advances, 23,638 patients, 57 percent rural or underserved), heart failure (Journal of Cardiac Failure) and Medicare utilization across heart failure, sepsis, arrhythmia and stroke (Mayo Clinic Proceedings, 15 states), while no validation is published for diabetes or COPD.