Hyro

The AI agents are the product. Scheduling, prescription refill requests, provider search, frequently asked questions, password resets and IT help desk requests are each handled by conversational agents across the call center, website, text and mobile apps, and the Care Intelligence reporting layer is built from the records those conversations leave.

The chief technology officer's August 2026 account describes the stack: semantic parsing mapped against a knowledge graph built from each health system's own data since 2019, scheduling broken into seven smaller models with one job each, and large language models from OpenAI, Gemini and others used inside that design with fallback between them. With the models removed, what remains is a set of integrations and a reporting screen with nothing to report on.

The agents act on their own for scheduling, prescription refill requests, provider search, password resets and questions answered from sources the health system defines, and hand complex calls to staff. The oversight structure is described: the customer defines and restricts the knowledge sources an agent may draw on, each conversation can be unpacked to show which sources produced each response, and every conversation is recorded for review.

The missing part is the threshold at which an agent stops and transfers to a person, and what it does when a caller raises a clinical question. The responsible AI page says the design will 'eliminate' hallucinations and misleading responses by restricting data sources.

That claim concerns what the agent adds, and the per conversation source trace is the mechanism that lets a reviewer catch an addition; the company's own figures ('10x Less Hallucinations', a 96 percent accuracy rate) concede that errors remain, so it is no promise that the output is always right. The home page also promises to handle unlimited calls and messages at full capacity, which sets the scope an escalation rule has to cover.

The approach and the suppliers are named; the version and update discipline is not. The company describes a knowledge graph built from each customer's website, provider directory and records, semantic parsing in use since 2019, and, for scheduling, seven focused models that each receive only the context their step needs, with foundation models from OpenAI, Gemini and others inside that design.

It also says its own models are trained on the ways real patients ask for care, drawn from interactions handled over several years. What a buyer cannot establish is which model version produced a given answer or when the system changed. The same account says a call switches foundation model provider when one times out, so the model behind an answer can change mid-conversation without notice.

The model layer is named. The chief technology officer writes that the stack uses OpenAI, Gemini 'and others', with fallback between vendors so a call can move from one to another without the caller noticing. The contact center and platform partners in the call path are named on the integrations page: Cisco, Five9, Genesys, NICE and Twilio for telephony and messaging, plus Salesforce, ServiceNow and the record systems, though these are often the customer's own systems rather than the vendor's suppliers.

What is missing is the full list. 'And others' leaves the model providers open, the speech recognition and synthesis providers are not named, and there is no subprocessor list or statement of the retention and training terms that govern patient data once it reaches OpenAI or Gemini.

Named deployments carry outcome figures, each linked to a case study. Tampa General Hospital is tied to a 56 percent fall in daily call abandonment and a 58 percent fall in wait times; Prisma Health to more than $300,000 in annual savings and an 80 percent chat resolution rate in its first year; Overlake to 90 percent chat resolution; Inova to 338,000 automated calls, around 4,000 staff hours a month and an 8.8 times return; Baptist Health to close to $1 million in savings; and Weill Cornell Medicine to 47 percent more appointments booked online.

Intermountain Health and Hackensack Meridian Health are also named, and the company reports serving 2,500 hospitals, clinics and facilities, which is scale of use rather than evidence of benefit. The headline figures (an automation rate above 85 percent, 5.5 times return within six months, 35 percent lower operating costs) carry no source, method or denominator, and none of the results is independently evaluated.

The training position is stated, and it runs toward use: the chief technology officer writes that the company's models are trained on the ways real patients ask for care, knowledge that lives in the interactions it has handled for years. The privacy policy adds that anonymized or de-identified information may be used and shared without restriction, for example to improve the services, and that an internal retention policy prevents perpetual retention, with no periods given.

By design the customer defines and restricts the knowledge sources an agent may use, and each foundation model step receives only the context it needs. Missing: whether conversations are de-identified before they train anything, what reaches OpenAI, Gemini and the other providers and whether they keep it, and retention windows for call audio and transcripts.

HIPAA compliance is claimed across the company's material, and nothing behind the claim is published. The responsible AI page lists 'HIPAA-compliant sources of information', the site footer carries a HIPAA Secured seal, and the Epic Showroom listing and the August 2026 ServiceNow announcement both describe 'fully HIPAA-compliant' agents. None of it states business associate status or says a business associate agreement is offered.

The privacy policy reaches the platform and its end users, including patients who reach a health system through the agents, but says nothing about protected health information and places consents and legal compliance on the customer.

The company presents a SOC 2 attestation as its own: its responsible AI page states 'SOC II Compliant', and the site footer carries the AICPA SOC logo beside a HIPAA Secured seal from HIPAA Secure Now and a GDPR badge. The SOC logo is the generic AICPA mark, so the report type, its period and the scope it covers are not stated and are the buyer's first questions; a Type I report tests control design at a point in time, a Type II tests operation over a period.

