Synthpop

Language models do the defining work in the agents, inside an orchestration layer that carries value of its own. Document intelligence that turns fax referral packets into payor-ready orders, conversational voice agents that sit through payor calls and take resupply orders, and payer-aware reasoning do not exist without the models.

Around them sit capabilities that would remain with the models removed: API-first eligibility and benefits checks through clearinghouses and payor portals, scheduled multichannel outreach, duplicate suppression, write-back to practice management and billing systems, and exportable audit trails and outreach reports for CMS and accreditation reviews. The payor agent itself tries the API and portal first and places a call only when those leave no answer.

The oversight structure is described and the threshold is missing. On the payor agent, every case ends either in a structured, code-level result or in a documented hand-off to the customer's team when a payor will not resolve it, with a transcript and reference numbers attached.

Resupply calls escalate exceptions to staff with full context, and the company lists the guardrails it tests against: staying on task, not being talked into shipping supplies that are not due, and handing off to a person when it should. Each customer's call flow goes through three to four review rounds with the customer's own team before production.

What is not published is where the line sits: what makes an exception real, at what point a voice agent hands a patient to a person, and what happens after a wrong eligibility result or order has been written back.

The suppliers are named and the approach is described, without version or update discipline. The home page names OpenAI, Claude and Gemini as the AI vendors the company partners with, and the platform is described as document intelligence, payer-aware reasoning and conversational voice agents under an orchestration layer. The company's 2023 launch announcement says it fine-tunes language models on de-identified medical records.

The payor agent's order of operations is published: API first, then portal, then a phone call that navigates the payor's phone menu and captures a code-level result. No model version, evaluation method or accuracy figure tied to a test set is published, and it is not stated which components are learned and which are rules.

Model providers are named, and nothing beyond them is. The home page lists OpenAI, Claude and Gemini under an enterprise-level partnership with leading AI vendors, and names the platforms the agents connect to: billing (Change Healthcare, Waystar, Experian Health), payor systems (Availity, Optum, Zelis), communications (Twilio, Five9, RingCentral) and record and resupply systems. The agent is also offered through Google Cloud's Gemini Enterprise. The company publishes no subprocessor list and no retention or training terms with those providers, and does not say which provider processes which workflow or whether record content reaches each.

Named customers and production figures, without a named deployment behind the numbers. Pivotal Health is quoted saying referrals are now processed in under a minute, MedBridge appears as a customer, and logos include Hampton Homecare, LabRetriever and UCLA. The resupply page gives the most method: 9.5 percent more orders per patient and 77 percent fewer staff hours against human-only outreach, across more than 10,000 patients and 12 operating accounts in a four-week window.

The payor page reports 58 percent of manual calls eliminated and roughly twice as fast time to therapy for PAP patients at an unnamed national sleep-care deployment in production for over two years. The home page adds 96 percent faster processing, 3,500 packets processed daily with half fully automated, and more than 9 million patients supported. No figure is tied to a named organization with a date and baseline.

One categorical commitment is on record, and it is dated. In the company's 2023 launch announcement its chief technology officer says it developed a method to de-identify medical records so that language models can be fine-tuned without risking exposure. Today's payor page adds that protected health information is minimized and access is logged.

Nothing current states whether that method still governs training now that the company works with OpenAI, Anthropic and Google models, which de-identification standard it meets, whether one customer's data improves agents serving another, or how long referral packets, call recordings and transcripts are kept. The privacy policy covers only the website and email, and allows aggregated or anonymized website data to be used for any purpose.

HIPAA compliance is claimed and nothing is published behind it. The home page carries a HIPAA Compliant badge, the payor and resupply pages answer HIPAA and SOC 2 to the security question, and the company's February 2026 funding release describes it as fully HIPAA compliant. The privacy policy (last updated April 2026) covers the website and marketing email only, and no business associate agreement, service privacy notice or statement of role is published.

Business associate status is structural for the work described: the agents read referral packets, verify coverage, call payors and patients, take payments and write results back to the customer's systems.

