Cadence

Supervised AI agents are the engine of a chronic care management offering that includes an embedded medical group; the AI platform and the care delivery wrapper are packaged together.

Supervised agent model with published autonomous resolution rate (55 percent of vitals alerts), 3.5 minute median response time, and clinical escalation paths.

The capabilities are described in operational terms and the technology is not described at all.

What is public: agents that monitor daily vitals from connected devices, surface risk between visits and coordinate clinical action; workflows that automate the predictable and escalate the complex; an automated assistant that contacts patients by text and telephone; and a stated proportion of vitals alerts resolved autonomously with a median response time. Those are useful operational facts and a buyer can reason about the workflow from them.

What is absent: any model or method, whether models are the company's own or licensed, where inference runs, what triggers an escalation, how thresholds are set or personalised, how versions are managed, and what evaluation supports either the autonomous resolution rate or the escalation logic.

The escalation logic is the part that most needs describing, because it is where the clinical safety of the whole model sits. A system that resolves a majority of alerts without a person is making a judgement each time that no clinician needs to see this, and the basis for that judgement is the product. Whether it is a rule set, a trained model, or a combination, and whether a clinician can inspect or adjust it for their own patients, are questions a partner health system is accountable for and cannot currently answer.

The company also states that its workflows improve with every patient interaction, which implies model change over time and raises the question of how a partner is told.

Ask what decides escalation, how it is validated, and how changes are notified.

Nothing identifies any party in the chain: no model or model family, no foundation model provider, no hosting arrangement and no sub processor list was located in two passes, and no retention or de identification position was found. Three distinct streams flow through that unnamed chain and each raises a different question.

Connected devices transmit daily physiological readings from patients' homes, which accumulates over years for a chronic care population rather than existing as an episodic record. An automated assistant contacts patients by text and telephone, producing conversation content and, for calls, potentially audio. And the company's own clinical staff work inside partner health system workflows, which means access to the record itself rather than to an extract.

Two structural features complicate it further. The service is delivered by a technology company together with several affiliated professional entities, so patient information sits across corporate and clinical entities whose obligations differ and a patient has no way to tell which holds what.

And the statement that workflows improve with every patient interaction is a training claim in ordinary language, with nothing stating whether patient data develops the models, whether a partner can decline, or whether improvements derived from one partner's population serve another's. Ask for retention across device readings, messages and call content, a written training position, and which entity holds what.

Peer reviewed publication record cited on the vendor's own site, spanning NEJM Catalyst, the Journal of the American College of Cardiology: Advances, the Journal of Cardiac Failure, Mayo Clinic Proceedings: Innovations, Quality and Outcomes, and a Circulation abstract, covering hypertension and heart failure programs with clinical outcome and utilization results. Peer reviewed publication in cardiology journals is the top anchor for this axis.

Graded on the existence and venue of the evidence; this index does not re verify the underlying studies, and the vendor's operational figures (55 percent autonomous alert resolution, 3.5 minute median response, $1,302 per patient cost reduction) remain vendor reported.

No stewardship framework, retention position, training use statement or de identification posture was located.

The data surface has three parts and each raises a different question. Connected devices transmit daily physiological readings from patients' homes, which is a continuous stream rather than an episodic record and accumulates over years for a chronic care population. An automated assistant contacts patients by text message and telephone, which produces conversation content and, for calls, potentially audio. And the company's own clinical staff work inside partner health system workflows, which means access to the record rather than to an extract.

Two further features complicate the picture. The service is delivered by a technology company together with several affiliated professional entities, so patient information sits across corporate and clinical entities whose obligations differ, and a patient would have no way to tell which holds what. And the company states that its workflows improve with every patient interaction, which is a training claim in ordinary language: nothing states whether patient data develops the models, whether a partner health system can decline, or whether improvements derived from one partner's population serve another's.

Retention deserves specific attention because chronic care is longitudinal by design. A patient enrolled for years generates a dense physiological history, and what happens to it when they disenrol or the partner contract ends is unaddressed.

Ask for retention across device readings, messages and call content, a written training position, and what data sits with which entity.

The patient terms of use state that Cadence must protect PHI in accordance with HIPAA and describe permitted uses for treatment, payment, and healthcare operations. That is a legal notice to patients rather than a vendor security disclosure to buyers. No trust center, no published BAA terms, and no security certifications were found on the vendor's site; a health system buyer must establish all of this in procurement.

No SOC 2, HITRUST, ISO 27001 or equivalent attestation and no trust centre were located.

The exposure has an unusual shape because this is a service business as much as a software one. Three surfaces need covering and a platform attestation would only reach one. The first is the connected device fleet: monitors placed in patients' homes, provisioned and supported by the company, transmitting daily. Consumer premises equipment is not a controlled environment, and how devices authenticate, how firmware is updated and what happens to a device when a patient disenrols are questions with no software equivalent. The second is the clinical workforce, which works inside partner health system systems, so access provisioning, credentialing and revocation as staff change are partner facing controls the partner should be able to audit. The third is the platform itself.

The segment norm is worth stating. Remote monitoring vendors selling into health systems routinely hold and advertise an independent attestation because partner procurement requires it, so the absence here is conspicuous against the category rather than merely unstated.

One factor probably explains part of it: the company operates through affiliated professional entities as well as a technology company, and attestation scope across such a structure is genuinely awkward to define. That is a reason to ask what is in scope, not a reason to assume it is covered.

Ask what attestation exists and over which entities, how home devices are secured and decommissioned, and how staff access to partner systems is controlled and logged.

Converted from a placeholder row created when a duplicate was repurposed. The prior note set out what the assessment needed and a second pass supplies it.

