Hippocratic AI

The product is the AI agent; customers buy agent conversations by the hour.

Agents barred from diagnosis and prescribing; escalation to humans when clinical judgment is required; constellation support agents can end non compliant conversations; certification pipeline includes simulated call testing by thousands of nurses and physicians plus customer sign off.

Polaris constellation architecture publicly documented, including design rationale and published safety research.

Substantial disclosure at the infrastructure and speech layers, and a deliberate silence at the base model. What is named is named clearly: a collaboration with NVIDIA covering inference microservices and speech technology for recognition and synthesis, a company operated cluster on named hardware generations, hosting through a multi account and multi cluster Amazon Web Services strategy that separates production from development, and further stated collaborations with Google, DigitalOcean and Modular.

The model layer is claimed as in house and it is claimed with detail rather than as a slogan: a patented constellation of a primary conversational model plus more than twenty specialist support models covering medication, labs, dosage, social determinants, privacy, compliance and safety, versioned publicly across three generations with parameter counts and accuracy figures attached.

For a voice product that is better than most, because the speech recognition and synthesis layer is where recorded patient audio actually goes and it is usually the least disclosed part of the stack. Held below the top grade on one specific and deliberate gap. The company says a frontier base model is not sufficient for clinical voice work, which implies the primary model begins from one, and it does not say which.

The published statement that it builds on the safest and smartest available technology is the sentence where a provider name would go, and no provider is given. No subprocessor list was located either, so a buyer cannot see which of the named collaborators processes protected health information and under what terms, and the two are not the same question as which vendors appear in an engineering blog. Ask for the base model lineage and a subprocessor list with change notification.

Published safety research with stated methodology on the Polaris constellation architecture, which is genuine technical evidence and rare in this lane. Operational scale is substantial and consistently reported (2.5 million patient calls rising to a reported 10 million, 8.95 out of 10 patient satisfaction, a reported 99.9 percent safety score).

Held back from A because the clinical outcome claims published across use cases (30 percent readmission reduction, 360 percent chronic care capacity increase) are vendor reported without methodology, cohort sizes, or comparison groups, and have not been independently validated.

Safety is the product thesis, not a disclosure afterthought. The Polaris constellation architecture deploys more than 25 task specific models with a dedicated safety focus, agents are restricted to non diagnostic patient facing tasks by design, every clinician authored script is safety tested before release, and the company publishes a safety score across its call volume. Combined with HITRUST e1 covering PHI controls, the stewardship posture is evidenced rather than asserted.

HITRUST e1 certification covers the 44 most critical security controls and is the healthcare specific framework built to evidence HIPAA safeguards, which is materially stronger than an unevidenced HIPAA alignment claim. Held back from A because BAA availability and execution terms are not published on the vendor's own site; a buyer must confirm BAA specifics in procurement.

The strongest certification posture in the index. SOC 2 Type II examination completed and issued by an independent AICPA accredited auditor over a multi month observation period (announced August 2025), on top of HITRUST e1 Certification covering the full agentic platform (announced July 2025), which the HITRUST Alliance publicly confirmed. Two independent third party validations, both named and dated, and the company employs a named Chief Information Security Officer. This is what an A looks like on this axis.

The regulatory posture is deliberate and stated consistently rather than left implicit. The product is positioned throughout as non diagnostic and confined to patient facing clinical tasks, which is a coherent strategy for operating outside medical device classification rather than an evasion of it. Several elements go further than positioning.

Adverse event handling and guardrails preventing discussion of off label use are built in for pharmaceutical clients, which engages promotional and pharmacovigilance regulation directly. Escalation paths exist for mandatory reporting situations including suicidal ideation and child protection. A safety evaluation framework has been published as a paper, and agents pass a multi step certification process before deployment.

Held below the top of the band because no formal regulatory determination is published. The non diagnostic claim is asserted rather than reasoned through against the clinical decision support software criteria, no clearance or submission is cited, and no external regulatory review is referenced. Worth asking for the written analysis behind the non diagnostic classification.

Substantive governance structure: a stated Do No Harm principle enforced through a rule that agents are only built where the company believes generative AI can operate safely, clinician validated script review, and a published safety score. Held back from A because no bias or fairness evaluation with stated methodology and no third party AI audit were retrieved, which is the specific evidence this axis is looking for.

This grade is worth reading against the cleared vendors in this index, because it is reached by an entirely different route. Hippocratic AI has no clearance and no device regulator, so none of the reporting, complaint handling and correction machinery that gives a cleared manufacturer its floor applies here. The band is earned by publication instead.

Clinical accuracy is published per model generation and the figures move, from 96.79 percent to 98.75 percent to 99.38 percent across three constellation versions, which is a falsifiable claim a buyer can hold the company to and almost nobody else in this category offers one at all. A safety evaluation framework is published as a paper, so the method behind the number is inspectable rather than asserted.

Agents pass a certification process including simulated call testing by clinicians and customer sign off before deployment, and customer sign off is the closest thing here to an accountability gate, since it puts a named party on the record accepting the configuration.

Escalation paths exist for mandatory reporting situations including suicidal ideation and child protection, which is genuine recourse for the patient at the moments that matter most, and adverse event handling built for pharmaceutical clients engages pharmacovigilance reporting obligations directly. Held below the top grade for three reasons. No indemnity, guarantee or remediation commitment toward the customer was located.

The safety score is vendor defined, vendor measured and not independently audited, so the transparency is real but the verification is internal. And the escalation design covers defined crisis categories rather than ordinary error: a patient given a wrong answer about a medication has no published route to contest it, and no relationship with the vendor through which to try. The published figure also carries a reading a buyer should do out loud.

At the stated accuracy, roughly one clinical claim in one hundred and sixty is wrong, and the company reports patient interactions in the hundreds of millions. Credit the disclosure, then ask what the wrong ones consist of, how severity is distributed across them, and what happens to the patient on the other end of one.

The deepest named integration coverage in this category. Systems of record explicitly integrated include Epic, Cerner and Salesforce, with stated support extending to Athenahealth, eClinicalWorks, NextGen, Modernizing Medicine, Allscripts and Meditech, which between them cover most of the US acute and ambulatory market.

Integration runs in both directions rather than read only: a dedicated agent extracts structured clinical data and notes from patient conversations so the care team receives documentation rather than a transcript. The platform also navigates the interactive voice systems of other providers, laboratories and pharmacies, which extends interoperability past the customer's own estate to the organisations it has to call.

One nuance worth understanding before contracting: a multi call memory layer carries patient context between encounters outside the EHR, so some patient specific information persists on vendor infrastructure rather than in the record. Worth asking how that layer is reconciled with the chart.

Cloud delivered and vendor hosted, with agents configured on the vendor platform rather than deployed into customer infrastructure. Beyond that the published detail is thin. No hosting provider is named, no cloud region or data residency commitment is offered, no tenant isolation model is described, and there is no on premises or customer managed option.

The gap matters more here than it would for a purely transactional service, because the platform maintains a persistent multi call memory of patient specific context, stored in what is described as a HIPAA compliant database and retrieved at the start of later conversations. That means patient derived information accumulates on vendor infrastructure between encounters. Worth establishing where that memory store is hosted, how long entries persist, whether a patient or customer can have them purged, and whether anything can be kept in region.

Headline rate of $9 per agent hour published, with usage based billing terms described.

More than 300 prebuilt agents across 25 specialties, use cases explicitly enumerated.