Clearstep

The models carry the product without carrying the clinical judgement, which is an unusual and defensible split.

What the models do is essential. Natural language processing interprets a free text symptom description rather than making the patient pick from a list, decides which follow up questions are worth asking, and matches expressed intent to a workflow and a care endpoint. The company's own competitive claim rests entirely on that: roughly ten fewer questions than alternatives, with completion in under five minutes. A static decision tree serving the same protocols would be a materially worse product and a slower one, and patients abandon slow triage flows. A case study describes models trained on intent data and provider expertise to match patients not just to a provider but to the right subspecialty, which is model work rather than lookup.

What the models do not do is decide acuity. That comes from licensed Schmitt telephone triage content, the same protocol set nurse call centres run, and the company positions this as a feature rather than hiding it. So the layer that determines whether a patient is told to call an ambulance or wait until morning is deterministic clinical content, and the artificial intelligence sits on either side of it, interpreting input and routing output.

Graded B rather than A because of that split, and above the handheld device records because here there is no product at all without the models. A buyer should understand the architecture precisely: they are buying a model driven interface to somebody else's clinical protocols, and both halves matter.

The strongest axis on this record, and the reason is architectural rather than promissory.

The product does not diagnose. It determines acuity and recommends a care setting, and the determination runs on licensed Schmitt telephone triage protocols, the content nurse call centres already use for the same decision. That means the autonomous step is bounded by clinical logic a nurse reviewer can inspect, that was authored outside the company, and that predates the product by decades. Very few autonomous patient facing tools can say the decision layer is not theirs.

The framing is consistent with the architecture. A co founder is quoted saying the technology reinforces clinical judgement rather than replacing it, the company describes the conversational flow as explaining to patients why particular care options are being recommended rather than simply issuing an instruction, and the navigation product is positioned as complementing an existing nurse or administrative call centre so that staff handle the complex interactions. A third party product review records the boundary explicitly, that this is not for final diagnosis, emergency reassurance, treatment decisions, or replacing nurse and clinician oversight without local validation.

Graded B rather than A on three unaddressed operational questions. Nothing published describes red flag escalation, meaning what the system does when a patient describes something time critical and then abandons the conversation. Nothing describes a human fallback path or the hours it operates. And nothing describes how under triage is monitored after go live, which is the failure mode that does not announce itself, because a patient told to stay home who deteriorates does not return through the same channel.

Clinical content provenance is disclosed unusually well and model disclosure is thin, which is the mirror image of most records in this index.

On the clinical side the company names its source specifically, stating that its triage is built on Schmitt content and naming Dr Barton Schmitt as the co author of the telephone triage protocols used in nearly all nurse call centres in the United States. That is a checkable provenance claim about the logic that determines acuity, and it lets a nurse leader evaluate the product against something they already know. Most triage vendors describe their content as clinically validated without saying by whom or from what, and this is materially better than that.

On the model side the description stops at function. Natural language processing understands free text and asks fewer questions. Models are trained on intent data and provider expertise. A hybrid artificial intelligence approach combines the two layers. Nothing describes architecture, training corpus, evaluation methodology, update cadence or versioning, and no accuracy figure of any kind is published for either the interpretation layer or the routing layer.

One structural question goes unanswered and matters commercially. Where the licensed protocol content ends and the company's own models begin is not defined anywhere located, so a buyer cannot tell which parts of the product are proprietary, which are licensed, and what would happen to the product if the licence changed.

Graded C: real transparency about the part that decides, none about the part that interprets.

One upstream dependency is named clearly and the rest of the stack is undisclosed, which is a better position than most records in this index reach.

The named dependency is the clinical content. The company states that its triage is built on Schmitt clinical content and identifies Dr Barton Schmitt as co author of the telephone triage protocols used in nearly all nurse call centres in the United States. That is a licensed third party component sitting at the most consequential point in the product, the acuity determination, and naming it lets a buyer evaluate the source independently and recognise it as content their own nursing staff already work with. A third party product profile also refers to clinical protocol terms as something a buyer should ask about, which supports reading this as a licence rather than an inspiration.

