RadiantGraph

There is no product here without the models, and that is true of both layers rather than one.

The personalisation layer is the original product and it is entirely model driven. The platform ingests member data from a customer's own infrastructure and builds per person models that decide who to contact, about what, and through which channel. The company's own scale metric is stated in those terms, processing personalisation models for more than 3.5 million people, which is a description of model throughput rather than of records held. Remove the models and what remains is a data pipeline with nowhere to send its output.

The voice layer is equally dependent. AI Voice Studio runs conversational agents capable of naturalistic dialogue, not recorded messages or dialler assisted human calling, and the published study measures the agents against human callers rather than against a manual process. A campaign tool without the agents would be an ordinary outbound calling platform of the kind this product exists to replace.

The data infrastructure integrations across four major cloud and warehouse platforms are the one substantial non model component, and they exist to feed the models rather than to deliver value on their own. That is the correct architecture for a payer side product, where the customer already holds the member data and the vendor's contribution is what it decides to do with it.

Graded A, the highest on this axis in this session, and the record is unambiguous: the company describes itself as an artificial intelligence engine and the description survives scrutiny.

The design is bounded and the boundary that matters most is undefined.

What is described is the right shape for outbound engagement. The published study characterises the agents as capable of naturalistic dialogue while maintaining strict adherence to conversation flows, which is bounded autonomy: the model handles the conversation and a defined script governs where it can go. For outreach the risk profile is genuinely lower than for triage, because the agent is driving enrolment and scheduling rather than dispositioning acuity, and a poor outcome is usually a member who does not book rather than a member who is harmed. The study reporting no safety incidents or escalations across nearly 8,800 members is a real observation and the kind vendors normally omit.

That finding is also ambiguous in a way worth stating. Zero escalations can mean the escalation path worked and was never needed, or that it was never exercised, or that nothing triggered it because no trigger was defined for the situations that arose. Nothing published distinguishes these, and a study reporting zero of something is weaker evidence than a study reporting a rate.

The unaddressed boundary is specific and it follows from the customer base. The platform runs outreach for substance use, mental health and chronic condition programmes, so agents will place calls to members who may disclose crisis, suicidal ideation or acute symptoms mid conversation. That is a foreseeable event at this volume, not an edge case. Nothing published describes crisis detection, escalation triggers, warm transfer to a clinician, or out of hours handling, and nothing states whether members are told they are speaking to an artificial agent or can request a person.

Graded C.

Architecture is described at the level of named components and nothing below that.

What is disclosed is more than the marketing minimum. The platform is separated into a personalisation layer branded as Intelligent Personalization and a voice layer branded AI Voice Studio, with the relationship between them stated: the voice agents build on the member data and personalisation capability rather than operating independently. Scale is given in model terms, personalisation models running for more than 3.5 million people. The data infrastructure is named specifically across Amazon Web Services, Google Cloud, Snowflake and Databricks, which tells a technical buyer exactly what the ingestion surface looks like. The published study adds real methodological detail on deployment design, population, duration and journey segmentation.

What is absent is everything about the models themselves. No architecture, no training data description, no accuracy or performance figures for the personalisation models, no evaluation method beyond the outcome study, no versioning and no update cadence. For the voice layer, no speech recognition, synthesis or language model provider is named, and no description of how strict adherence to conversation flows is enforced technically, which is the control the whole safety case rests on.

That last omission is the substantive one. The study's safety claim depends on the agents staying inside defined flows, and whether that is achieved through prompt constraints, a state machine, output filtering or something else determines how robust it is. A reader is asked to accept the outcome without the mechanism.

Graded C.

No upstream model dependency is named anywhere, and for a voice product that is the disclosure a buyer most needs.

An outbound conversational agent requires at minimum speech recognition, language understanding and generation, speech synthesis and telephony. Several of those are almost never built in house by a company of roughly 21 employees, and the platform's own launch material describes the voice capability as arriving in June 2025 as an addition to an existing personalisation platform, which is a timeline consistent with integration rather than ground up development. Nothing published names a provider for any layer, and nothing states which components are the company's own.

The question is not academic. Each unnamed model provider is a party in the data path handling member voice audio, each carries its own retention and training terms, and a health plan conducting diligence needs to enumerate them to satisfy its own obligations. It also bears on the certification scope question raised on the security axis, since a certified platform integrating an uncertified model service has a boundary a buyer should locate.

The personalisation layer is equally undisclosed upstream, with no statement on whether models are built in house, derived from open frameworks or licensed.