The company presents a SOC 2 attestation as its own. The home page carries a SOC 2 Certified badge, the payor and resupply pages answer HIPAA and SOC 2 to the security question, and its February 2026 funding release describes it as SOC 2 audited. The resupply page adds that payments are PCI DSS compliant. The report type, period, auditor and scope go unstated, and no trust center or security page is linked.

No FDA position is stated and none is needed for the work described: referral intake, eligibility and benefits verification, authorization checks and resupply ordering are administrative. Different rules govern this work, and answering to them is not a regulatory pathway.

Resupply of CPAP masks, glucose sensors and other consumables is subject to Medicare DMEPOS documentation and refill-contact rules, which the company addresses by exporting outreach records for CMS and accreditation reviews. Outbound AI voice calls and texts to patients carry Telephone Consumer Protection Act consent requirements, and prior authorization work sits under CMS's 2026 electronic prior authorization rules. When a resupply patient raises a problem such as a leaking mask, the agent is described as adapting rather than escalating.

Process is described and nothing measured is published. The resupply page lists the guardrails each voice agent is tested against and describes three to four review rounds with the customer's own team before production, and the payor and resupply products keep full audit trails of every attempt. No governance framework, evaluation method, error rate by workflow or subgroup result is published.

The groups that matter here are callers and documents: voice agents calling older CPAP and diabetes patients meet hearing loss, accents and second languages, and a patient the agent cannot understand may simply not reorder, leaving a gap in therapy with no visible error; referral extraction degrades on poor faxes and handwriting.

Mechanisms exist to check an output, and nothing published stands behind one. The payor agent returns a structured, code-level result for every check, with call and confirmation IDs and a full transcript for every phone verification, and every attempt across API, portal and phone is logged and exportable. Resupply keeps a timestamped record of every call and text for CMS and referral-source questions.

No terms of service, warranty, accuracy commitment or remediation is published: the footer links only a privacy policy covering the website. The person harmed by a wrong eligibility result or a missed resupply order is a patient whose therapy is delayed or whose supplies do not arrive.

Named systems with write-back, without verification of depth. The home page names the platforms the agents connect to through APIs: Epic, Cerner, Brightree, NikoHealth, Bonafide, Salesforce, Practice Fusion and DrChrono on the record and resupply side; Change Healthcare, Waystar, Experian Health, Availity and R1 on billing; Availity, Optum, Zelis and TriZetto on the payor side; and Twilio, Five9, RingCentral, Talkdesk and Zoom Contact Center for communications.

The payor and resupply pages say every result and attempt is written back to the customer's practice management, record and billing systems. No marketplace listing, integration documentation or per-system method is published, and nothing shows which of these connections run in production at which customer.

A hosted, API-first service with its location implied. Customers log in to a hosted application, the agents reach customer systems through APIs, and Google Cloud appears among the company's partners, with the agent also offered inside Google Cloud's Gemini Enterprise. No hosting provider, region, residency commitment, tenancy model or single-tenant option is stated. The processing path matters because three model providers are named, so record content from referral packets and call transcripts may reach OpenAI, Anthropic or Google depending on the workflow.

No price is published. The product is sold through a working session that maps the customer's workflows and quantifies return, and the agent is also offered through Google Cloud's Gemini Enterprise, which may carry its own commercial terms. The resupply page states demo readiness within a day and production within two weeks.

Value is framed per patient ($8 to $16 saved per patient onboarded) and per order captured, yet the unit of charge is left open: nothing says whether pricing runs per workflow, per transaction, per call minute, per patient or per order, or whether any part is tied to orders captured or authorizations obtained.

Coverage is named, with validation behind part of it. The company now leads with durable medical equipment and home care: resupply of CPAP and diabetes consumables, and payor verification for PAP therapy at a national sleep-care deployment. It also sells referral and order intake, coverage and authorization, and patient engagement to providers more broadly.

Resupply is supported by production figures across 12 operating accounts and payor verification by a two-year deployment, both unnamed; named customers include Pivotal Health on referral intake and MedBridge, with Hampton Homecare, LabRetriever and UCLA among the logos. Hospital and payer settings are claimed at platform level with nothing shown.