Everything else is dark. No language model, vendor, framework or hosted service is named for the natural language layer, nothing states whether interpretation runs on a commercial foundation model or on models trained in house, and no component inventory of any kind is published. For a product whose patient facing conversation is its entire interface, the identity of the model conducting that conversation is a reasonable thing for a buyer to know, not least because it determines where the conversation is processed.

Graded C: a clear disclosure of the component that decides, and no disclosure of the component that talks.

Deployment evidence is strong and performance evidence is entirely self reported, and the gap between the two is wider here than on most records.

The customer list is the substance. Ochsner Health, Mount Sinai, Tufts Medicine, Novant Health, HCA, CVS Health, Duly Health and Care and the Defense Health Agency are named, and several are corroborated outside the company's own marketing. Ochsner's deployment is described in operational detail by its own clinical informatics leadership, including how the triage conversation works and how the organisation intends to extend routing by subspecialty. For a company reported at 29 employees, that roster is remarkable and it is the best evidence on this record that the product works in production at scale.

Every performance figure located is the company's own and none carries a method. A net promoter score stated as four times the healthcare average, a tenfold return on investment, ten fewer questions than competitors, reduced call centre volume and reduced emergency department utilisation all appear without a denominator, a comparison group, a time period or an independent evaluator. No peer reviewed publication was located, and no independent evaluation of any kind.

The missing number is the one that matters most. This product decides whether a patient is directed to an emergency department, an urgent care, a scheduled visit or self care, and its own marketing correctly identifies over triage and under triage as the twin failure modes. No triage accuracy, safety, concordance with nurse decisions or adverse event data was located in any form. A tool that changes where sick people go should be able to show how often it sends them to the right place.

Nothing published addresses what happens to the conversations. No statement was located on whether patient symptom dialogues are retained, for how long, whether they are used to train or tune models, whether any de identification is applied before secondary use, or whether a customer can opt out of any of it.

The content at stake is unusually sensitive even by the standards of this index. A self triage conversation is a patient's unedited account of a symptom, typed in free text before any clinician has framed it, and it routinely contains disclosures a person would not put in a portal message to a named doctor. Mental health presentations, sexual health, substance use and domestic circumstances all surface in triage conversations, and the tool is deliberately positioned at the moment before a patient has decided to see anyone.

The free text design compounds it. Structured symptom pickers constrain what a patient can enter, whereas natural language input invites narrative, and narrative carries identifiers a de identification pass will not necessarily catch.

Nothing here suggests the company handles this badly. Nothing published allows a buyer to determine that it handles it well either, and the customers named on this record will have satisfied themselves privately.

Graded D on the absence. The single most useful disclosure this vendor could make is a plain statement of whether patient triage conversations inform model development, because that is the question a health system privacy office will ask first and the answer is currently only available under a mutual non disclosure agreement.

No published position was located. The site navigation carries a privacy policy and terms of use and nothing else of this kind: no security page, no protected data handling description, no business associate statement, no trust centre and no compliance page.

The absence is conspicuous rather than ordinary because of what the product handles and who buys it. A patient describing symptoms in free text to a triage agent is generating protected health information of a particularly sensitive sort, the conversation is initiated before any clinical encounter exists, and the vendor is unambiguously a business associate the moment that conversation touches a health system's systems. The customer list includes an academic medical centre, a national hospital operator, a retail pharmacy chain and a federal health agency, every one of which will have executed contractual protections and conducted a security review. None of that is visible to the next buyer.

That point deserves emphasis rather than softening. The existence of those customers is strong indirect evidence that the vendor can satisfy a demanding privacy review, because organisations of that kind do not sign without one. Indirect evidence is not disclosure, and this index grades what a buyer can read rather than what a buyer may reasonably infer.

Graded D. Publishing a short protected data handling summary and a statement that business associate agreements are available would move this record several grades at essentially no cost, which is the most useful thing that can be said about it.

No published security posture of any kind was located. The site navigation runs to platform, case studies, blog, about and contact, with a privacy policy and terms of use in the footer and nothing else. There is no security page, no trust centre, no certification claim, no independent audit reference, no penetration testing statement, no vulnerability disclosure policy and no incident notification commitment.

That is a complete absence rather than a thin page, and it is the sharpest contrast available within this part of the index. In the adjacent inbox category one vendor operates a public trust centre and another states certification to a recognised service organisation control standard at the type two level, so buyers in this market plainly do ask and some vendors plainly do answer.

