Switchboard, MD

Models do the work end to end. Language models read patient messages, calls, voicemails, faxes and forms, then summarise, classify, prioritise, flag clinical urgency, assess sentiment and route to the right person, with conversational agents handling callers and orchestration deciding what happens next. The founding artefact was itself a model rather than a workflow product: a triage model trained to sort messages inside the medical record. Nothing about this offering survives the removal of the models, and the company describes itself as a data science company rather than as software with artificial intelligence added.

Mixed and mostly appropriate. Triage and routing place work in front of a human who decides, which is the right posture for message handling. Conversational agents guide callers without a person present, and autonomous scheduling for high volume visit types acts on its own.

The oversight feature worth naming is that flagging clinical urgency is built in as a first class capability rather than added later, so the escalation concept exists in the product design rather than only in the sales conversation.

Held at B because no escalation criteria, no emergency handling protocol and no disclosure practice about whether callers know they are speaking to a machine were located, and because the same urgency model that raises a genuine emergency is the one that decides a message is routine.

The strongest transparency position in this segment, and it rests on something none of its competitors has. The results of the message triage model at Emory Healthcare were published in the New England Journal of Medicine artificial intelligence journal in 2025, which means external referees examined the method and the performance and any reader can go and check them.

Across two source lists covering more than fifty contact centre and conversation intelligence vendors, this is the only peer reviewed publication located for any of them. Held at B rather than A because the paper covers one component, message triage, at one academic site, while the current platform has grown to include voice handling, scheduling, referral capture and orchestration, none of which has equivalent published evidence, and no model card or architecture description for the wider platform was located.

The architecture offers a real answer for one deployment route and no answer for the other. The platform is deployable on premise or in the cloud, and the company states plainly that the customer's data remains the customer's, so an organisation can choose never to let patient communication leave its own environment.

Offering that choice at all is rare in this index and rarer still at this size, because supporting an on premise path costs a small engineering team a great deal, and it is a structural bound rather than a promise for customers who take it. The limit is that it is an option rather than the architecture. What happens in the cloud configuration is unaddressed, and most customers will take the cloud.

On enumeration there is nothing either way: no model or model family, no foundation model provider, no hosting arrangement for the cloud path and no sub processor list was located, and no retention schedule or position on whether customer conversations improve the models was found. Ask which configuration you are being quoted, and for the cloud path ask for the hosting arrangement, a sub processor list and an explicit training position.

Peer reviewed evidence plus named customers, which is a combination almost nothing in this segment offers. The published Emory work reports a concrete effect, roughly 40 percent of messages that had been reaching clinicians no longer doing so, and it is the effect that matters for a product sold on clinician burden.

Two customers are named on the record with their own executives quoted: Kirby Medical Center, a critical access hospital in Illinois that first ran a focused assessment on prescription refill intake before expanding to an enterprise partnership, and Georgia Vision Institute. The staged pattern of a bounded assessment followed by expansion is itself good evidence, because it means someone measured before committing.

Held at B rather than A because the publication covers one component at one academic centre, no independent evaluation of the voice, scheduling or referral components was located, and no outcome data from either named customer has been published.

An architectural answer rather than a policy one. The platform is deployable on premise or in the cloud, and the company states plainly that the customer's data remains the customer's, which gives an organisation the option of never letting patient communication leave its own environment. Offering that choice at all is rare in this index and rarer still among companies of this size, since supporting on premise deployment costs a small engineering team a great deal.

Held at B rather than A because no retention schedule, minimisation statement or position on whether customer conversations improve the models was located, and because the choice is an option rather than the architecture, so what happens in the cloud configuration is unaddressed.

HIPAA compliance is stated consistently and in formal announcements rather than only on a marketing page, and it appears alongside an independently audited security certification rather than standing alone as an assertion, which is what raises its credibility.

The on premise deployment option also materially changes the posture, since an organisation running the platform inside its own environment has a narrower business associate relationship to manage than one sending communication to a vendor. Held at B rather than A because no agreement terms, privacy page or subprocessor list was located, and the subprocessor question is live for any product using language models, which frequently route through external providers.

Certified to SOC 2 Type 2, stated repeatedly and specifically. The type matters and this company names the strong one: a Type 2 report tests whether controls operated effectively across a period, rather than whether they were suitably designed at a single moment.

Placed on the scale this index has now built across five vendors selling conversation intelligence into healthcare, this sits second: an audited portfolio naming three standards with their versions and the report type earns an A, this single named Type 2 certification earns a B, a published badge with the type unstated earns a B, a Type 1 examination from 2021 with the auditing firm named earns a C, a claim of operating with Type 2 controls but no report earns a C, and an unevidenced assertion of being secure earns a D. Held at B rather than A because no examining firm, examination period, report availability, trust centre or vulnerability disclosure policy was located.