What is named is infrastructure rather than models. Four cloud and warehouse platforms are identified as integration points, which describes where customer data is read from and says nothing about what processes it afterwards.

Graded D as an absence of disclosure rather than evidence of a problem, and it is the sharpest contrast with the certification posture recorded elsewhere on this record.

The strongest evidence produced by any patient engagement vendor in this session, and the qualifications matter.

The study is real and it went through peer review. Conducted with Oshi Health, a virtual first gastrointestinal specialty clinic, it ran seven weeks across nearly 8,800 members and three distinct journey stages in the fourth quarter of 2025, and was presented at the 2026 International Conference on Artificial Intelligence in Medicine and published in Lecture Notes in Artificial Intelligence with a citable identifier. Reported outcomes are specific and operationally meaningful: appointment completion up 3.4 times overall, 24.7 times among members who had registered without scheduling, a 9 times increase in outbound call volume against human agents, and roughly 75 hours of staff time saved per 1,000 contacts. The company also states results have held as the population grew. Taking a deployment to a peer reviewed venue is rare enough in this index to be worth crediting on its own.

Four things hold this at B. The work is a collaboration between the vendor and its own customer, so it is company associated rather than independent, and both parties benefit from a positive result. The headline figures are multiples with no absolute baselines stated, so a 24.7 times improvement on a very low starting rate is a different result from the same multiple on a moderate one, and a reader cannot tell which this is. It is a five page conference paper rather than a full trial, and no control arm is described, so the comparison appears to be before and after rather than concurrent. And the other performance claims in circulation, a 70 percent team efficiency gain and a brand impact ten times that of human agents, carry no method and should not be read as evidence.

One finding deserves separate weight because vendors rarely report it: no safety incidents or escalations were identified during the study.

Certification covers the handling and nothing covers the learning, which is the split this axis exists to separate.

On handling, the position is better than most. HITRUST certification is a validated external assessment of controls over protected data, and it addresses access, encryption, logging and retention practice in a way a self asserted compliance claim does not. A plan's privacy office has something real to work with before any conversation begins.

On stewardship nothing was located. No statement addresses whether member data, call recordings or transcripts inform model development, whether personalisation models are trained per customer or across the book, what retention applies to voice audio, or whether a customer can decline secondary use. The cross customer question is the pointed one, because a personalisation engine improves with population scale and a vendor processing models for millions of people across multiple plans has an obvious technical incentive to learn across them. Nothing states whether it does.

The voice layer raises the harder version of the same question. Outbound calls produce audio of members discussing their health, and audio cannot be de identified the way a structured field can. The named programmes include substance use and mental health, where the recording captures the most sensitive category of health disclosure and where federal confidentiality rules for substance use records impose obligations beyond the general privacy statute.

One detail cuts in the vendor's favour. Integration with the customer's own cloud and warehouse infrastructure suggests member data may largely stay in the customer's environment, which would be a genuinely good architectural answer. Nothing published confirms it.

Graded C: strong certified handling, no published stewardship position.

The strongest certification backed privacy posture of any non device vendor in this session, with the contractual layer still unpublished.

HITRUST certification is the material fact. The framework consolidates the health privacy statute's requirements alongside other recognised standards into a single certifiable control set, it is earned through a validated assessment rather than asserted, and it is expensive enough that a company of roughly 21 employees holding it represents a deliberate allocation of scarce resources. For a vendor processing member data on behalf of health plans, it is the credential a plan's privacy office is most likely to ask for by name. Service organisation control type 2 and health privacy statute compliance are stated alongside it.

The data flows make the vendor unambiguously a business associate. The platform ingests member data from the customer's own warehouse infrastructure, builds person level models on it, and places outbound calls in which members discuss their care, including in substance use and mental health programmes where a disclosure carries additional statutory protection beyond ordinary health information.

What is missing is the contract. No business associate agreement is offered or described, no terms are published, no protected data handling summary exists, and nothing states retention for call recordings and transcripts. The substance use point deserves particular attention, because federal confidentiality rules for substance use disorder records impose obligations beyond the general privacy statute, and nothing published indicates whether the platform is configured for them.

Graded B on the certification, held from A by the absence of any published contractual or retention position.

The strongest certification stack held by any software vendor built in this session, without the surface that would let a buyer verify it.

Three credentials are claimed together: HITRUST certification, service organisation control type 2 and health privacy statute compliance. The first carries the weight. HITRUST consolidates the health privacy statute alongside other recognised standards into a single certifiable control set, it is earned through a validated assessment with maturity scoring rather than asserted, and it is registered with the certifying body so its existence is externally checkable. Health plans routinely ask for it by name and many will not contract without it. Holding it at roughly 21 employees is a serious commitment of resources by a company of that size, and it signals that enterprise payer procurement was designed for rather than encountered.