The customer roster makes the silence genuinely puzzling rather than merely disappointing. An academic medical centre, a national hospital operator, a retail pharmacy chain and a federal health agency will each have run a security review before signing, and a company that has passed reviews at that level has almost certainly accumulated the artefacts a trust page is built from. Whatever exists is being shown under agreement to individual prospects rather than published.

Graded D on published evidence, and stated plainly as an absence of disclosure rather than an assertion about the underlying engineering. Of all the gaps on this record this is the one most likely to be a publishing decision rather than a capability gap, and the one most cheaply closed.

No clearance is claimed and, on the current design, none appears to be required. The tool recommends a care setting rather than offering a diagnosis or a treatment, which is the distinction that keeps self triage software outside device regulation, and building on established telephone triage protocols reinforces the position, since the equivalent human process has operated in nurse call centres for decades without device oversight. The company does not overclaim a regulatory credential it does not hold, which is worth crediting because vendors in adjacent categories frequently do.

What is missing is any articulation of where the company believes the line sits and how it stays on its own side of it. Nothing published addresses the regulatory analysis, the intended use statement, or the controls that keep the product from drifting across.

That matters because the company has said publicly where it is heading. Its own guidance to health systems describes layering in decision support where ambiguity or risk is high, and a case study describes evolving toward ever more sophisticated guidance matched by subspecialty and clinical fit. Both directions move toward the boundary rather than away from it. Software that recommends whether to seek care is one thing; software that offers clinical guidance in ambiguous, high risk presentations is a different regulatory question, and the transition between them can happen through incremental product releases rather than a single decision.

Graded C: a correct and unremarkable position today, undocumented, on a product roadmap pointed at the boundary.

Nothing was located. No model card, no training data description, no accuracy figures, no subgroup analysis and no bias statement of any kind.

The exposure here is specific rather than generic. Triage outcome depends on how a patient describes a symptom, and how a patient describes a symptom varies systematically with first language, health literacy, education and cultural framing of illness. A natural language system that interprets free text will handle a fluent, clinically literate description better than a halting one, and the consequence of misreading the second is a patient sent to the wrong place. This is the classic mechanism by which an access tool built to widen access narrows it instead.

The deployment settings sharpen it further. Retail pharmacy and population health programmes reach exactly the populations most likely to be affected, and a third party product review flags language support and accessibility as open questions a buyer must confirm, alongside paediatric and adult differences and the risk of false reassurance.

The protocol foundation provides partial mitigation and only partial. Schmitt content is validated clinical logic, but it was designed to be applied by a trained nurse conducting a conversation, and substituting a model for the nurse changes where the variance lives. The protocols do not become biased; the interpretation feeding them can be.

Graded D. A published breakdown of routing accuracy by language and by presentation type would be the single most valuable disclosure this vendor could make on the clinical side, in the same way a protected data statement would be on the privacy side.

Nothing published addresses responsibility for a triage outcome, and the stakes on this axis are the highest of any record built in this session.

The failure modes are asymmetric and both are real. Over triage sends a patient to an emergency department they did not need, which costs money and consumes capacity. Under triage tells a patient with a time critical presentation to wait, and the harm from that is measured in outcomes rather than dollars. The company's own marketing names both, describing an over and under triage crisis the product exists to solve, so the risk is acknowledged in the sales material and unaddressed in anything a buyer could rely on.

Nothing was located on indemnity, on limitation of liability, on how responsibility divides between the vendor, the health system deploying the tool and the protocol licensor whose content determines the disposition, or on what recourse a customer has if routing accuracy degrades after a model update. That three way division is a genuinely difficult question and it is specific to products built this way: the acuity decision comes from licensed content, the interpretation that feeds it comes from the vendor's models, and the patient facing brand belongs to the health system.

A terms of use document exists on the site and was not retrieved, so a limitation of liability clause may well sit inside it. A general terms of use is not a published position on clinical liability, and a health system's counsel will need to construct the answer from a contract rather than read it.

Graded D.

Deeper than most records in this category, because the product is built to finish an action rather than to hand one off.

