RadiantGraph

The models are the engine of the core product. The personalization layer builds per person models that decide which members to contact, about what and through which channel, and the AI Voice Agents and Care Guidance Agents hold the conversations themselves; the published study measures those agents against human callers. Outreach at this volume and specificity does not exist without the models.

The platform around them carries value of its own. The platform page lists a Health Data Engine and Member 360 that unify a plan's fragmented member data, Smart Cohorts for segmentation, Smart Spend, a Journey and Content Studio for campaigns, and omnichannel orchestration across voice, text and voicemail. Remove the models and a health plan still has a data unification layer, segmentation and a campaign orchestration tool.

Autonomy is described and oversight is stated as a phrase. The published study describes agents that hold natural conversations while adhering strictly to set conversation flows, the platform page promises healthcare specific guardrails, and every call in the study was audited afterward by human healthcare professionals. The September 2026 announcement says Care Guidance Agents work within a non diagnostic scope and integrate with a human care management team for more complex situations. An audit after the call and a handoff 'for more complex situations' leave the moment that matters unrouted: what happens during a call when a member discloses a crisis.

That moment is foreseeable here. The home page promises to reach every member, the agents run outreach and coaching for behavioral health, substance use and chronic condition programs, and at this volume some members will disclose suicidal thoughts or acute symptoms mid call. Nothing published says what the agent listens for, when it transfers, to whom, at what hours, or whether members are told they are speaking with an AI agent and can ask for a person. The study's zero safety escalations cannot tell a buyer whether the path worked or was never triggered.

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 personalization 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 personalization capability rather than operating independently. Scale is given in model terms, personalization 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 shows 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 personalization 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.

The architecture is described and no model provider is named. The platform page sets out the parts (a Health Data Engine, personalization models, AI Voice Agents and Care Guidance Agents) and carries a Databricks endorsement of RadiantGraph's use of the Databricks Data Intelligence Platform, and the company integrates with customers' AWS, Google Cloud, Snowflake and Databricks environments. That places data platforms in the chain; it does not say which model hears or answers a member.

An outbound voice agent needs speech recognition, language understanding and generation, speech synthesis and telephony, and nothing published says which of those are RadiantGraph's own and which come from a provider. Each unnamed provider would handle member voice audio under its own retention and training terms, including audio from behavioral health and substance use programs. The privacy policy names service providers only by category and publishes no subprocessor list.

A deployment has been through peer review. The study, run with Oshi Health, a virtual first gastrointestinal clinic, covered nearly 8,800 members across three journey stages over seven weeks in late 2025 and was published in Lecture Notes in Artificial Intelligence from the 2026 International Conference on Artificial Intelligence in Medicine. It reports appointment completion up 3.4 times overall and 24.7 times among members who had registered without scheduling, a ninefold increase in outbound call volume against human agents, about 75 hours of staff time saved per 1,000 contacts, and no safety escalations. The home page states the result as over 370 percent improvement in program adoption and up to 2,400 percent in the best cohorts, with a 0.08 percent opt out rate, figures that do not match the multiples exactly.

The evidence has four limits. It is a collaboration between the vendor and its own customer; the headline figures are multiples with no absolute baselines; it is a short conference paper with no concurrent control arm described; and the other figures in circulation (15.8 times enrollment lift, 70 percent team efficiency, brand impact ten times that of human agents) carry no method.

Certification covers the handling and nothing covers the learning.

On handling there is something real. 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 a basis to work from before any conversation begins.

On stewardship nothing is published. No statement addresses whether member data, call recordings or transcripts inform model development, whether personalization 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 personalization 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 deidentified the way a structured field can. The named programs 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 favor. 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 good architectural answer. Nothing published confirms it.

A HIPAA status is stated and supported. The company's announcements describe the platform as HITRUST certified, SOC 2 Type 2 compliant and HIPAA compliant, and the privacy policy (updated 27 March 2026) says RadiantGraph processes protected health information about the patients and members of its customers and directs those members to the customer for questions about it, which is the business associate position in practice.

What is missing is the contract. No business associate agreement is offered or described, no terms are published, and nothing states retention for call recordings and transcripts. The substance use programs named among the use cases bring in federal confidentiality rules for substance use disorder records, which impose obligations beyond HIPAA, and no published material shows the platform configured for them.

RadiantGraph states HITRUST certification and SOC 2 Type 2 in its own announcements, from the September 2024 cloud integration release on its own library to the September 2026 Care Guidance Agents release, which describes the company as HITRUST certified, SOC 2 Type 2 compliant and HIPAA compliant. Those are recognized attestations presented as the company's own.

Nothing published lets anyone examine them: no trust center, no report request process, no audit period, auditor or certification scope, no penetration testing statement, vulnerability disclosure policy or subprocessor list. Scope matters especially here, because a certification earned on the personalization platform may not extend to the voice agents launched in 2025 or the Care Guidance Agents launched in 2026, and the 2026 release says the platform is 'built to' HITRUST standards, which is not the same statement.

