Counsel Health

The model handles the first contact and resolves most of it, inside a business that would still exist without it.

What the model does is real and it is the reason a buyer chooses this over a conventional telehealth service. Medical artificial intelligence gathers context and offers initial evidence based guidance, most concerns are reported as resolved on the platform without a physician joining, and the company describes itself as built from the ground up with artificial intelligence native workflows rather than layering a model onto existing processes. A published quality assurance framework evaluates every interaction automatically, which is itself model work at scale.

What would remain is a telehealth company. The company operates its own medical group of board certified physicians licensed across all fifty states, employs them, and treats patients asynchronously through chat and voice messaging between 8am and 9pm every day. It built its own record system for asynchronous care. It documents every interaction and shares it with health information exchanges and record systems. It sells to payers, employers and health systems with embeddable interfaces, single sign on and rollout kits. Remove every model and that is a fifty state asynchronous physician practice with a purpose built record system and enterprise distribution, which is a product a buyer would pay for and which existed as a category before this one.

What survives without the model matters more than the company's framing. Its own description, physicians as the safety layer rather than the throughput engine, is a statement about intent and about where the leverage sits; it is not a statement about what survives.

The model is the engine of the first contact and of the quality assurance layer, and both are core.

The boundary is drawn at a specific and defensible line, published in both directions, with a threshold, a route to a person and a stated response time.

What the system may do: gather context and offer initial evidence based guidance. What it may not do: diagnose, prescribe or refer. Those three acts require a board certified physician from an in house medical group licensed in all fifty states, and a company physician has stated publicly that they are the decisions only a physician can make. That is a boundary defined by clinical act rather than by disclaimer, which is the form that actually constrains a product.

The routing is published with numbers attached. A physician joins when medically necessary, usually in under fifteen minutes, during stated clinical hours of 8am to 9pm seven days a week. Publishing the hours a person is reachable and the time it takes to reach them answers two questions directly.

A second oversight layer operates after the interaction and is documented in detail. The company publishes a quality assurance framework in which specialized automated judges evaluate every interaction against clinician defined rubrics, scoring adherence to evidence based protocols and flagging safety deviations, with composite labels running from excellent to near miss. Evaluating every thread rather than a sample is what lets a low frequency, high severity failure be found at all, and the company publishes the agreement figures for that system rather than only asserting it.

Continuity supports the design: patients message over days as symptoms evolve, which is a safer failure mode than a single stateless answer.

One question is open. Nothing published describes what the system does when a patient presents something urgent, or what happens outside the stated hours, and no instruction telling the patient that this is not an emergency service is published.

A supplier is named and the quality assurance architecture is documented in detail, while the model that actually talks to the patient is not.

The agent infrastructure has been publicly described as built on a named third party agent framework with encrypted at rest storage shared with physician and patient conversations, so a buyer can read that supplier's own commitments rather than inferring them. The workflow layer is described too: a clinician cockpit integrating patient records, specialty guidelines and academic research, and a purpose built record system for asynchronous care.

The quality assurance side is documented to a level almost nothing here matches. Judge designs are published per condition, including inclusion and exclusion logic, contextual analysis of symptom timing and progression, and a suite of seven independent binary judges aggregated by a clinical rubric into a four level composite label. Performance is published with denominators: 230 threads and 92 threads, accuracy, weighted precision, recall, F1 and Cohen's kappa. That is method a technical reader can argue with.

What is not published is the part that matters most for a product giving medical information. No model, model family or foundation model provider is named for the clinical guidance itself, no version, no update or release cadence, no model card, and no accuracy measurement of the guidance as distinct from the judges that score it. A buyer can learn how the company checks its own work and not what produces the work.

The version question has weight here because the guidance is patient facing and changes silently. Nothing tells a customer which model answered a patient last month, or how they would learn it had changed.

A third party in the chain is publicly identified.

