Ellipsis Health

The agent is the product. Sage conducts inbound and outbound patient calls end to end, running engagement and enrollment, clinical assessments and surveys, and care coordination, adapting to what the patient says as the conversation goes. There is no portal, no workflow suite and no services layer that would retain value without it: the company sells conversations that a person would otherwise have to hold, and the model holds them.

The company's own framing is added staffing capacity rather than software, with clinicians reserved for complex high touch intervention, which is a description of the model doing the work rather than assisting someone doing it.

What sits underneath is proprietary and is the company's origin. The published frequently asked questions describe Sage as powered by the Empathy Engine, a combination of patented vocal biomarker technology, therapeutic techniques, and training on millions of real clinical patient calls, and state that the agent detects subtle vocal cues and adapts its tone in real time. So the model is not only generating the conversation, it is reading the acoustic signal while the conversation happens, which is a materially heavier dependency than a scripted voice agent carries.

The product pages lead with the care management use cases rather than with the acoustic technology, which is a marketing choice rather than a change in what runs. Both descriptions come from the company and both are current.

Positioning is consistent and correct in shape, and the escalation rule is not published.

What is stated: the product is added staffing capacity rather than a chatbot, with the explicit purpose of reserving skilled clinicians for complex, high touch intervention while the agent handles high volume outreach. Clinical oversight of all artificial intelligence operations is claimed, a human in the loop approach is stated for complex cases, and the transparency principle commits to telling patients when they are speaking to artificial intelligence. A clinical team validates every product update before deployment, and joint monitoring of calls with the customer is offered during implementation and afterward. Those are real controls on the system's behavior over time.

What is missing is the control during a call. No escalation criterion is published, no handoff path, no statement of when a live person becomes available, and nothing describing what the agent does if a member discloses acute distress during a routine medication adherence or discharge call. The company's own product list includes triage, which is the function where the threshold matters most.

The gap matters more here than for an inbound line a person chose to call. These are outbound calls to health plan members about care gaps and discharge follow up, so the population includes people who are unwell, recently hospitalized, or isolated, and the agent reaches them rather than the other way round. Complex cases are said to involve a human; nothing defines complex.

The training basis is stated and nothing about the system that produces the output is.

What the company publishes is the corpus and the mechanism rather than the architecture. Its frequently asked questions describe the agent as powered by the Empathy Engine, a combination of patented vocal biomarker technology, therapeutic techniques and training on millions of real clinical patient calls, and state that it detects subtle vocal cues and adapts tone in real time. The acoustic work is grounded in more than ten peer reviewed publications on detecting mental health conditions from voice, and a patents page is maintained. Naming the scale and kind of material a model learned from, and saying what signal it reads, is substantive disclosure, and the published research means a technical reader has somewhere real to go.

A buyer reading only the home page and the product page would not find any of that. Those pages describe an agentic care manager with an API first architecture and adaptive conversation, with no mention of the acoustic engine, the corpus or the research. The capability is disclosed; it is disclosed somewhere else.

What is absent throughout is the deployed artifact. No model, model family or foundation model provider is named, no version or update cadence is stated, no model card exists, and no accuracy, sensitivity or specificity figure is published for any inference the system makes. A named partnership with a large artificial intelligence hardware company is announced for more natural conversation, which is the one supplier relationship disclosed and says nothing about what produces a clinical or emotional judgment.

Detecting a mental health condition from voice is a strong claim and needs numbers. So does an assertion that an agent conducts a competent discharge call. Neither has any.

The corpus is described and no party is named.

The company has stated that its engine was trained on millions of live clinical patient calls, with the underlying acoustic work grounded in published research on detecting mental health conditions from voice, so a buyer has been told what kind of material the model learned from and can read the literature. That disclosure makes provenance the question rather than an accusation: millions of live clinical calls is a corpus of recorded patients, and whose calls those were, on what consent basis, and whether a current customer's calls join it are all unanswered.

Voice raises this above the ordinary training question, because a recording carries speaker identity independently of what is said, so a voice corpus is identifying in a way a text corpus is not, and de identifying transcripts does nothing about it.

On enumeration there is nothing. No model, model family or foundation model provider, no hosting arrangement and no sub processor register is published, and no statement covers retention of recordings or of the signals derived from them. A partnership with a large artificial intelligence hardware company is announced for more natural conversation, which is the single supplier relationship disclosed anywhere and does not identify what produces a judgment about a member.

