Popai Health

Remove the model and nothing at all remains. The product is speech and language processing applied to patient phone conversations: transcription and understanding of medical terminology across languages, detection of clinical, operational and social events inside unstructured dialogue, generation of documentation from that dialogue, and real time suggestion to the person on the call.

There is no curated content asset, no rules library and no services layer underneath it, and unlike several vendors in this index the surrounding workflow product is thin rather than dominant. The company also builds rather than borrows, describing the models as healthcare trained rather than adapted, though it publishes nothing to substantiate that characterisation, which is graded on the transparency axis rather than here.

Mostly assistive, with one component that is not and is not discussed. The coaching function suggests talking points to a human coordinator during a live call and that person decides what to say, which is the correct posture. The escalation function triggers automatically on detected events, which is appropriate for a safety net.

The component that deserves scrutiny is documentation: the platform states that it automates documentation into the electronic health record from what it heard on a call. Ambient scribes, which do the same thing for a clinical encounter, universally place a clinician review step before anything enters the chart, and this index has graded them on it.

No published material describes whether a human reviews Popai's generated documentation before it is written, who is accountable for an error in it, or how a coordinator would know the summary had misheard something. That is the first question to put to this vendor.

Two retrieval passes including a fetch of the company's own site located no architecture, no training data description, no validation methodology, no accuracy figures and no publication. The recurring phrases are healthcare trained voice artificial intelligence and advanced artificial intelligence, neither of which specifies anything.

Two performance claims circulate, a better than 20 percent improvement in performance and a 20 percent or greater reduction in documentation burden, both without a method, a baseline or a comparison. The absence matters more than usual for a speech product, because transcription and understanding accuracy on accented, distressed, elderly or multilingual speech is the single property that determines whether everything downstream works, and it is the property most easily measured and most conspicuously unpublished.

Nothing identifies any party in the chain: no model or model family, no hosting arrangement and no sub processor list was located, and no retention, recording or training position was found. A privacy policy is published and compliance badges appear on the site, which is more than several vendors manage, and the detail behind them is absent.

The unaddressed question is consent, and it is sharper for this product than for almost anything else assessed here because of the stated proposition: that every patient conversation is captured and analysed. Total capture is a different proposition from sampling, and patients say things on the phone to a nurse that they would never write on a form, so the material being collected is candid in a way a form response is not.

Recording law also differs by jurisdiction with several requiring all parties to consent, so the question is not only ethical. Nothing published describes what patients are told, whether analysis by a third party model is disclosed separately from the fact of recording, how long audio is retained as distinct from derived text, or whether conversations are used to improve models.

Those are four separate disclosures and a vendor whose entire input is recorded human speech has the strongest reason of anyone to publish all of them. Ask what the call preamble actually says, and for retention on audio specifically.

Real deployment, nothing checkable. Named customers are unusually good for a company founded in 2024: Essen Healthcare, a large provider organisation which also participated in the funding round, and Clover Health, a listed Medicare Advantage plan. A stated volume of more than 18 million minutes of analysed conversation is a meaningful operating scale.

Against that, two passes located no peer reviewed publication, no independent evaluation, no registered study and no customer confirmed outcome. The documentation burden claim is the vendor's own with no denominator. One structural note a buyer should weigh rather than a criticism: the largest named provider customer is also an investor, which is a common and legitimate arrangement in early healthcare software and does mean the reference is not fully arm's length.

A privacy policy is published and the compliance badges are on the site, which is more than most of this index manages, and the detail behind them is absent. The unaddressed question here is consent, and it is sharper for this product than for almost anything else assessed in this index. The stated proposition is that 100 percent of patient conversations are captured and analysed.

Patients say things on the phone to a nurse that they would never write on a form, and recording law differs by state with several requiring all parties to consent. Nothing published describes what patients are told, whether analysis by a third party model is disclosed separately from call recording, how long audio is retained, or whether conversations are used to improve models. A vendor whose entire input is recorded human speech has the strongest reason of anyone to publish its position on all four.

The site displays a HIPAA compliance badge alongside a published privacy policy, and deployments at a provider organisation and a health plan mean Business Associate Agreements certainly exist. Held at C rather than higher because a HIPAA compliance badge is a self assertion: no body certifies HIPAA compliance, so the badge conveys an intention rather than an audited finding, and it should not be read as equivalent to the SOC 2 attestation displayed next to it.

No Business Associate Agreement terms, subprocessor list or breach notification commitment was located. The subprocessor question is worth asking specifically, because speech products frequently route audio or text through external model providers.

A SOC 2 badge is published on the company's own site, which is a third party audited attestation and puts this vendor ahead of the large majority assessed in this index, where the recurring finding has been that nothing is published at all. For a company founded in 2024 to have completed an audit and to display it is a deliberate investment rather than an accident of maturity.