The agent infrastructure has been described publicly as built on a named third party agent framework, with encrypted at rest storage shared with physician and patient conversations, so it is established that a supplier exists, what it does, and that clinical conversation content sits inside it. Naming the framework rather than describing the technology as proprietary lets a reviewer read that supplier's own commitments rather than inferring them.

The onward data flow is documented rather than hidden. The company states it documents every patient interaction and shares it with health information exchanges and record systems so the rest of the care team stays informed. That is a real safety benefit, and it widens the set of parties holding the content.

The model layer and the rest of the chain are unnamed. No foundation model provider, class or version is identified for the clinical guidance, and no subprocessor register covers the wider set of parties, including the exchanges and record systems that receive interaction records and whatever handles voice messages.

Evidence is utilization and self reported avoidance rather than clinical outcome.

The company reports that survey data shows it helped 77 percent of users avoid in person care, and describes serving tens of thousands of patients through health plan and provider partnerships across multiple states. A Medicare focused chronic care program has been announced. Avoided care is a cost metric that cuts two ways: it demonstrates resolution but does not distinguish appropriate avoidance from care that should have happened, and for a front door product that distinction is the whole safety question.

What the company does publish in quantity is quality assurance rather than outcome, and the two should be kept apart. Its report describes evaluating more than 6,000 physician led threads and over 136,000 messages with automated judges, and publishes those judges' agreement with physician reviewers at 96.59 percent accuracy on 230 viral respiratory threads and 81.52 percent on 92 sinusitis threads, with weighted precision, recall, F1 and a Cohen's kappa of about 0.55 on each. Those figures describe how well the oversight system agrees with physicians. They say nothing about whether patients treated through the service did better, or whether the guidance given before a physician joined was correct.

So the company has built and published a measurement apparatus, and pointed it at its own quality process rather than at patient outcome. That counts for something, and it is still not clinical evidence.

No published clinical quality, safety or diagnostic concordance evidence for the care itself exists, and no comparison against an alternative pathway.

The platform holds longitudinal clinical conversations and medical history, which the company states is used only to deliver personalized care and kept private. HIPAA and SOC 2 compliance are stated on the company's own site alongside encryption commitments, and its agent infrastructure has been described as HIPAA compliant and SOC 2 secure with encrypted at rest storage.

The company also states it documents every patient interaction and shares it with health information exchanges and EMRs so the rest of the care team stays informed, which is a real safety benefit and also widens the data flow.

A published notice of privacy practices, which is the right instrument for what this company actually is.

The company operates its own physician group, so it is a covered entity in its own right rather than only a business associate, and it publishes a HIPAA notice of privacy practices accordingly. That is the document a covered entity owes its patients, and publishing it is the core of the posture. Compliance with the health privacy rule and a SOC 2 are both stated on the company's own site, alongside encryption commitments, and a public status page is maintained.

The covered entity posture changes the contracting question for payer, employer and health system buyers rather than removing it. Where the company's clinicians are the treating providers, the clinical record is the medical group's and the purchasing organization is not the covered entity for that care. Where the company processes the purchasing organization's own data, a business associate agreement is the instrument, and no template, execution requirement, scope description or subcontractor flow down position is published.

The embedded distribution model sharpens it. The service is surfaced inside a plan's or employer's properties, so a patient may reach it believing they are using their sponsor's service while the treating relationship and the record sit with the medical group. An employer in particular is sponsoring care it must not see, and the separation should be explicit in the contract rather than assumed from the architecture.

One onward flow belongs in the same conversation: every interaction is documented and shared with health information exchanges and record systems, which widens the set of parties holding the content beyond the two obvious ones.

SOC 2 compliance is stated publicly and corroborated in a third party engineering account describing the platform's agents as SOC 2 secure, alongside stated encryption commitments and a public status page. The SOC 2 type is not specified, and the distinction decides what the claim means: a Type II tests whether controls operated effectively across a period, a Type I assesses design at a point. No trust center, downloadable attestation, audit period, assessor or scope statement is published, so confirming any of it runs through sales rather than self service.

