Attune

The models are the engine of the core module, and the platform around them carries its own value. Agentic Voice depends on learned components throughout: a self hosted speech transcription layer, language understanding, and natural multilingual synthetic voices. What the agent says and does, though, is set by deterministic orchestration the customer configures: the company markets conversations that follow the customer's rules without deviation, built in its Flow Builder, and sells Campaigns and Analytics modules beside the voice agent. Remove the models and a campaign manager, a flow builder and reporting remain, with no voice to carry them.

The record's earlier reading, an agent given a goal and latitude over how to reach it, no longer matches how the company describes the product, which it now positions against unpredictable generative behavior.

Deterministic flows and warm transfers are described, and the company's own copy promises the output is right. Conversations follow rules the customer configures, the agent passes the full conversational context to staff when escalation is necessary, and it detects voicemail and retries; outbound campaigns call members with no person on the line. What is not published is the escalation criterion: what the agent does when a post discharge patient answers that they feel worse, or discloses distress, in the check in the Agentic Voice page itself demonstrates.

The home page says conversations follow your rules "with complete accuracy. No guesswork, no hallucinations, no deviations", and that every interaction is accurate. That promises the output is right, which covers what a reviewer would miss as well as what the agent adds, so it grades here whatever the flow design: deterministic logic fixes what the agent is allowed to say, not whether it heard the caller correctly, and a misheard answer written back to a record is still wrong.

The architecture is described in general terms and nothing is identified. The home page describes deterministic orchestration that follows the customer's rules, a self hosted medical grade transcription layer, natural multilingual voices and clinical language accuracy, and the Agentic Voice page says the logic is deterministic when you need it to be. No model, version, training data or evaluation method is named on the home, Agentic Voice or privacy pages, and "when you need it to be" implies parts of the conversation are generated rather than fixed, without saying which.

The site does not say which models transcribe, understand and speak, what they were trained or tuned on, or how customers learn that a model changed. For a voice product the figure that matters most is recognition accuracy on the population called, and medical grade transcription is asserted without it.

The architecture is described, part of the chain is in house, and no model provider is named. The home page says transcription runs on a self hosted layer, which keeps call audio from reaching an outside speech recognition service at that step, and that conversations run on deterministic orchestration. The providers behind language understanding and the synthetic voices are not named on the home, Agentic Voice or privacy pages, and the privacy notice mentions data processors without naming any. The trust center, a hosted portal, may carry a subprocessor list.

The hosting provider is not named either, and nothing states the retention and training terms each provider is bound to once a call reaches it. With transcription in house the chain may be short, and a published list would show how short.

A quantified unnamed deployment and a named customer, with no method. The Agentic Voice page reports that a major health plan ran the agent across its Medicare population to complete health risk assessments: every member dialed, 49 percent answered and a 75 percent completion rate, with client agents moved to other work. The home page quotes the chief executive of Gold Kidney on the platform's reliability, without figures. The company is the claimed source.

The plan is unnamed, no period or baseline is given, the completion rate's denominator is not stated (answered calls or all members), and completion is not correctness, since an assessment completed on misheard answers counts the same as one completed well. No peer reviewed or independent evaluation appears on the home, Agentic Voice or privacy pages.

General assurances, and one fact about where audio goes. The home page states that data is encrypted in transit and at rest, that transcription runs on a self hosted layer, and that the company is HIPAA compliant. The only privacy document is a website notice: its retention clause (no longer than necessary, then deleted or anonymized) and its rights section cover website and marketing data, and it does not mention patient calls or protected health information.

No retention period is published for call audio as distinct from transcripts and structured results, nothing says whether calls train the agent or any model, and nothing covers consent. The system places calls patients did not ask for and records what they say, so three questions a patient would ask go unanswered: whether they are told the call is recorded, where several states require every party to consent; whether the agent says it is software; and what happens to the audio. The sample calls on the home page open with the caller's organization and a name check, not a statement that the caller is automated. No safety engineering is published beyond deterministic flows and warm transfers.

Compliance is claimed, and the only privacy document covers the website. The home page carries a HIPAA Compliant badge in its security section, alongside HITRUST and ISO 27001, and the navigation and footer link a trust center. The website privacy notice does not mention protected health information, patient calls or business associate status; the trust center may carry a business associate position.

No business associate agreement, subcontractor flow down or breach timetable is published, so the claim stands without the document that would support it. The agent places calls to patients and members and writes results into the customer's systems, which is business associate work for every provider and plan it serves, and the patients called are not party to that agreement.

Recognized certifications are named as the company's own, and the artifacts sit in a trust center. The home page's security section says Attune is HITRUST certified and ISO 27001 certified, with badges for HITRUST, HIPAA and ISO 27001, and a trust center is linked from the navigation and footer. The trust center may release reports openly or only on request; the open site does not say.

The HITRUST assessment type (e1, i1 or r2) and dates, the ISO 27001 certification body and scope, and any SOC 2 report are not named on the home page, and a certificate's scope decides whether it covers the voice platform or only the company's own office systems.