No device pathway applies to the service and none is claimed. The software monitors vitals, surfaces risk and coordinates clinical action, while clinicians decide what to do, so it sits outside device regulation on the ordinary reasoning. The connected measurement devices in the pathway are a separate matter and the company describes supplying approved devices, which a buyer should read as cleared instruments procured rather than manufactured.

What actually governs is two things and the company's own patient terms make the first of them explicit. The services are delivered by a technology company together with a set of affiliated professional entities incorporated state by state, which is the standard structure for delivering clinical services across jurisdictions. That places the substantive compliance surface in medical practice regulation: licensure in each state, scope of practice for the advanced practice providers who lead the programme, supervision requirements, and the standards governing care delivered without an in person encounter.

The second is reimbursement. Remote monitoring and chronic care management codes carry their own conditions on measurement frequency, the amount of clinician time spent and how it is documented, and those conditions function as the operative compliance requirement for the programme. Where automation resolves a proportion of alerts without a person, whether the time claimed reflects human clinical work is a question that matters more, not less.

Ask which entity treats in each state, and how billable clinical time is captured where an agent acted.

No governance framework, model documentation or performance disclosure was located, and the company publishes one figure that makes the absence consequential.

It reports that a majority of vitals alerts are resolved autonomously, with a short median response time. That is the operating claim the economics rest on, and it is also an allocation statement: for every alert, something decides whether a person is involved. Which alerts are resolved without a clinician, and on what basis, is the substantive governance question, and nothing published addresses it.

The measurement layer adds a second question that is physical rather than statistical. This programme runs on home devices measuring blood pressure, weight, glucose and oxygen saturation in older adults with hypertension, diabetes, heart failure and chronic lung disease. Blood pressure accuracy depends on correct cuff sizing and placement by the patient, and oximetry is known to perform unevenly across skin pigmentation, a finding this index has recorded elsewhere in monitoring. A triage layer built on those inputs inherits their error structure, and a patient whose readings are systematically less accurate is triaged on systematically worse information.

The third question is who enrols. A programme reaching patients who own a smartphone, have reliable connectivity and can manage a device selects for exactly the population least likely to be underserved.

Ask which alert categories resolve autonomously, the false negative rate on escalation, and enrolment and outcome data by patient demographic and connectivity.

One operational figure is published and the number that would make it meaningful is not. The company states a proportion of vitals alerts resolved autonomously and a median response time, which tells a buyer how much work the system absorbs. It does not tell them how much of that work should have been absorbed. A system that resolves a majority of alerts without a person is making a judgement each time that no clinician needs to see this, and the basis for that judgement is the product.

Nothing published describes what decides an escalation, whether it is a rule set, a trained model or a combination, how thresholds are set or personalised, how the logic was validated, or whether a partner clinician can inspect or adjust it for their own patients. That is the point at which the clinical safety of the whole arrangement sits, and a partner health system is accountable for it while being unable to answer basic questions about it.

The failure direction that matters is invisible by construction: an alert wrongly resolved produces no artefact, no complaint and no record that anything should have happened, and it surfaces later as a deterioration nobody traces back. The company also states that its workflows improve with every patient interaction, which implies the escalation logic changes over time, and nothing describes how a partner is told when it does. Ask what decides escalation, how it was validated, the rate of alerts wrongly closed, and how changes are notified.

Deep EMR integration claimed; no marketplace listing or integration documentation verified.

No hosting location, region, tenancy model, retention schedule or subprocessor list was located.

What can be said about the architecture comes from the service description rather than any technical disclosure, and it is more distributed than a cloud platform. Devices sit in patients' homes and transmit over the patient's own connectivity or a cellular path the company provisions. A platform receives and analyses those readings. Clinical staff employed by the company work inside partner health system environments. An automated assistant reaches patients over text and voice, which necessarily involves telecommunications carriers. And the company reports working with more than twenty health systems, so the tenancy question is real.

The carrier layer deserves the same treatment this index has applied to other outbound messaging vendors: text and voice traverse third party networks, so message content passes through parties the company does not control, and that belongs in the residency answer rather than being treated as plumbing.

One question is specific to the corporate structure. Patient data is generated in a clinical relationship with an affiliated professional entity and processed on infrastructure operated by the technology company. Where the clinical record sits, who controls it, and what happens to it if a partnership or an entity relationship ends are questions a health system should settle before enrolling its patients.

Ask where the platform runs, how partner health systems are separated, the retention schedule, the carrier and subprocessor list, and which entity holds the clinical record.

Converted from a placeholder row. The prior note set out what the assessment needed to establish, and a second pass answers most of it.

The company describes its model as fees tied to performance and actual patient utilisation, with the company making the upfront capital investment to stand the programme up, including devices, deployment configuration, enrolment staff, clinical staff, technical support and logistics, and invoicing partners only for services rendered to patients. That answers the structural question the prior note posed: this is neither a platform licence nor a pure revenue share, but a managed service in which the vendor funds implementation and is paid per patient actually served. Naming who carries the upfront capital, and stating that invoicing follows delivery rather than enrolment, is a more substantive disclosure than most of this category offers.

Two things a buyer should carry alongside it. The company has signalled participation in a Medicare innovation model that pays a contingent per patient annual amount tied to outcomes, materially less than the monthly codes conventionally billed for this work. That is a considerable shift in economics, and it is only viable with heavy automation, which connects directly to the autonomy claims on this record. And the company's own patient terms state that the patient is financially responsible for the services, which for a Medicare beneficiary means recurring coinsurance on a monthly service rather than a one off.

Held at B because no rate or range is published, and the structural description sits on a marketplace profile rather than the company's own materials.

Population and conditions explicitly enumerated: older adults, home based, hypertension, type 2 diabetes, heart failure, COPD.