The claims are also made consistently in company issued material over time rather than appearing once in a marketing page.

What is missing is the apparatus around them. No trust centre exists, no process is described for requesting reports under agreement, no audit period, auditor or certification scope is named, and no penetration testing statement, vulnerability disclosure policy, subprocessor list or incident notification commitment was located. Certification scope matters especially here: a certification covering the personalisation platform may or may not extend to the voice infrastructure launched later, and nothing published resolves that.

Graded B rather than A because the credentials are strong and independently meaningful while the means to examine them are entirely absent. Publishing a trust page with scope and a report request process would make this an A at negligible cost.

No clearance is claimed and none is required. Member outreach, enrolment and engagement are administrative rather than clinical functions, the agents drive appointment completion rather than offering assessment or advice, and nothing in the product as described approaches a device claim. This is correctly positioned and does not overclaim.

The regulatory exposure sits elsewhere and it is real for this business model rather than theoretical. Automated outbound calling is governed by federal telephone consumer protection rules, which impose consent requirements, calling window restrictions and revocation handling, and the penalties are assessed per call. A platform whose value proposition is a nine fold increase in outbound call volume is scaling exactly the activity those rules govern, and nothing published addresses consent capture, how a customer's existing consent record is honoured, do not call handling or opt out processing.

A second and newer exposure concerns disclosure. Several states now require that a person be told when they are interacting with an artificial agent rather than a human, and more are legislating. For an outbound voice product calling members across a national plan footprint this is a per jurisdiction compliance question, and nothing located describes whether disclosure is made, how it is configured, or how state variation is handled.

A third applies to the substance use programmes named among the platform's use cases, where federal confidentiality rules restrict even the fact of a person's participation in treatment, which constrains what an outbound caller may say if someone other than the member answers.

Graded C: a clean position on the regulation that does not apply, and silence on three that do.

One genuine governance artefact exists and the subgroup question is untouched.

The artefact is the peer reviewed study. Submitting a real world deployment to an academic conference, describing the method, reporting a safety observation and publishing in a citable venue is a governance act rather than a marketing one, and it exposes the work to review by people with no stake in the result. Almost nothing else in this index has been through that process. The company also demonstrates it segments its population, reporting results separately across three journey stages, which shows the analytical capability to break results down exists.

What it does not do is break them down by person. No performance figures by age, sex, race, language, insurance type or geography were located, and no model card, training data description or bias statement exists.

Two mechanisms make that gap concrete rather than formal. Voice recognition performance varies with accent, dialect, speech rate and age, and a substantial part of this platform's stated market is Medicare onboarding, meaning an older population where recognition degrades most and where a failed conversation means a member misses an enrolment window. And a personalisation engine optimised on engagement will learn who responds, which tracks phone access, work schedule flexibility and language, so an outreach programme optimised purely on conversion can quietly concentrate effort on the members already easiest to reach and withdraw it from those hardest to.

That second mechanism is the important one here, because unlike a triage tool this platform decides who gets contacted at all. Graded C, above the records with no artefact and below any vendor publishing subgroup performance.

Nothing published addresses responsibility when an automated outbound conversation goes wrong.

The failure modes here are distinctive because the vendor initiates contact rather than responding to it. An agent that gives a member incorrect information about benefits, eligibility or a programme creates a reliance the member acts on, and the member has no reason to doubt it because the call came from their health plan. An agent that calls at a prohibited time, calls a number after consent was revoked, or discloses programme participation to whoever answers the phone creates statutory exposure that lands on the plan rather than on the vendor. And an agent that encounters a crisis disclosure and does not escalate produces a harm nobody may learn about.

None of these is addressed. No indemnity, limitation, performance warranty or accuracy commitment was located, and nothing describes who bears telephone consumer protection exposure when the vendor's platform places the calls using the customer's consent records.

That last division is the specific and unanswered question on this record. The plan holds the consent, the vendor operates the dialling, and the models decide who to call and when. A per call statutory penalty regime multiplied by a nine fold volume increase is a material exposure, and the published material describes the volume increase as a benefit without addressing where the risk sits.

One mitigating fact belongs on the record. The published study reported no safety incidents or escalations, which is an operational observation rather than a contractual position, and it is the only thing here pointing in the vendor's favour.

Graded D.

The integration surface is well chosen for a payer product and it points away from the clinical record entirely.

