Hippocratic AI

The AI is the product. Every line the company sells is a language model agent or a layer that runs on them: patient-facing voice agents for providers, payers and life sciences (the company lists more than 1,000), orchestrators that coordinate teams of agents toward an outcome, an AI Call Supervisor that scores every call, and a builder that drafts and tests new agents. They run on Polaris 5.0, described as a 700-billion-parameter primary conversation model surrounded by more than 30 specialized support models that check medications, labs, escalation and scheduling in real time.

Remove the models and nothing remains to buy: there is no scheduling system, record or content library underneath. What a buyer purchases is the conversation itself.

The limits are published and the escalation rule is not. The agents do not diagnose, prescribe, handle hospice, mental health disorders or children under two, and hand the call to a human nurse mid-call when a situation needs one, passing the context along; the safety page lists the situations tested (suicidal ideation, domestic abuse, chronically ill patients downplaying symptoms), and the company reports more than 31,000 patients escalated to nursing care and escalation accuracy across eight categories. Every agent passes simulated calls, clinician evaluation and a mandatory clinical review gate before it goes live.

The missing part is the threshold. The eight categories and the criteria are not listed; the company says a separate set of models decides whether a concerning reading is an emergency or a nonurgent signal for standard channels; over 95 percent of calls run on autopilot, with 0.5 to 1 percent sampled for safety afterward.

The model is identified, versioned in public and characterized by its training and test data. Polaris is the company's own constellation, released in numbered versions with what changed: 3.0 (March 2025) as 22 models and 4.2 trillion parameters, 5.0 (April 2026) as a 700-billion-parameter primary model with more than 30 support models and over 5 trillion parameters in all. Release notes give per-version accuracy on clinical advice, scheduling, speech recognition subtasks and languages, and the company describes its training data as production patient calls and clinician-scored test calls. The constellation post explains the design: narrow support models with verifiable jobs, and two-level verifiers (a real-time judge and a slower offline model) on every tool call.

Not stated: whether the primary model started from an outside base model, and how a customer is told when a new version reaches its agents.

A structurally short chain at the model layer, with the host named and the rest unlisted. The company builds and versions its own models, and a 2024 patent covers the constellation design. The May 2026 press item names DigitalOcean's cloud as the platform the calls run on, and NVIDIA is an investor and collaborator on speech and inference. The constellation post credits the company's own speech work with a 50 percent reduction in recognition errors over off-the-shelf enterprise speech recognition, without saying whether an outside speech service remains in the path.

No subprocessor list is published, and whether the primary model began from an outside base model is not stated.

Published research is company-led and answers narrower questions than the headline outcomes. The peer reviewed work covers three studies: a retrospective WellSpan study (JMIR 2025, company and WellSpan authors) of 1,878 patients in September 2024, where fecal test opt in was 18.2 percent among Spanish speakers and 7.1 percent among English speakers; a retrospective study of 4,415 calls (JMIR Formative Research 2026) finding multicall memory lengthened calls without changing satisfaction; and a qualitative usability study with Roche and Genentech on education for macular degeneration injections (JAMA Ophthalmology 2026). The safety framework is a preprint (medRxiv 2025).

The headline outcomes are the company's own and come without method: a 30 percent reduction in readmissions, a 360 percent increase in chronic care team capacity, 12 times average ROI, 275 percent more survey returns. "Zero incidents of harm since inception" is stated without a definition of harm or how incidents are detected.

Safety engineering is described in detail; how patient data is kept and used is not. The company publishes the parts that catch a wrong answer: support models that check medications, labs and escalation during the call, an online judge before a patient is told something and an offline model that re-reads calls, a call supervisor that scores every call, and clinician testing before an agent goes live.

Data handling is the weak side. The only privacy policy published covers the website and mobile apps, and says information is kept as long as required and that de-identified data may be used. The company says Polaris 3.0 was built from over 1.85 million real patient calls and Polaris 5.0 on more than 180 million patient interactions, so customer calls feed model development; multicall memory keeps patient details between calls; and marketing plays deidentified excerpts of real patient calls. No retention period, de-identification method or opt-out is published.

HIPAA compliance is claimed, and the document that would support it covers the website. The FAQ says the company is HIPAA compliant and holds HITRUST and SOC 2 certifications with consent-aware handling of patient data, and marketing call excerpts are de-identified in compliance with HIPAA. No business associate status, agreement or tier is stated, and the privacy policy applies to the website and mobile apps, not to the patient calls the agents make. The trust portal at trust.hippocraticai.com may cover the business associate question.

The agents take in protected health information on every call.