One scope question is specific to this company and worth settling early. It operates a physician group, a purpose built record system and an enterprise distribution layer embedded inside payer and employer properties. Whether an attestation covers all three, or only the production service, is not stated, and the answer matters because the clinical record is the medical group's rather than the purchasing organization's.

The company makes no device claim, scopes the product accordingly, and answers to a regime that is a heavier compliance surface than device regulation.

The device position is simple, and the architecture is what keeps it there. The model provides information; a licensed physician makes every diagnostic and prescribing decision. No pathway is claimed, no classification, and no registration is presented as a credential.

What governs instead is medical practice regulation, and it is worth setting out because it is substantial. The company operates its own medical group with physicians licensed across all fifty states. Medicine is licensed state by state, so a national asynchronous service must match each patient to a clinician licensed where that patient is located at the time of the encounter, and maintain that mapping as clinicians and patients move. States also differ on whether a physician patient relationship can be established asynchronously at all, on what modality a first encounter requires, and on documentation standards.

Prescribing adds a further layer. Authority to prescribe following an asynchronous encounter varies by state and by drug class, and controlled substances sit under a separate federal framework with its own telemedicine conditions that have been repeatedly revised.

That is a real and demanding regime, and answering to it is not a regulatory position on the software. Nothing is published about the product's own regulatory status. A published intended use statement, or a written determination that the guidance component sits outside device regulation and why, is the artifact that is missing and the one a buyer's counsel will ask for.

The corporate structure deserves a question too. An affiliated medical group employing the clinicians is the standard arrangement, and where it sits relative to the technology company determines who carries clinical liability and who holds the record.

A published evaluation framework with a named process, a stated method and figures attached, and no result broken out by anyone.

The process is documented rather than asserted, and that is the main strength. The company publishes a quality assurance framework in which specialized automated judges evaluate every clinical interaction against rubrics its clinical team defined, built from standardized clinical criteria and evidence based protocols. Judge designs are described concretely: inclusion and exclusion logic for respiratory conditions, contextual analysis of symptom duration and clinical progression for sinusitis, and a suite of seven independent binary judges for urinary and vaginal conditions each checking one safety criterion, aggregated into a composite label of excellent, acceptable, inadequate or near miss. The stated analysis covers more than 6,000 physician led threads and over 136,000 messages. Agreement with physician reviewers is published with denominators, and the company reports that the judges apply stricter criteria than physicians on lower quality interactions, which is a finding against its own convenience.

That is governance with named process behind it, and the explicit framing, that scalable deployment of medical artificial intelligence requires defensible governance, is matched by an apparatus rather than a policy page. Clinical structure reinforces it: board certified physicians hold the three consequential acts, and the company positions its product against consumer tools that cannot tell a member what is safe or covered.

What is absent is any breakdown by population. No performance is reported by age, sex, race, language or presentation, no calibration analysis, and no external audit. The gap has a specific edge here: the company notes that some of its most engaged users are women over 65, a group whose symptom presentation and model performance frequently diverge from training distributions, and the published judge work covers four high volume acute conditions rather than the range the service actually receives.

Quality scores broken out by patient group are not published, though the apparatus to produce them already runs on every interaction.

Real mechanisms stand between the guidance and the patient, and nothing stands behind the guidance itself.

The mechanisms are strong for a patient facing product. Diagnosis, prescription and referral are reserved to a board certified physician, so the acts with consequences pass through a licensed human before they reach the patient. Every interaction is documented and shared with health information exchanges and record systems, which is an audit trail a downstream clinician can read. And a published quality assurance framework evaluates every interaction automatically against clinician defined rubrics, scoring adherence and flagging safety deviations, with results aggregated into labels running from excellent to near miss. Review before the act, a durable record, and systematic post hoc detection work together.

What is absent is a commitment about the guidance. No accuracy figure, error rate or limitation is published for the information the model gives a patient before a physician joins, and no warranty, indemnity or remediation commitment is stated.