The stated integration set covers electronic health and medical record systems with Epic named specifically, customer relationship management systems, and scheduling. The workflow reflects it: triage determines acuity, navigation matches the patient to a service, insurance verification runs, and the patient is booked, with referral handoff also named. That is a chain ending in an appointment rather than in a recommendation the patient must then act on, and it is the reason the deflection claims are plausible, since a patient who books does not call.

The portal embed and post visit text message channels imply working integration with patient engagement infrastructure, and the ability to run alongside an existing nurse call centre implies routing into contact centre systems as well.

Two things hold this at B. No integration mechanism is described anywhere located, so whether this runs on modern interoperability standards, on a proprietary interface or on point to point work per customer is unknown, and that determines implementation effort. And no marketplace listing was located, where two vendors in the adjacent inbox category appear in both the Epic and athenahealth marketplaces, which is both a distribution signal and an independent confirmation that an integration has been reviewed by the record vendor.

The scheduling and insurance verification depth is the genuine differentiator here and it is why this grades above the rest of the record.

Nothing published. No deployment model is described, no hosting arrangement, no region, no residency option, no subprocessor list, no uptime or availability commitment and no disaster recovery position.

The architecture is inferable and inference is not disclosure. A patient facing widget embedded in a health system's website and portal, reachable by text message link and by voice, is a vendor hosted multi tenant service by construction. That is the ordinary and reasonable model for this product. A buyer still cannot learn from anything published where their patients' triage conversations are processed or stored.

One customer makes the omission harder to overlook. The Defense Health Agency is named among the organisations served, and federal health deployments carry hosting, authorisation and residency requirements that a commercial multi tenant service does not meet by default. Either a separate arrangement exists for that customer or the standard service satisfies those requirements, and both possibilities are interesting to a prospective buyer. Neither is described.

Availability deserves its own mention on a product of this kind. Triage is used when a patient feels unwell, which is disproportionately outside business hours, and the deflection value depends on the tool being there at two in the morning. No availability commitment was located.

Graded D on the absence rather than on any evidence of a problem.

Nothing is published. No price, no unit of charge, no tier or module structure, no implementation fee guidance, no contract length and no minimum. The only commercial description located comes from a third party product profile, which characterises the arrangement as enterprise implementation pricing and advises a prospective buyer to request current terms covering modules, channels, integrations and clinical protocol licensing. That last item is worth noting on its own, because it implies the licensed triage content may carry its own commercial terms, and nothing published confirms or denies it.

The website carries no pricing page and no request a quote flow beyond a general contact form. Case studies describe outcomes without cost. Return on investment is asserted as a multiple with no cost side to the calculation, which is a claim a buyer cannot check or model.

Graded D as the floor, and the floor is correct here rather than harsh. Several vendors in the adjacent inbox category publish real structure, including a full per provider per month ladder in one case, so this is not a category where disclosure is impossible or unusual. It is a choice.

One consequence for buyers is worth stating. Where a vendor sells modules across suites, as here with a triage suite, a navigation product and a capacity optimisation suite, the absence of any published structure means a buyer cannot tell which capabilities described in marketing material are included in the product they are being quoted for, and that is a question to settle in writing before signature rather than after.

Organisation types and channels are broad, and clinical breadth is inherited rather than built.

The company names health systems and hospitals, population health management organisations, retail pharmacy chains, multi specialty medical groups and digital health companies as buyer types, and the customer roster bears that out across an academic system, a large for profit hospital operator, a retail pharmacy chain and a federal health agency. Reaching those four buyer categories with one product is genuinely unusual.

Channels are equally wide. The experience runs as a website widget, an embed inside an existing patient portal, a link in post visit text messages, a voice interaction, and as an augmentation to an existing nurse or administrative call centre rather than a replacement for it, with the company stating that the navigation product can be bought without the triage product and vice versa.

Clinical coverage comes from the underlying Schmitt protocol library, which spans adult and paediatric telephone triage across a very wide symptom set and is the standard content in nurse call centres. That is real breadth and the company did not create it.

Two gaps hold this at B. Nothing published describes language coverage, which is a first order question for a patient facing triage tool serving retail pharmacy and safety net populations, and a third party product profile flags language support and accessibility as open items a buyer must confirm. And paediatric versus adult handling, another item that profile raises, is not addressed in anything located despite the protocol library covering both.