One link in the chain is unavoidable and undescribed. Inbound and outbound calling runs through carriers and voice infrastructure providers, so audio passes through parties the company does not own, and that layer belongs in the answer rather than being treated as plumbing.

Distribution and investor signals are strong, and nothing measures whether the product works.

The commercial evidence is real. A partnership embeds the agent into a major care management suite drawing on member data and returning outcomes to the platform, a second places it inside a large customer relationship platform's health cloud, and a third, announced with a leading artificial intelligence hardware company, concerns more natural conversation. Named customers on the company's site include two of the largest national health plans, a large academic health system and a regional plan. Named testimonials come from two named executives at named customers. Investors include a major pharmacy and health services venture arm.

What is absent is measurement. The operational figures, a 60 percent reduction in administrative tasks, six times faster program enrollment and a fourfold return on investment, are company reported with no site, period, denominator, baseline or method, and the company does not claim otherwise. No clinical validation of any inference capability in deployment is published, and no outcome from any named customer is described.

The deeper gap sits between the science and the product. The company's scientific foundation, more than ten peer reviewed publications, concerns detecting mental health conditions from voice. The product now performs health risk assessments, enrollment, discharge follow up, medication adherence and quality measure work. Evidence that a voice signal carries information about depression does not establish that the agent conducts a good discharge call, and no evidence of the second kind is published.

Disclosure runs at the level of principle and stops short of the questions recorded speech raises.

The company publishes healthcare grade security standards described as covering the health privacy rule and a SOC 2 Type II, encryption in transit and at rest, role based access and data minimization practices, clinical oversight of all artificial intelligence operations, continuous validation and monitoring, and a commitment to tell patients and providers when they are interacting with artificial intelligence. A privacy policy, a terms of use and, separately, a consumer health data privacy policy are all published, the last being the notice that state consumer health data statutes require.

What is absent is everything specific to the material this product generates. No retention schedule for call recordings, no destruction process, no statement of whether audio persists after a call is processed, no position on whether derived acoustic features outlive the recording, and no description of how a signal inferred about a member is stored, who inside the plan can see it, or whether the member can.

The consent question is not hypothetical. The company states that its agent detects subtle vocal cues and adapts in real time, powered by patented vocal biomarker technology, so voice is measured and not merely transcribed. A health plan may lawfully call a member about their care without separate consent. That permission covers the call; it does not by itself cover the collection of a measurement of the person's voice, which several state statutes treat as a biometric identifier with its own consent, retention and destruction requirements. Nothing published describes what consent is obtained before voice is analyzed, or how it is captured on a call the member did not initiate.

The narrower version of the question is unanswered too: whether a member who rang about a prescription refill has a mental health signal inferred from their voice, and if so, where that signal goes.

Compliance with the health privacy rule is stated on the company's own materials, and customers are health plans and at risk providers, so business associate agreements very likely exist as a matter of course. No template, availability statement, scope description, contracting entity or subcontractor flow down position is published.

A separate consumer health data privacy policy is published alongside the general privacy policy and the terms of use, which is the notice that recent state consumer health data statutes require.

A second framework applies here, and it is stricter than the first. The company states that its agent is powered by patented vocal biomarker technology and detects subtle vocal cues in real time, so voice is measured rather than only transcribed. A measurement of vocal characteristics used to characterize a person is a biometric identifier under several state biometric privacy statutes. Those generally require informed written consent before collection, disclosure of a retention schedule and destruction policy, and prohibit sale, and one carries a private right of action with damages assessed per violation, which has produced substantial class litigation against companies that assumed their existing consent covered it.

The critical point is that the two frameworks do different work. The health privacy rule permits use for treatment and operations without specific consent, so a plan may lawfully call a member without one. That permission does not extend to collecting a biometric identifier, where the state statute asks its own question and expects its own consent. A program entirely compliant on the health privacy side can still be exposed on the biometric side, and this exposure is live rather than hypothetical.

A real and independently examined posture, described with care.

The company states a SOC 2 Type II, which it explains correctly as demonstrating that controls were operationally effective over a period rather than present at a point in time; compliance with the health privacy rule; alignment with European data protection law; and regular third party security audits. It also publishes a plain language explanation of what each framework is and does, useful to a reader without a security function of their own.