No device claim is made, and the product is scoped as non clinical, with one marketed workflow at the edge of that scope. The home page says Attune automates the work that does not require clinical judgment and lists non clinical questions and routing; most workflows are scheduling, reminders, outreach, benefit and claims calls and structured assessments. The Agentic Voice page, though, demonstrates a post discharge check in that asks whether the patient feels better, the same or worse and whether they are taking their medications as prescribed, and the home page plays a voicemail relaying test results. The voicemail passes on a result the organization determined and the check in asks set questions; what the agent does when a patient answers worse is not stated. That is a tension with the scope line rather than a contradiction, and it is where a stated regulatory position would matter.

The law that bites is better defined: telephone consumer protection rules for automated calls and texts, HIPAA, and Medicare Advantage rules where assessments feed risk adjustment.

Nothing is published on how the agent's behavior is tested, and multilingual operation is marketed as a strength. The home page offers natural, multilingual voices whose pace, tone and language adapt to context and culture, and clear communication across populations; the Agentic Voice page says the voice supports diverse patient populations. No per language, per accent or per age result is published on the home, Agentic Voice or privacy pages. Governance appears as phrases (every decision is governed; continuous improvement with governance), the audit trail keeps each call, and the analytics report how callers feel rather than whether they were understood, so nothing published reads calls for misunderstanding.

Two exposures follow. The flagship deployment called an entire Medicare membership, where hearing loss, slower turn taking, dysarthria and accented speech are more common and speech recognition degrades. And health risk assessments completed outside a treating encounter draw federal scrutiny because they can generate risk adjusted payment; nothing says what a completed assessment is used for or what separates gap closure from coding.

Mechanisms exist to find a bad call afterward, and nothing stands behind the output. The home page describes audit ready traceability, every interaction auditable, and deterministic logic the company says makes the agent's behavior predictable; warm transfers pass the full conversational context to staff when escalation is necessary, and analytics report on calls. They give a way to contest a bad call: a customer can reconstruct what the agent said and did.

No warranty, indemnity, accuracy commitment or remediation obligation appears on the home, Agentic Voice or privacy pages, and the only terms linked from the footer cover Attune's own text messages. The home page's "every word is accountable" names no remedy and no one to hold to it, and no error rate is published, only a 75 percent assessment completion rate. A patient given wrong information by the agent has no route against the vendor on anything published.

The product must integrate to function at all, since booking an appointment requires reading a live schedule and writing to it, and the company describes reading and writing through interfaces, webhooks and secure transfers. One capability stands out: the agent navigates phone trees and manages being placed on hold, which means it can operate as a caller into organizations that offer no interface at all, substituting a phone call for an integration where none exists. That is a pragmatic answer to a real problem in this market. No record system, scheduling platform or telephony partner is named, so the depth is entirely asserted.

A hosted voice platform with its location implied. The agent sits in the telephony path, reads and writes through interfaces, webhooks and secure transfers, and runs transcription on a self hosted layer, so call audio is transcribed on infrastructure the company runs rather than sent to an outside recognition service; data is encrypted in transit and at rest. The company is based in Chicago and sells to United States organizations. The website privacy notice says personal information may be transferred outside the country where it was collected, under safeguards, and it does not reach patient calls.

No cloud provider, region, tenancy model or residency commitment for call audio or transcripts is published on the home, Agentic Voice or privacy pages; the trust center may state some of it. For a system calling a whole Medicare membership, whether audio is kept or only transcripts and results is the question that matters, since a recording identifies the speaker by voice alone.

No price is published, and how the product is sold is discoverable. Attune sells a platform of four modules (Agentic Voice, Flow Builder, Campaigns, and Analytics and Reporting) to providers and to health plans, with named workflows for each: scheduling, referrals, transitions of care and preventive outreach for providers; onboarding, benefit education, claims status, Stars aligned campaigns and assessment administration for plans. The company says organizations usually start with one workflow and scale, and the route in is a demo.

What a buyer cannot establish before contact is the unit of charge: per call, per minute, per completed task or per member, which shift the cost of a long or failed call between the parties. The company launched publicly in January 2026 and does not state its funding, so its runway is unknown to a plan handing it outreach to a whole Medicare membership.

The product is a voice agent for patient access: scheduling, reminders and referral coordination for providers, and onboarding, benefit and claims status calls for health plans.

Coverage is stated clearly by function, and the one quantified deployment is unnamed. For providers the company names scheduling and reminders, transitions of care, referral coordination, preventive care outreach, pre operative and post discharge check ins and non clinical questions; for health plans, onboarding and benefit guidance, claims status, Stars aligned campaigns, outreach for screenings and gaps in care, and health risk assessment administration, which is where the value based care work sits. The setting is the telephone, so the patient's home; voices are multilingual with no language count given; geography is the United States.

The only deployment with figures, Medicare assessment outreach at a major health plan, names no plan, and the customer quoted on the home page, Gold Kidney, is not tied to a workflow. No function is validated at a named site.