What exists is real and named. The platform integrates with Amazon Web Services, Google Cloud, Snowflake and Databricks, which is where a health plan's member data actually lives, and announcing those integrations before the Series A indicates they were treated as a prerequisite rather than a later addition. For a buyer whose data estate is a warehouse rather than a record system, this is the correct and harder integration to have built, and it means the personalisation models run against the customer's own curated data rather than a partial extract.

The platform also closes a loop on the outbound side, since the published study measures appointment completion, which implies visibility into whether a scheduled appointment occurred and therefore some connection to a scheduling source.

What is absent is the clinical record in both directions. No electronic health record integration is named, no fast healthcare interoperability resources or health level seven support was located, and nothing describes writing an outreach outcome back into a clinical system. For a payer product that is defensible scope rather than a failure, and it does limit what the platform can do: a member's conversation with a voice agent does not become part of any clinical record, so a clinician later seeing that member has no visibility into what was discussed or agreed.

That matters most for the chronic condition, mental health and substance use programmes, where an outreach conversation may contain clinically relevant information that stops at the plan boundary.

Graded C: strong warehouse integration, no clinical interoperability.

One architectural fact is disclosed and it is the useful one, with the specifics unpublished.

The fact is the integration surface. Naming Amazon Web Services, Google Cloud, Snowflake and Databricks tells a buyer that the platform connects to the customer's own data infrastructure rather than requiring a bulk export into a vendor environment. That is a materially better starting position than an opaque managed service, because it implies member data can remain in an environment the customer already governs, already has agreements for, and already knows the residency of. For a health plan with an established cloud posture this answers a large part of the question before it is asked.

What is not stated is whether that implication holds. Nothing describes what the platform copies out of those systems, what it retains, where the personalisation models are trained and hosted, or where call audio and transcripts come to rest. A platform that reads from a customer warehouse may still process and store extensively on its own infrastructure, and the two arrangements have very different residency consequences.

The voice layer is the harder gap. Placing outbound calls requires telephony infrastructure and speech processing, neither of which is described, and both introduce parties and locations the cloud integration story does not cover. No region, residency option, subprocessor list or retention position for recordings was located.

HITRUST certification implies a controlled environment behind all of this and does not tell a buyer where it is.

Graded C, above the records with no architectural disclosure at all, and short of any published residency position.

Nothing is published. No price, no unit of charge, no tier structure, no implementation fee, no contract term and no minimum was located in vendor or third party material.

The unit question is the substantive one and it is entirely open. This platform could reasonably be priced per member under management, per personalisation model, per call minute, per completed contact, or as a platform subscription, and each produces a wildly different cost curve for the same programme. A per minute or per contact model exposes a health plan to cost that scales with outreach volume, which is precisely the thing the product encourages a customer to increase. A per member model does the opposite. Nothing published indicates which applies, and this is the single most consequential missing fact on the record.

The published study gives a buyer partial material for a business case without the cost side. Roughly 75 hours of staff time saved per 1,000 contacts is a genuine efficiency figure a plan can value against its own loaded labour rate, and a 9 times increase in outbound volume against human agents implies a favourable unit economics story the company has evidently modelled internally. Publishing the saving while withholding the price leaves the return calculable in one direction only.

The adjacent voice agent market publishes rates, with per minute pricing common and stated openly by several platforms, so this is a disclosure choice rather than a category norm.

Graded D.

Deep on the payer side, absent on the provider side, and that is a deliberate positioning rather than a gap in execution.

Buyer coverage is genuinely broad within its chosen half of the market: health plans, payors, health services organisations and digital health companies, with named programme types spanning pharmacy operations, Medicare and individual plan onboarding, preventative care outreach and benefit utilisation. Those are distinct operational functions with different regulatory calendars and different member populations, and supporting all four is real breadth.

Clinical coverage is stated across substance use, mental health, chronic conditions and musculoskeletal care, and the published deployment sits in gastrointestinal specialty care. That spread suggests the platform is genuinely condition agnostic, which follows from what it does: the models decide who to contact and when, and the clinical content of the conversation is configured per programme rather than built into the product.

Journey coverage is the most specific evidence available. The published study segmented three distinct patient journey stages and reported the largest effect among members who had registered without scheduling, which demonstrates the platform operates across a member lifecycle rather than at a single touchpoint.

What is absent is the provider side entirely. No health system, hospital or medical group deployment was located, and the platform's integration surface points at data warehouses rather than clinical systems. A provider organisation evaluating this would be an atypical customer.

Graded B: excellent coverage of one half of the market, none of the other.