The published figures do not fill that gap. The metrics in the quality assurance report, accuracy of 96.59 percent on 230 viral respiratory threads and 81.52 percent on 92 sinusitis threads with weighted precision, recall and F1, measure how well the automated judges agree with physician reviewers. They are an evaluation of the oversight apparatus, not of the care. The company also publishes the agreement statistic: Cohen's kappa of about 0.55 on both judges indicates moderate agreement once chance is accounted for, which is a more sober number than the headline accuracy.

One design choice cuts both ways. Sharing every interaction with exchanges and record systems means a downstream clinician sees what was said rather than discovering a gap later, and it means an inaccurate exchange propagates into the permanent record and into systems the vendor does not control, where correcting it is somebody else's process.

The company built its own record system for asynchronous care rather than adopting an existing one, and states it documents every patient interaction and shares that documentation with health information exchanges and record systems so a patient's other clinicians stay informed. Outbound exchange participation is a meaningful interoperability commitment that many virtual care companies skip, and it is the difference between a front door that leaves a gap in the record and one that closes it.

For enterprise buyers the company offers application programming interfaces, single sign on and rollout kits to embed the experience inside an existing platform, which is the integration that matters for a product distributed through payers, employers and health systems.

No record system is named as connected, no interoperability standard is cited for the exchange participation, and no list of the exchanges and systems that receive interaction records is published. That last omission is the sharpest: nothing published lets a patient or a purchasing organization see which systems hold a copy of an asynchronous consultation, and a correction to a record that has already propagated is somebody else's process in each of them.

Delivered as a cloud service with embeddable interfaces and single sign on for enterprise partners, and no hosting, region, tenancy, retention or sub processor terms are published.

The embedded distribution model is what makes this worth more than a standard absence note. The service is sold to payers, employers and health systems and surfaced inside their properties, so a patient may reach it believing they are using their health plan's or employer's service. That produces two questions a purchasing organization should settle before launch.

The first is where the encounter record lives. The company describes a purpose built record system of its own design for asynchronous care, and its clinicians are the treating providers, so the clinical record is the medical group's rather than the purchasing organization's. A health system embedding this is therefore creating records it does not hold, and an employer is in the unusual position of having sponsored care it must not see.

The second is what the enterprise partner receives. Single sign on means the partner knows who used the service. Whether it learns anything about what was discussed is a design decision with real consequences where the partner is an employer, and the separation should be explicit rather than assumed.

Retention is the third and is unaddressed. Asynchronous care produces threads that run over days and include voice messages, and nothing states how long conversations or audio persist, or what happens to them at contract end.

No published pricing for the payer, employer, or health system offerings. The company frames value in terms of avoided high cost claims and per engaged member savings from avoided care and utilization, which suggests a per member arrangement, but no rate or structure is disclosed.

Positioned as enterprise virtual primary care: a front door spanning general medical questions, symptom evaluation, lab result interpretation, medication refills and lifestyle guidance, with escalation to in network care when needed. The physician group is licensed in all fifty states, which removes the state by state coverage gaps common in telehealth, and service runs 8am to 9pm seven days a week. Buyer types are named and addressed separately: health plans, health systems, employers, consumer health partners and individuals, with a Medicare focused chronic care program announced.

The conditions the company describes handling are acute primary care. Its published quality assurance work covers respiratory infections, sinusitis, urinary tract infections and vaginitis, and its consumer content covers dermatology prescribing. Specialty depth is handled by steering to point solutions and in network specialists rather than in house, so this is breadth of access rather than depth of specialty care, and the company is clear about that.

One scope gap stands out precisely because the product is a front door. Nothing published describes behavioral health content of any kind, and a general front door will meet people whose presenting problem is behavioral whether it is offered or not. How such a presentation is handled, and where it is steered, is not described, and for a service that reserves diagnosis and referral to a physician it is a routing question rather than a coverage one.

No language coverage is published, which matters for an asynchronous text and voice service sold to plans and employers with diverse member populations.