Hippocratic AI
250 million patient calls at $9 an agent-hour with the deepest guardrail architecture in healthcare AI, every safety number self-published and unverifiable, and a CEO who said on the record he is not building for Europe because the AI Act makes his agents high-risk.
PRICING
| Polaris Pro | from $9/agent-hour | clinical, non-diagnostic voice AI for patient-facing conversations; billed on active agent time only |
| Polaris Flash | from $5/agent-hour | lower-cost tier published on the vendor's product page; what it gives up versus Pro is not documented |
| App Store creator share | 5% of base + 70% of premium | licensed US clinicians who build and certify agents; earnings capped at $5,000 per agent |
| Enterprise deployment | Quoted | integration, safety certification and scoping run through custom contracts on top of the hourly rate; no free trial |
The hourly rates are vendor-published floors, not contracted prices, and the vendor's /pricing URL returns 404. Usage-based billing means the bill tracks call volume, so a busy quarter costs more. Enterprise contracts may be structured on covered patients rather than hours.
checked 2026-08-27 · vendor pricing page
Element scores
Strengths
Two numbers carry this audit. 250 million completed patient voice interactions across 300+ live clinical use cases, and $9 an agent-hour against a US registered-nurse median near $39. That combination is why 50+ healthcare organisations across provider, payer and life sciences signed up, and why the company raised $444 million at a $3.5 billion valuation. The architecture is the interesting part. Polaris 5.0 is a 5-trillion-parameter constellation with a 700-billion core and more than 30 supervising models watching for privacy breaches, medication errors, adverse events and escalation triggers, any of which can end a call. Agents are structurally barred from diagnosing or prescribing and hand off to a human nurse in real time. Certification runs through 7,700+ US-licensed clinicians and 775,000+ test calls before anything reaches a patient, and the RWE-LLM evaluation framework behind it is published rather than hidden. On top of that, August 2026 brought Agentic Orchestrators: teams of agents under a supervising orchestration brain that decides who calls which patient when, across 30+ orchestrators for readmission reduction, Star Ratings, chronic care and trial enrolment. That is the most credible multi-agent deployment in this database. The voice engineering is also real: cough detection, background-voice separation, post-stroke speech handling, mid-call Spanish switching, 1.5 second time-to-first-audio.
Honest dings
Every safety number in the previous paragraph comes from the vendor. Polaris is proprietary, so as one independent critic put it, independent benchmarks are impossible; the 50+ comparisons against GPT, Claude and Gemini are Hippocratic's own tests of Hippocratic's own model, and the clinician testing was simulated calls, not real patients. '99.89% correct advice, zero severe harm' is a self-graded exam, however many people sat it. Second, the EU problem, which for a European buyer is the whole story. In July 2025 CEO Munjal Shah said on the record that the company would not focus on European expansion because the AI Act classifies its agents as high-risk, and that Europe had moved too fast to regulate. Fourteen months later the site names no EU deployment and the 'six countries' claim goes unbroken-out. So the EU-facing read is: not sold here, by choice, on regulatory grounds. Third, the developer surface is closed. No public API, no SDK, no MCP, no A2A. You build inside their console or you do not build. Fourth, the meter. Usage-based billing on active agent time means budgets move with call volume, and $9 an hour is a published floor, not a quote. And fifth, the labour politics are not noise: National Nurses United has campaigned against exactly this deployment pattern, and the AI Now Institute flags the thinness of independent scrutiny across healthcare AI generally. A telling critique from 2025 still stands, that the tasks these bots do best (patient education, outreach calls) are not the tasks eating most nursing time, which is documentation and hunting for people and equipment.