Recognized certifications are named; the reports behind them are not on the main site. The company announced HITRUST e1 certification for its full agentic platform in July 2025 (press release on its press page), and its FAQ and trust badges name HIPAA, HITRUST and SOC 2. A trust portal exists at trust.hippocraticai.com and may hold the reports, though whether they can be retrieved without a request is not stated. The SOC 2 type, period and scope are not given on the main site. HITRUST e1 is the entry tier, covering a core set of controls rather than the fuller r2 assessment.

A scope statement is published, and one step inside it is triage. The company states throughout that its agents do not diagnose or prescribe, do not handle hospice, mental health disorders or children under two, and focus on outreach, intake, care navigation and support, escalating clinical judgment to licensed humans. That is a statement of intended purpose disclaiming diagnosis and treatment. For pharmaceutical clients, agents capture adverse events for pharmacovigilance and avoid off-label discussion.

The company's May 2026 constellation post says that when a patient reports a concerning value, separate models "triage severity", deciding between immediate escalation and a nonurgent route through standard channels, and over 95 percent of calls run without a person on the line. No regulatory position addresses that step, and no reasoning is published for why the product sits outside device regulation.

A governance process with a method is published, with results in the group most at risk. The RWE-LLM framework (a medRxiv preprint, 2025, company authors) describes clinician output testing: 6,234 US licensed clinicians scored 307,038 test calls against harm categories, with nurse review and physician adjudication feeding fixes between model versions. For callers who do not speak English, the Polaris 3.0 release reports 99.83 percent accuracy in Spanish and 99.09 percent across nine other languages against 99.38 percent overall, and a peer-reviewed WellSpan study (JMIR 2025) found Spanish-speaking patients engaged more than English-speaking ones (fecal test opt-in 18.2 versus 7.1 percent, 1,878 patients).

The nine languages are pooled, and no results are published by age, hearing, accent, cognitive status or health literacy, the groups where a phone agent is most likely to mishear or be misunderstood. The testing is run and reported by the company.

An error rate with its method and denominator is published, and the liability the company says it takes is not written down in public. The RWE-LLM preprint reports, across 307,038 clinician test calls, correct medical advice rising from about 80 percent before Polaris to 96.79, 98.75 and 99.38 percent over three versions, advice with potential minor harm falling from 1.32 to 0.13 and 0.07 percent, and severe harm at 0.00 percent in the latest; the company now cites a 99.9 percent clinical safety score. Those are simulated calls scored by clinicians the company hired, not an independent audit of live calls.

The home page says "Liable: Our contracts take responsibility" and the agent builder page says "Your ideas. Our guarantee."; neither states what is covered, the remedy or its limits, so both are recorded rather than graded. No route is published for a patient given a wrong answer.

Named record systems, with direction and production evidence left to the buyer. The home page reports more than 50 EMR integrations including Epic, and the Polaris 3.0 release says the agents have been integrated with Epic, Cerner and Salesforce and can integrate with athenahealth, eClinicalWorks, NextGen, Modernizing Medicine, Allscripts and Meditech. Named health systems run the agents (WellSpan, OhioHealth, Universal Health Services, University Hospitals), and the agents also navigate the phone menus of pharmacies, labs and other providers.

What is read from and written to each system is not described, no marketplace listing or integration documentation is cited, and no customer is tied to a named connection.

A single hosted service, with the host named and the location not committed. The agents run on the company's platform, and a May 2026 press item describes the platform holding its safety score across 10 million patient calls on DigitalOcean's cloud. No region, data location, tenant separation or customer-hosted option is published, although the company deploys with partners in the UAE, Japan, Israel and the UK. Multi-call memory stores patient details between calls on the platform.

No price is on the company's pages now, and how it sells is discoverable. The $9 per agent hour rate the company announced in 2024 does not appear on the current site, which instead offers a meeting that scopes a first use case, integrations and an ROI model for the buyer's population.

Buyers are named by type (providers, payers, pharma and medtech), deployments start with one use case and expand, and the constellation post notes that running a frontier model 30 times per conversation would cost over $100 per hour, which frames the cost argument without giving a rate.

Coverage is named precisely, with published results behind part of it. The company lists agents for providers, payers and life sciences across 17 specialty filters and use cases from post-discharge calls and screening outreach to adverse event intake and trial retention, and it names what it will not do: hospice, mental health disorders and children under two. Named deployments include WellSpan, OhioHealth, Cincinnati Children's, Universal Health Services, University Hospitals, Medical Mutual, Sheba Medical Center and Burjeel Holdings.

Published results cover colorectal screening outreach, chronic care calls and patient education in ophthalmology; each agent passes clinician testing before launch, but those results are not published by setting.