Laguna Health

Language models are the whole product. The platform listens to care management conversations, extracts the life circumstances and obstacles that do not appear in a record, generates the documentation, and scores whether the care manager pursued the clues a member offered.

The company is specific about the technology rather than gesturing at it, stating that it uses large language models and that it attends to the bidirectional structure of a conversation rather than treating it as a transcript to summarise. There is no content library, no rules engine and no meaningful software product underneath: strip out the models and what remains is a dashboard with nothing to display.

Assistive on the member side and more consequential on the staff side. Care managers receive preparation, in conversation prompts and drafted documentation, and they decide what to say and what to record. The same question raised against Popai Health applies here: documentation generated from a conversation and written into the member record needs a review step, and none is described. The distinct issue here is the supervisory product.

Laguna Insight analyses 100 percent of member interactions and produces barrier detection scores assessing whether a care manager recognised and explored clues to obstacles, explicitly so that team leaders no longer listen to recordings themselves. That is a model evaluating an employee's professional judgement at scale, which is a different kind of autonomy from advising a clinician, and no published material describes whether a care manager can see, question or contest a score.

Two retrieval passes located no architecture description, no validation methodology, no accuracy figures and no publication. What is published instead is a claim that should not be made: a company representative states that the solution is accurate and hallucination free. No system built on large language models can be guaranteed free of fabrication, no method exists to demonstrate it, and the claim is unfalsifiable as stated.

It is graded here rather than treated as marketing because the whole product writes clinical documentation into a member record, which is precisely the setting where a fabricated detail does damage, and because a buyer who believes the guarantee will not build the review step that would catch one. A published error rate on generated documentation, however unflattering, would be worth more than the assurance.

The published policy discloses more than most in this segment and one permission deserves reading twice. Confirmed in the company's own words: the service collects what a person provides when communicating with a care team by chat, email, video or phone, with call recordings named explicitly; a health insurance carrier may append contact details, medical information and protected health information into the company's systems; consumer information is used for call analysis and summaries; and aggregate, anonymised or de identified information may be shared with third parties for marketing, advertising, business, research or similar purposes.

That last item is the one to read twice, because a company whose product is analysis of recorded care conversations at scale is describing a derived data asset, and what counts as de identified for conversational audio is a far harder question than it is for a claims table: voice is biometric, and the content of a care conversation is distinctive in ways a de identification process built for structured fields does not address.

Real credit for an honest caveat: the deletion route covers protected information and the policy states plainly that aggregate and anonymous derivatives may remain on the company's servers. Two gaps hold it at C. Retention of raw audio as distinct from derived text is unaddressed, and nothing states whether a member is told a call is recorded and analysed by software, which matters because several states require all party consent to record.

There is a real evidence base underneath this product and it is not evidence for this product, which is a distinction the company's own phrasing blurs. Its material states that the solution is built on a proven contextual care model and separately that its suite is proven in published randomised clinical trials driving 50 percent cost savings and tenfold productivity gains, with no citation given and none located in two passes.

Contextual care as a clinical concept does have a published academic basis in work on contextual error, so the model being proven is credible. That a software product applying the concept produces those specific figures is a different claim requiring different evidence. This is the same split recorded against Creyos, where 400 independent studies validate a task battery and a much thinner body of work supports the commercial screener. Held at C rather than lower because the underlying concept is genuinely evidenced; the citation for the product claim is owed.

Revised on 5 August 2026 with what the published privacy policy actually discloses, which is more than the earlier assessment credited and sharper in one place than expected. Confirmed in the company's own words: the service collects the personal information a person provides when communicating with a care team by chat, email, video or phone, and call recordings are named explicitly among what is collected; a health insurance carrier may append contact details, medical information and protected health information into the company's systems; consumer information is used for call analysis and summaries; and aggregate, anonymised or de identified information may be shared with third parties for marketing, advertising, business, research or similar purposes.

That last item is the one to read twice, because a company whose product is analysis of recorded care conversations at scale is describing a derived data asset, and what counts as de identified for conversational audio is a harder question than it is for a claims table. A deletion route exists and covers protected information, with the honest caveat published that aggregate and anonymous derivatives may remain on the company's servers. What holds this at C is unchanged.

Nothing published states whether a member is told a call is being recorded and analysed by software, retention periods for raw audio as distinct from derived text are absent, and there is no statement on whether customer conversations improve models. The recording consent point is not academic: several states require all party consent to record a call, and the population on these calls is disclosing things to a nurse they would not write on a form.

Corrected on 5 August 2026 after the published privacy policy was read in full, which the earlier assessment had not done. It is substantially more detailed than that assessment allowed. The company sets out that it may act as a business associate under the federal health privacy rule, explains what that status means and links to the regulator's own definition, states that it will use and disclose protected information only as the rule permits, confirms that business associate agreements are in place with service providers that access such information, and provides a deletion route that expressly covers it.

It also addresses international transfer of personal information and states that transfers are made only to jurisdictions offering comparable protection. That is a genuine posture rather than a badge. Held at C rather than higher because of two carve outs written into the policy itself, both of which narrow the protection rather than describe it.

The policy states that information collected when a person signs up directly, outside a provider or insurer relationship, is not protected health information, so the same health facts about the same person carry different protection depending on which commercial route brought them in.