Sources (14) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- newswise.com/articles/hippocratic-ai-announces-next… — Official announcement 13/08/2026: Agentic Orchestrators with a supervising orchestration brain; 30+ orchestrators across payer, provider, pharma and med tech; 250M+ patient interactions, 300+ live clinical use cases, $444M raised; 700B primary model plus 30+ supervising models; validated by 7,700+ US clinicians across 775,000 calls; 99.89% correct advice with zero severe harm claimed; contracts allow the company to take responsibility (accessed 2026-08-27)
- prnewswire.com/news-releases/hippocratic-ai-launche… — Official Polaris 5.0 release 30/04/2026: 5-trillion-parameter constellation on a 700B core; contextual ASR halving word error rate on drug names; cough detection and background-voice separation; IVR navigation; clinical escalation across seven body systems; interdependent document intelligence; mid-call English/Spanish switching; 1.5s time-to-first-audio; frontier thinking models (GPT 5.4 Pro, Claude Opus 4.7, Gemini 3.1 Pro) benchmarked as 'too slow for voice'. Note this is a corrected release: the original headline claimed 'outperform every frontier model', amended to 'major frontier models' (accessed 2026-08-27)
- hippocraticai.com/new-products — Official product page: Polaris Pro 'as low as $9/hour' and Polaris Flash 'as low as $5/hour'; cross-platform products AI Front Door, AI Physician Front Door, Nurse Co-Pilot, AI Anti-abrasion, AI Call Supervisor, AI Certified Agent Builder; full orchestrator catalogue for provider, payer, life sciences and med tech (accessed 2026-08-27)
- hippocraticai.com/safety — Official five-phase safety framework: constellation architecture, output testing with 7,500+ US-licensed clinicians and 725,000 test calls, human clinical supervision, real-time escalation to human nurses, cross-validation of simulated against real-world performance over 200M+ interactions (accessed 2026-08-27)
- hippocraticai.com/real-world-evaluation-llm — Official RWE-LLM framework: 307,000+ unique calls evaluated with multi-tier error flagging by severity, nursing review then physician adjudication, continuous feedback loop; company argues output testing beats input validation (accessed 2026-08-27)
- fiercehealthcare.com/ai-and-machine-learning/hippoc… — Independent (Fierce Healthcare) 09/07/2025, and the single most important source for an EU buyer: CEO Munjal Shah states the company will NOT focus on European expansion because the EU AI Act classifies its agents as high-risk, asserting Europe moved too fast to regulate and the law stifled business. Also documents the KPMG international collaboration and the value-based-care / single-payer fit (accessed 2026-08-27)
- hippocraticai.com/hippocratic-ai-announces-series-c… — Official Series C release: $126M at $3.5B valuation, $404M total at the time, led by Avenir with CapitalG, General Catalyst, a16z, Kleiner Perkins, Premji, UHS, Cincinnati Children's, WellSpan; 50+ enterprise healthcare organisations across provider, payor and life sciences (accessed 2026-08-27)
- telecareaware.com/are-hippocratic-ai-and-ai-nurses-… — Independent critical analysis summarising Thomas W. Dinsmore's machine-learning critique: testing not done on real patients; accuracy claims not externally benchmarked and 'independent benchmarks are impossible with a proprietary model like Polaris'; many named partners likely non-paying; the WHO 15-million shortage figure is global and dominated by roles that cannot be done virtually; the tasks the bots do best are not the tasks consuming most nursing time per McKinsey workload data (accessed 2026-08-27)
- usagepricing.com/blueprint/hippocratic-ai — Independent pricing analysis 10/06/2026: ~$9 per agent-hour billed on active patient time only, no seats and no tokens, /pricing URL 404s; App Store creator share of 5% of the ~$10 base plus 70% of premium capped at $5,000 per agent; enterprise contracts may be structured on covered patients rather than hours (accessed 2026-08-27)
- agenticindex.io/vendors/hippocratic-ai — Independent vendor profile 06/07/2026: 300+ prebuilt agents across 25 specialties; certification pipeline of automated evals, simulated testing by 6,237 nurses and 308 doctors, then customer sign-off; no-code Agent Trainer; running one agent around the clock costs about $6,480 per month; usage-based billing makes budgets spike with call volume (accessed 2026-08-27)
- siliconangle.com/2025/11/03/hippocratic-ais-valuati… — Independent (SiliconANGLE) 03/11/2025: clinicians can prototype an agent in under 30 minutes and start testing within three to four hours; agents specialised for chronic care management, medication checks and post-discharge follow-up on kidney failure and congestive heart failure (accessed 2026-08-27)
- nationalnursesunited.org/press/national-nurses-unit… — Independent labour-side evidence. National Nurses United survey: AI often contradicts nurses' clinical judgment; 29% of nurses cannot change software-generated assessments or categorisations; 40% cannot modify predictive scores to reflect clinical judgment. Context for any patient-facing agent deployment, not a finding about Hippocratic specifically (accessed 2026-08-27)
- ainowinstitute.org/publications/research/expanding-… — Independent research-institute position: names Hippocratic AI as a VC-backed firm premised on AI replacing nurses, notes its presence at a US Congressional hearing on a state-AI-regulation moratorium, and argues there remains limited independent scrutiny of how healthcare AI tools are actually integrated and who bears the risk (accessed 2026-08-27)
- hippocraticai.com/polaris-3 — Official Polaris 3.0 page, kept for the version trail and the integration list: 4.2T parameters across 22 models, clinical accuracy 96.79% (1.0) to 98.75% (2.0) to 99.38% (3.0), 6,234 clinicians and 307,038 test calls, integrations with Epic, Cerner, Salesforce, athenahealth, eClinicalWorks, NextGen, Modernizing Medicine, Allscripts and MEDITECH; policy-document quoting accuracy 86.4% to 99.4% and complex scheduling error rate 8% to 0.5% between 2.0 and 3.0 (accessed 2026-08-27)