The HIPAA seal is a compliance program vendor's mark, not an audit of how security is operated. There is no trust center, security page, penetration testing statement or subprocessor list on the site, and the privacy policy says the company cannot guarantee information will not be compromised. The responsible AI page's promise of 'air-tight locks on patient privacy and system security' is an assurance, not a control.

No device claim is made, and the core work is administrative: scheduling, prescription refill requests and status updates, provider and location search, password resets and questions answered from sources the health system approves. The boundary questions sit at the edges. The integrations page pairs the agents with Infermedica's symptom checker to offer 'personalized health assessments'; the assessment in that pairing is Infermedica's own product.

Physician search promises to 'match patients to the right point of care', and the October 2025 funding announcement names clinical AI agents as a development area. No regulatory position is stated anywhere, and no page describes what an agent says when a caller describes symptoms or asks a medication question.

Multilingual operation is sold as the answer for community health centers ('Expand access for your community with multilingual AI'), and nothing is published on how the agents perform by language, accent or caller group. A caller the system fails to understand leaves no complaint and no record, so that silence is the central gap.

What is published is policy and product design: a responsible AI page built on three pillars (explainability, control and compliance), the ability to trace each response to its sources, customer control over which sources the agents use, and three figures (a 96 percent accuracy rate, '10x Less Hallucinations' and '78% More Resolutions') with no method, baseline, sample or period. No governance body, evaluation process or AI management system certification is described.

Mechanisms exist to challenge an output, and nothing stands behind it. Each conversation can be unpacked to show which knowledge sources produced a response, so a health system can trace a wrong answer to its cause and correct the source; every conversation is recorded, and the chief technology officer says the company can show a health system exactly what happened on a specific call.

No warranty, remedy or service credit attached to the output is described, and the published figures (a 96 percent accuracy rate and '10x Less Hallucinations') carry no method or denominator, so neither can be tested. A patient given a slot that does not exist, or whose refill request went nowhere, has no route against the vendor described anywhere on the site.

The Epic integration is bidirectional and verifiable in Epic's own marketplace. Hyro is listed on Epic's Showroom, where the listing describes appointment scheduling and management, prescription refills, MyChart password resets and provider and location searches, and the home page describes using Epic's APIs for patient record identification, scheduling management and prescription support: reading the record and writing to it.

Named Epic customers are live, Tampa General Hospital among them. Beyond Epic, the integrations page names Oracle for real time patient data, athenahealth, eClinicalWorks, Kyruus Health for provider search and scheduling, Salesforce Health Cloud with a two-way sync, the Cisco, Five9, Genesys, NICE and Twilio contact center platforms, and, from August 2026, ServiceNow, where an app on the ServiceNow Store creates IT service desk tickets from voice requests. The integrations outside Epic are named without the same marketplace evidence.

A hosted service that layers over the health system's existing phone, web and record systems; no customer-hosted option is described, and location is implied rather than committed. The privacy policy allows personal data to be transferred to the company's local or foreign subsidiaries and affiliates for storage or processing, and to service providers in other countries.

The chief technology officer describes failover across foundation model vendors and across regions ('if a region goes down, we route around it'), so processing can move between providers and locations during a call. No hosting provider, region, residency commitment or tenant isolation model is published.

No price, rate card or unit of charge is published, and the product is sold through a demo. What is public is the shape of the offer: the modules (AI agents, call center automation, Care Intelligence reporting and Proactive Px outbound campaigns), the skills sold within them (scheduling, prescription refill requests, provider search, frequently asked questions, IT help desk, site search and ACA and Medicaid coverage outreach), and the buyers they are packaged for: health systems, federally qualified health centers, children's hospitals and payers. A public return on investment calculator estimates time and cost savings, and the home page claims a 5.5 times return within six months, but neither shows what the product costs.

Its work is healthcare administration automation, centered on patient access in the health system call center and website: appointment scheduling and management, provider and location search, prescription refill requests, MyChart password resets and, since August 2026, IT service desk tickets created from voice requests through ServiceNow.

Coverage is organized by organization and channel rather than specialty: health systems, federally qualified health centers (sold on multilingual access), children's hospitals and payers (member and agent experience), across calls, websites, text and mobile apps, with an outbound line for ACA and Medicaid coverage renewal outreach.

Validation sits behind the health system part, with named deployments at Tampa General Hospital, Prisma Health, Inova, Baptist Health and others; the home page names no customer for the children's hospital, community health center or payer lines. No specialty-specific scheduling capability is described.