No clearance is claimed and the stated scope is non diagnostic: the September 2026 announcement says Care Guidance Agents operate within a non diagnostic scope and hand complex situations to a human care management team. Outreach, enrollment, scheduling and reminders are administrative.

Care Guidance Agents go further, delivering condition specific coaching, education and adherence support directly to members with diabetes, hypertension, heart disease, asthma, behavioral health and musculoskeletal conditions, including under the CMS ACCESS model. Talking to a member about their own condition is where a regulatory question opens if the conversation becomes advice about treatment, and the company does not say where the scripts stop or who clinically approves them.

The regulation that plainly applies sits elsewhere. Automated outbound calling is governed by federal telephone consumer protection rules on consent, calling windows and revocation, with penalties per call, and nothing published addresses consent capture or opt out handling. Several states require disclosure that a caller is an AI agent. For the substance use programs named among the use cases, federal confidentiality rules restrict even confirming a person's participation to whoever answers the phone.

RadiantGraph describes a governance process rather than only principles. Its home page reports that 100 percent of calls in its published deployment were audited by human healthcare professionals; its September 2026 announcement says each condition specific Care Guidance Agent undergoes hundreds of hours of testing and validation by US based healthcare providers and human expert auditing, overseen with a named clinical advisory board; and the platform page promises full visibility into every AI decision, with audit trails and guardrails. That is a described evaluation process with no results from it published.

What is missing is any breakdown by person. No performance by age, sex, race, language, insurance type or geography is published, and two mechanisms make that gap concrete. Voice recognition varies with accent, dialect, speech rate and age, and much of the stated market is Medicare, where a failed conversation can mean a missed enrollment or redetermination window. And a personalization engine that learns who responds will learn phone access, work schedules and language, so outreach optimized on conversion can drift toward the members easiest to reach. Because this platform decides who gets contacted at all, the second matters more.

Mechanisms exist to catch a wrong conversation after the fact, and nothing stands behind the output. The platform page offers full visibility into every AI decision and interaction, with audit trails, and the home page reports that every call in the published deployment was audited by human healthcare professionals. That lets a plan find a bad call; nothing published commits RadiantGraph to anything when one happens, and the member has no route against it.

The failure modes are distinctive because the vendor initiates contact. An agent that gives wrong information about benefits, eligibility or a program creates a reliance the member acts on, because the call came from their plan. An agent that calls outside permitted hours, calls after consent was revoked, or discloses program participation to whoever answers creates per call statutory exposure that lands on the plan. The plan holds the consent, the vendor operates the calling and the models decide who to call and when, and nothing published says who bears that exposure.

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 suggests they were treated as a prerequisite rather than a later addition. For a health plan 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 personalization 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 is published, 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 programs, where an outreach conversation may contain clinically relevant information that stops at the plan boundary.

The privacy policy (updated 27 March 2026) says RadiantGraph's servers are generally located in the United States, with 'certain exceptions' it does not describe. The integration surface is also stated: the platform connects to a customer's own AWS, Google Cloud, Snowflake and Databricks environments, and the platform page describes a composable layer over the customer's existing EHR, CRM and data warehouse. That suggests member data may largely stay where the plan already governs it, which would be a good answer; nothing confirms what the platform copies out, where personalization models are trained and hosted, or where call audio and transcripts rest.

The voice layer is the harder gap. Outbound calling needs telephony and speech processing, neither of which is described, and both add parties and locations the cloud integration story does not cover.

No price is published, but how the product is sold is discoverable. RadiantGraph sells a modular platform (AI Voice Agents, Care Guidance Agents, Health Data Engine, Smart Cohorts, Smart Spend, Journey and Content Studio) to health plans and health systems, for named programs: care gap closure, health risk assessment completion, new member onboarding, retention, clinical program enrollment, risk adjustment visits and Medicaid redetermination.

The unit of charge is the open question and it changes the economics. The platform could be priced per member, per personalization model, per call minute, per completed contact or as a subscription; a usage based price rises with outreach volume, which is exactly what the product is designed to increase, and the study reports a ninefold rise in outbound calls. The study's 75 hours of staff time saved per 1,000 contacts gives a buyer the saving side of a business case and not the cost side.

Deep on the payer side, with health systems now named as buyers alongside health plans. The value based care tie is the product's own list of programs: closing gaps in care before they affect quality scores, health risk assessment completion, risk adjustment visits, new member onboarding and retention, clinical program enrollment, and Medicaid redetermination. In September 2026 the company announced approval for the CMS ACCESS model, a Medicare program that pays for technology enabled chronic care with payments tied to patient outcomes, with Care Guidance Agents supporting behavioral health and musculoskeletal care there.

Clinical coverage is stated across substance use, mental health, musculoskeletal care and chronic conditions including diabetes, hypertension, heart disease and asthma, and the published deployment sits in gastrointestinal specialty care, so validation stands behind one part of that range. Provider side deployments are not described in detail, and the integration surface points at data warehouses rather than clinical systems.