And it states that where a person signs a separate authorisation to release health information to the company, that information is no longer subject to the rule and is instead governed by the terms of the authorisation. Both are lawful constructions. Both are worth a buyer reading closely, because a health plan contracting for care management is unlikely to have considered that either route exists.

Corrected on 5 August 2026. An earlier assessment of this record stated that the company published nothing on this axis while asserting that its solution was secure. That was wrong on the first half. The company maintains a trust and security section within its enterprise platform page, linked from both the site footer and the resources menu, and carries a compliance badge row in the footer including the mark of the accountancy body that governs service organisation attestations.

The grade moves up because the earlier position overstated the absence, not because a strong disclosure was found. It stays at C rather than moving higher for three reasons. A badge bearing a standards body's mark is not a statement of which report was obtained, and the difference between a design only examination and one testing whether controls operated across a period is the whole question a buyer is asking.

No report type, no examination period, no auditor name and no scope statement were located. And the trust and security section itself was not read in this pass, so this grade is a correction of a factual error rather than a settled assessment, and it should be revisited on the next pass.

What does not change is the substantive point already on this record: this company competes directly against vendors in the same segment that publish an audited report with the type and the standard version named.

No clearance, authorisation or submission located and none needed. Supporting a care manager's conversation, drafting documentation and surfacing obstacles is not device functionality. As with others in this segment, the regulator that matters is not the FDA: the company names Medicare star ratings and HEDIS quality measures as explicit value drivers, which places the product inside plan quality and payment rules administered by the Centers for Medicare and Medicaid Services.

A buyer assessing this against a device framework would be examining the wrong body of law, and the questions worth asking concern documentation integrity and quality measure attribution rather than clinical safety review.

Nothing published, and two exposures compound each other in a way that is specific to this product and has not been recorded before in this index. The first is familiar: speech and language model performance varies with accent, dialect, first language, age and emotional state, and the members on these calls are disproportionately in the affected groups, since the product targets post discharge, complex care and social obstacle populations. The second is new.

Laguna Insight scores care managers on whether they noticed and explored the clues a member offered. If the model hears some members less reliably, it will detect fewer clues in those conversations, and a care manager assigned to those members will appear to have missed less because less was ever registered, or will be scored against cues the model itself misread.

Model bias therefore propagates into staff performance evaluation, and nothing published describes validation of the scoring, whether a care manager can see or contest a score, or whether scores inform employment decisions. Any vendor scoring employee conversational performance should be asked all three.

Two retrieval passes located no architecture description, no validation methodology, no accuracy figures, no publication and no warranty, indemnity or remediation commitment. What is published instead is a claim that should not be made: a company representative states the solution is accurate and hallucination free. No system built on large language models can be guaranteed free of fabrication, no method exists to demonstrate it, and the claim is unfalsifiable as stated.

It is graded here rather than dismissed as marketing for two reasons. The product writes clinical documentation into a member record, which is precisely the setting where a fabricated detail does damage and persists, and a claim of this kind causes harm through the mitigation it suppresses: a buyer who believes the guarantee will not build the review step that would catch a fabrication, so the assurance removes the control that would have made it survivable.

That is worse than silence, which at least leaves a buyer's caution intact. A published error rate on generated documentation, however unflattering, would be worth more than the assurance and would cost the company only the discomfort of publishing it. Ask for the measured fabrication rate against a reviewed sample, what review sits between generation and the record, and what the vendor commits to when a fabricated detail is written into a member's file.

The platform states that it updates member information automatically and eliminates post call administrative work, which requires writing into whatever system of record the care management team uses, and it tailors documentation format to each organisation's conventions, which implies real integration work per customer. Two passes located no named record system, no care management platform partner, no telephony integration and no interoperability standard. The care management systems this product must sit beside are a fragmented market of their own, and which ones are supported is the first practical question a health plan will ask.

No published hosting architecture, named region or residency commitment located in two passes. As with the comparable record in this segment, the material question is whether raw audio is retained or only derived text, because a voice recording identifies the speaker independently of its content and carries a materially larger exposure than a transcript.

Nothing published distinguishes them, and nothing describes whether any component of the language model pipeline runs through an external provider, which for a product built on large language models is the first thing a security reviewer will ask.

No pricing, pricing mechanism or basis of charge located, and the claimed economics are unusually specific without being verifiable, since 50 percent cost savings and tenfold productivity are strong numbers attached to an uncited study. Total funding is stated at more than 21 million dollars including a 15 million dollar Series A co led by SemperVirens and HC9 Ventures, and no round after 2023 was located, which is a long interval for a company in a segment that has become crowded since. A health plan embedding this in care management operations should establish current funding position directly, on the supplier continuity ground this index applies from the Behold.ai precedent.

Focused on a phase of care rather than a disease, which is a coherent scope. The starting point is the transition out of hospital, where roughly two thirds of discharged patients go home to self care with paper instructions and little oversight, and coverage has extended into complex and ongoing care management.

Buyers span health plans, employers, post acute providers, home health providers and integrated delivery networks, and a separate application extends the product to family caregivers, which is a group almost nothing else in this index addresses directly. Setting is therefore the member's home and the telephone rather than any clinical facility. Held at B because it is United States only and because depth across the named buyer types is asserted rather than evidenced by named customers.