No clearance, authorisation or submission located and none claimed. The boundary is closer here than for most administrative products and deserves stating. A model that reads a patient message and flags clinical urgency is performing triage, which is a clinical judgement about how quickly someone needs to be seen, even though it is delivered inside a routing tool.

The exclusion that keeps clinical decision support outside device regulation turns on a clinician being able to independently review the basis of a recommendation, and a language model's assessment of urgency in free text does not produce a basis that can be reconstructed. Nothing here is irregular and this is the ordinary position of the whole segment. It is recorded because urgency detection is the function on which patient harm would turn, and it is the one operating with the least regulatory scrutiny.

Lifted above the segment norm by peer review and held below a higher grade by what peer review did not cover. Publishing the triage results in a refereed journal subjected at least one model to external examination, which no competitor in this segment has done. But no subgroup performance, bias testing methodology, model card or drift monitoring concept was located for any component. Two exposures are specific.

Voice handling carries the accent, dialect, first language and age variation this index has now recorded eight times before this record. And message triage carries a failure mode drawn from this index's own case finding work: a model that removes 40 percent of messages from clinician inboxes is deciding which patients a clinician never reads, and a misrouted urgent message produces no complaint, no incident report and no record of the miss. The published work is exactly the right vehicle for reporting how often that happens, and reporting it would be worth more than any assurance.

The route into this band is external peer review, and the scarcity of it in this market is worth stating precisely. The results of the message triage model at a named academic health system were published in a peer reviewed medical artificial intelligence journal, so external referees examined the method and the performance and any reader can go and check them.

Across two source lists covering more than fifty contact centre and conversation intelligence vendors, this is the only peer reviewed publication located for any of them. That is what a falsifiable claim looks like: a stated method, a defined population, a result someone else can contest, and a venue that applied its own standard before it appeared. Held below the top grade on scope rather than on quality.

The paper covers one component, message triage, at one site, while the platform has since grown to include voice handling, scheduling, referral capture and orchestration, none of which has equivalent published evidence, and no model card or architecture description for the wider platform was located. A buyer should not read the publication as covering the product they are being sold unless the component they are buying is the one that was studied.

No warranty, indemnity or remediation commitment was located either. Ask which components have published evidence and which do not, and whether the deployed triage model is the one the paper describes.

Real, named and bidirectional. The founding model ran inside the medical record rather than beside it, triaging messages in place, which is a materially deeper position than reading an extract. The company joined the athenahealth partner community as a solution partner in July 2026 with its solution listed in that marketplace, which is a verifiable distribution route rather than a claimed integration.

Referral handling writes as well as reads, classifying an inbound referral arriving by fax, voicemail, email, portal message or web form, then creating or updating the patient record and routing it to a provider. Held at B because one named record system partnership is a narrow base, and no interoperability standard, volume figure or second named platform was located.

One of very few vendors in this index to offer on premise deployment as a real alternative rather than a theoretical one, with the stated position that the customer's data stays theirs. That gives a security conscious organisation an answer to the residency question that requires no contract term at all, which is the same structural advantage recorded against Carenostics, though that vendor runs only in place while this one offers a choice.

Held at B rather than A because no hosting region, architecture description or account of what differs between the two deployment modes was located, and because the choice creates a question of its own: whether the on premise configuration carries the same model capabilities as the cloud one, which for a language model product is not automatic.

No pricing, pricing mechanism or basis of charge located. The funding picture is the more pressing gap and it is genuinely hard to read: the most recent figure locatable is a seed round of roughly 730,000 dollars raised around two years ago, which is very small against the enterprise hospital partnerships the company is now signing, so either the figure is stale or the business is unusually capital efficient.

A critical access hospital or a specialty group placing its patient communication on a supplier this lightly capitalised should establish the current funding position directly, on the supplier continuity ground this index applies from the Behold.ai precedent. The on premise deployment option partly mitigates it, since software running inside the customer's own environment is less immediately lost if a supplier fails.

The range of organisation type is the notable part. Deployments span an academic health system where the founding research was done, a rural critical access hospital in Illinois, and a specialty ophthalmology group, which are three very different operating scales served by one platform, and the smallest of them is the kind of organisation most healthcare artificial intelligence never reaches.

Channel coverage is wide, spanning the portal inbox, live calls, voicemail, fax, text, email, web forms and chat, and fax in particular matters because referrals still arrive that way in much of the country. Functionally it runs from triage and routing through scheduling, reminders, wayfinding and referral capture to booking. Not tied to a clinical specialty, since patient communication is not. Geography is the United States. Held at B because depth in each setting rests on one named customer apiece.