One status is unclear. Across materials published roughly a year apart, certification under the healthcare control framework is described as in process. That may simply reflect undated pages, and the certification may since have completed. It may also mean the process has been open for some time. Either way, in process is not certified.

What is missing is a trust center or documented request path, a scope statement covering which systems and entities fall inside the attestation, and penetration testing disclosure.

One scope question is specific to this product. The company holds recordings and derived acoustic features from millions of clinical calls. Nothing says whether that corpus and the model training environment sit inside the audited boundary, or only the production service does.

No clearance is published, no regulatory rationale, and the company's own descriptions of the technology span both sides of the boundary. That gap is the finding.

On one side, the operating product as described on the home and product pages is care management: engagement, enrollment, assessments and surveys, discharge follow up, medication adherence and coordination, delivered as staffing capacity for health plans with clinicians reserved for complex intervention. Framed that way it is administrative and no device pathway attaches. Those pages mention no mental health condition at all.

On the other, the capability is still claimed on the same estate. The frequently asked questions describe the agent as powered by patented vocal biomarker technology that detects subtle vocal cues in real time. The company's research library carries mental health screening work, including screening chronically ill patients for depression with a score mapped to a standard depression instrument. The launch announcement describes the agent as supporting patients with complex physical, behavioral and social needs. Software that produces a depression score mapped to a validated instrument is doing what a screening tool does.

So this is not marketing language that predates the product. Two descriptions of the same product sit on the same estate, one administrative and one clinical, and a buyer cannot tell from the outside which governs their deployment. That distinction determines whether what they are buying is a screening instrument requiring clinical governance or an outreach tool requiring operational governance.

One listed function keeps the question open on its own terms. Triage is among the agent's stated capabilities, and triage of a health plan member from a conversation no clinician witnessed is the shape where the decision support question becomes live. No intended use statement, classification, rationale or out of scope statement is published anywhere.

Two named governance bodies with a gate on every release, and no result from any of it.

The published detail is specific and procedural. A cross functional artificial intelligence ethics committee reviews and approves all healthcare applications against stated principles. A cross functional clinical team of physicians, nurses and care managers validates the clinical safety and effectiveness of every product update before deployment, which is a named gate on a named cadence rather than a value. Ongoing reviews of all active projects are described, and a feedback loop invites patients and care managers to report issues which are integrated into model updates. Five principles are set out covering patient centered care, privacy and security, fairness and inclusivity, human oversight and transparency, the last committing to telling patients and providers when they are interacting with artificial intelligence.

That is a governance framework with process behind it, and the route for patients to report issues is a control in its own right.

What is absent is any output. The fairness principle states that the product is designed to serve diverse populations equitably and that biases are identified and mitigated through regular audits and diverse training data. No audit result, no performance breakdown by any group, no calibration data and no external assessment is published, so the equity claim remains an assertion.

The acoustic risk is what makes that absence matter, and it is specific rather than abstract. A system that reads a voice is sensitive to accent, dialect, primary language, age, vocal pathology and the audio channel itself, since telephony compresses and degrades the exact signal such a model reads. The consequence is asymmetric and invisible: a member whose voice is well characterized gets accurate prioritization, and one whose voice is not may be scored as lower need and simply receive less outreach, with nothing in the workflow revealing it.

Nothing stands behind an output that reaches a person who did not ask for it.

No accuracy, sensitivity or specificity figure is published for any inference the system makes, no published limitations, and no warranty, indemnity or remediation commitment. Terms of use and a privacy policy are published; neither carries a performance commitment.

The structural problem is who absorbs an error. The person on the call is a health plan member. They did not buy the product, have no relationship with the vendor, and in most cases did not choose to be called. Whatever the system concludes about them returns to their plan alongside the rest of their record and shapes what outreach they receive next. The member gets no notice from the vendor, no route to see what was inferred, and no way to contest it.

The asymmetry is not theoretical, because the company states the capability runs. Its published frequently asked questions describe the agent as powered by patented vocal biomarker technology and as detecting subtle vocal cues in real time, and its own research library carries mental health screening work including depression screening in chronically ill patients scored against a standard depression instrument. So acoustic inference about state is a live part of the product rather than a legacy claim. A false positive attaches a mental health signal to a member who rang about a prescription refill. A false negative in a product sold partly on surfacing risk means a plan believes it has screened someone it has not. Neither failure has a published rate.