Held at B rather than A because the badge alone leaves the useful detail unstated: whether the report is Type I or Type II, what scope it covers, how a prospect obtains it, and whether penetration testing is performed. There is no trust centre and no vulnerability disclosure policy. Publishing the report type and scope would cost nothing and would move this grade.

No clearance, authorisation or submission located and none needed. Analysing conversations, drafting documentation and prompting a human coordinator is not device functionality and sits outside device regulation comfortably. The point worth making is that the relevant regulator for this product is not the FDA at all.

Its risk adjustment and quality functions place it inside Medicare Advantage payment rules, where oversight belongs to the Centers for Medicare and Medicaid Services and, in enforcement, to the Department of Justice. A buyer evaluating this product against a device framework would be checking the wrong body of law, which is graded and explained on the governance row.

Nothing published, and two exposures are specific enough to name. The first is speech performance across populations. Recognition and understanding accuracy varies with accent, dialect, first language, age and emotional state, and this index has recorded that the affected groups overlap with those already underserved. The population on these calls is precisely that group: care management for high risk patients, social drivers work, transitions of care, multilingual households.

A model that hears some patients less well produces thinner documentation and fewer detected risks for exactly the patients the programme exists to reach, and no subgroup performance is published. The second is the risk adjustment function. The platform offers to use artificial intelligence as a safety net so that conditions mentioned on calls are captured and hierarchical condition category documentation improves.

Surfacing a condition a patient mentions so a clinician can properly assess it is good care. Allowing a condition detected in a phone call with a coordinator to strengthen a risk adjusted payment claim is a different act, because risk adjustment requires diagnosis by a qualified provider in a qualifying encounter, and this is an area of active federal enforcement. Nothing published describes which of those two the product does, or what controls separate them. That distinction should be established in writing before deployment.

Two retrieval passes including a fetch of the company's own site located no architecture, no training data description, no validation methodology, no accuracy figures, no publication and no warranty, indemnity or remediation commitment. The recurring phrases specify nothing, and two performance claims circulate as percentage improvements with no method, baseline or comparison attached. The absence matters more than usual for a speech product and the reason is specific.

Transcription and understanding accuracy on accented, distressed, elderly or multilingual speech is the single property that determines whether everything downstream works, because a system that mishears the input cannot be rescued by anything that follows it, and those four categories are heavily represented among the patients who telephone a healthcare organisation.

It is also the property most easily measured, since a labelled test set of real calls answers it directly, and the most conspicuously unpublished across this whole segment. A vendor that has not measured it does not know how its product performs for the callers most likely to be failed by it; a vendor that has measured it and not published has chosen not to say.

Either way a buyer cannot establish it and should measure it themselves during a pilot on their own call population rather than accepting an aggregate. Ask for word error rate stratified by accent, age and language, the escalation rate when confidence is low, and what the system does when it cannot understand a caller.

The claims are broad and none is named. The platform states that it integrates into any call channel, writes generated documentation into the electronic health record, and drives downstream tasks through its own integration engine. Real deployments at a provider organisation and a health plan imply working integrations exist, since neither could use the product otherwise.

But two passes located no named record system, no interoperability standard, no telephony platform partner and no implementation detail, so the depth is asserted rather than demonstrated. The telephony side matters as much as the record side here and is discussed even less: a product that must sit in the call path has a dependency on whatever contact centre platform the customer already runs.

No published hosting architecture, named region or residency commitment located in two passes. The SOC 2 attestation implies infrastructure controls exist but says nothing about where anything sits. The question is more consequential than usual because the data is audio: recorded patient speech is both large and biometrically identifying, since a voice identifies a speaker independently of what was said, and the retention of raw audio rather than derived text is a materially different exposure. Nothing published distinguishes whether audio is retained, for how long, or where.

No pricing, pricing mechanism, basis of charge or contracting model located, and the basis question is genuinely open for a product like this, since per minute, per seat, per call and per member economics would produce very different bills from the same deployment. The visible commercial facts are an 11 million dollar round in November 2025 led by Team8 and New Enterprise Associates with a customer participating.

That is a reasonable early round and the company is roughly two years old, so a buyer placing call centre operations on it should weigh supplier continuity on the Behold.ai precedent, particularly given that removing a product embedded in the call path is more disruptive than removing an analytics tool.

Broad across the settings where patient telephone contact concentrates, which is a coherent scope rather than a scattered one. Named audiences are care management and case management covering transitions of care, discharge planning, chronic care management, social drivers of health, quality and plan compliance; patient access and call centre operations; specialty care named as cardiology, oncology, endocrinology and neurology; pharmacy led engagement and adherence; and scheduling and referrals including prediction of no show risk.

Buyers span providers and health plans, which is unusual and reflects that both sides run large telephone operations against the same patients. Multiple languages are supported. Held at B rather than A because coverage is United States only and the depth behind each named audience is not evidenced.