One mitigation is published. The company's transparency principle commits to making patients and providers aware when they are interacting with artificial intelligence, and a feedback route invites patients and care managers to report issues. Notice that you are speaking to software is not notice that your voice is being analyzed, and neither is recourse.

The right integrations for a payer facing product, named specifically and in one case bidirectional.

This is not an electronic health record story. The buyer is a health plan or an at risk provider organization, and the systems that matter are care management platforms rather than clinical record systems. The company names two: a major customer relationship platform's health cloud, where it maintains a listing in that vendor's marketplace, and through a 2026 partnership, a care management suite where it draws on member data and returns outcomes to the platform.

That second arrangement is the more substantive one, because the return path is what makes an outreach tool useful rather than merely busy. An assessment completed by voice is only valuable if its result lands in the system the care manager works in, attached to the right member, in a form the next workflow can act on. Naming a partnership that includes both directions is a real capability claim.

The product page adds an architectural position: an API first design with a stated ability to connect to record and case management systems, and customization to existing workflows.

That connectivity is asserted rather than evidenced. No interface documentation, supported data specification or list of production integrations beyond the two named partnerships is published, and a marketplace listing is a lower bar than a deployed integration. Nothing establishes connectivity to clinical record systems specifically, which matters if a plan wants outreach outcomes visible to treating clinicians rather than only to its own care team.

Delivered as a cloud service, with no hosting location, region, tenancy model, retention schedule or subprocessor list published.

Three questions are specific to this product rather than generic.

Retention of voice is the first and most important. The company describes technology trained on millions of live clinical patient calls, which means recordings and derived acoustic features are held at scale. How long raw audio is kept, whether derived features persist after audio is deleted, and whether a member can have their voice data destroyed are questions with legal weight as well as operational weight, because state biometric statutes impose retention and destruction requirements independently of health privacy law.

The second is the telephony layer. Inbound and outbound calling necessarily involves carriers and voice infrastructure providers, so audio passes through parties the company does not own. That layer belongs in the residency answer rather than being treated as plumbing.

The third is training separation. Where a company improves models from the calls it conducts, the boundary between one health plan's member conversations and the models served to another plan is a commercial question as much as a privacy one, and several of this company's named customers compete directly with one another.

No public pricing. Sold to health plans, at risk providers, and care management organizations, increasingly through platform partners including HealthEdge and Salesforce, which means some buyers will encounter it as a component of a larger contract rather than a standalone purchase. It is not stated whether the biomarker capability is separately priced or bundled.

Clearly bounded by customer and workflow, with the evidence and the marketing pointing at different parts of it.

The customer boundary is explicit: health plans and at risk provider organizations, not consumers and not clinical settings. The workflow boundary is equally clear and enumerated on the product page across three groups. Engagement and enrollment covers program registration, benefits overview, eligibility verification, copay checks and patient queries. Assessments and surveys covers health risk assessments, care and discharge assessments, a social needs screening instrument, and health outcomes, satisfaction and experience surveys. Clinical support covers care coordination, medication adherence, quality and star rating work, pre and post discharge check ins and transitions of care. The company is also explicit about what it is not, positioning the product as added staffing capacity rather than a chatbot, which is a stated ceiling.

Behavioral scope is claimed, and it sits away from the main pages. The product page and home page mention no mental health condition at all, while the frequently asked questions describe the agent as powered by patented vocal biomarker technology detecting subtle vocal cues, the company's research library carries mental health screening work including depression screening in chronically ill patients scored against a standard depression instrument, and the launch announcement describes the agent as supporting patients with complex physical, behavioral and social needs. So the behavioral capability is current, and the home and product pages do not carry it.

The asymmetry that remains is between evidence and application. The peer reviewed foundation concerns detecting depression and anxiety from voice. The product performs health risk assessments, enrollment, discharge follow up, medication adherence and quality measure work. Evidence that a voice signal carries information about depression does not establish that the agent conducts a good discharge call, and nothing says which capabilities rest on the published research and which on general conversational competence.

No language coverage is published, which matters for a product whose delivery mechanism is a spoken call to a member population that is not uniformly English speaking.