Cohere Health

The models are the engine of the core module, and the platform carries other value. Cohere's clinically trained models, fine-tuned on clinical documentation, drive touchless prior authorization: 85 percent of requests approved in real time on its home page, 89 percent at Humana in its August 2026 case study.

Around them the company sells a provider portal, clinical policy management, delegated utilization management run by its own nurses and physicians, and managed payment integrity audits, all of which carry value without the models.

Oversight is laid out in outline, with the threshold left out. The AI approves or pends; it does not deny. Cohere's home page states 85 percent of prior authorization requests are approved in real time and the rest are reviewed by a clinician before final determination, and its AI principles say pended requests go to clinical experts and the requesting provider is told the clinical basis.

What is not published is the confidence threshold that decides which requests approve automatically, or what happens after an automated approval is found wrong. The AI Hub also calls the system always clinician reviewed, which sits beside touchless approval, and which approvals a clinician actually sees is not stated.

The approach is named without the version discipline behind it. Cohere's AI Hub describes fine-tuned models trained on real clinical documentation from more than 40 million utilization management records, with reinforcement from clinicians' feedback and performance driven retraining, and says they outperform general large language models on lab values, diagnosis detail and treatment verification.

No base model, version, accuracy figure or confidence threshold is published, and the automation rate appears as 85 percent on the home page and 89 percent in the Humana case study, on different denominators.

A structurally short chain for the core models, with the rest unnamed. Cohere describes its own fine-tuned models, trained on utilization management records and retrained on its clinicians' feedback, which keeps the determination models in house. The base model they were adapted from, the host they run on, and the model behind the reviewers' case review chatbot are not named, and the privacy policy lists service providers by function only, with no subprocessor list.

Named deployments with dated outcome figures, short of independent evaluation. The Humana case study (August 2026) reports a partnership since January 2021 across all 50 states and 5.1 million Medicare Advantage and commercial members, with 95 percent provider digital adoption, an 89 percent auto approval rate, time to service five days shorter, 54 percent fewer nurse review hours and 94 percent of providers satisfied.

The home page reports 17 clients, 25 million covered lives, 47 million plan and provider interactions, 50 percent faster reviews and 30 percent more accurate decisions, and a HEDIS extraction product reports 20 to 25 percent closure rates for high value measures. The AI Hub publishes clinicians' ratings of the reviewers' chatbot (79 percent correct, 60 percent complete, 47 percent would trust it without verification).

None is peer reviewed or independently evaluated, and the claim that the models match or beat expert nurse reviewers carries no figures. The satisfaction figure was measured among providers, the counterparty.

Categorical commitments without the schedule behind them. Cohere's AI principles state the AI never uses explicit patient identifiers, such as names or locations, for care recommendations, and its AI Hub describes a minimum necessary approach that reads only the clinical information a decision needs.

The AI Hub also says the models are trained on real clinical documentation from more than 40 million utilization management records and retrained on clinicians' feedback, so case data shapes the models; whether one plan's data trains models used for other plans is not stated. No retention or deletion schedule for submitted clinical documents is published. Safety engineering is not published either: guardrails, how a wrong automated approval is caught, and how a safety event is handled. The machine learning that fast tracks requests from high performing practices profiles individual clinicians, which patient privacy rules may not reach, and nothing says what governs it.

Business associate status is stated and supported by a privacy document. The privacy policy (effective 18 September 2026) states that Cohere receives and processes patient information, including protected health information, as a business associate to its clients and at their direction, with the client determining how that information is used and shared. The AI principles add that the AI never uses explicit patient identifiers such as names or locations for care recommendations. The business associate agreement itself and its subcontractor terms are not published.

A recognized certification is named without the artifact. Cohere's AI Hub states the platform is HITRUST r2 certified, and its news pages announce the original r2 certification and a recertification. No trust portal or certificate is published for a buyer to open, and no SOC 2 report is claimed, so the certification date and scope are the questions to raise. HITRUST's mapping to other frameworks does not add further certifications.

No device claim is made and the product is scoped as health plan utilization management, which FDA does not regulate as a device. The AI approves or pends and does not deny; adverse determinations are made by clinicians. The position Cohere states is compliance with HIPAA, CMS interoperability and state prior authorization rules, and it announces URAC accreditation and NCQA utilization management accreditation and reaccreditation. Those accreditations examine how the utilization management operation runs, and they are not a clearance.

Responsible AI is committed to in policy language, with no evaluation published. Cohere's AI principles state that an AI Governance Board oversees equitable treatment regardless of background, demographics or geography, that models are continuously monitored for bias, drift and accuracy, and that standards of care are applied with fairness across members. No membership or charter for the board, no monitoring method or cadence, and no subgroup results are published. Prior authorization models trained on historical case decisions can carry past patterns forward, and fast tracking practices with a strong approval history can widen gaps between practices.

Mechanisms let a provider challenge an output, and nothing stands behind it. Cohere links its review criteria for providers from its home page, explains the clinical basis to the requesting provider when its AI approves or pends a request, and routes every request it does not approve to a clinician before a final decision. Nothing published gives an error rate, a commitment on what happens when an automated approval or a pend proves wrong, or a remedy toward the plan or the patient. The AI Hub reports clinicians rated the reviewers' case review chatbot correct 79 percent of the time, with no method or sample size.

A named system with one direction of flow. Cohere's April 2024 announcement describes an integration with Epic's Payer Platform that lets providers submit prior authorization requests inside their Epic workflow, stated as available with Epic instances nationwide. Its API product embeds utilization management into plan systems, and providers can also submit through its portal. No listing or integration documentation shows determinations or status returning into the EHR, and no health system using the Epic route is named.

A hosted service with location implied. Cohere offers engagement options (a platform, delegated services and APIs), but each runs as its hosted service, and no hosting provider, region or tenancy is published. The privacy policy says personal information may be processed and stored in the United States and other countries; that sentence covers the website and services generally, not the patient data Cohere handles as a business associate.

The selling model is published; the price is not. Cohere sells to health plans in three shapes: a platform the plan's own clinicians run, delegated utilization management run by Cohere's clinicians for named specialties, and APIs embedded into plan systems; payment integrity is sold in house or as a managed audit service. The only economics published are returns, up to 18x for prior authorization and 9x for payment integrity, which say what a plan gets back and not what it pays. The fee basis, whether volume, cycle time or savings, is not stated, and fees that scale with savings from denials would pull against the approve or pend design.

Coverage is named with validation behind part of it. Cohere names delegated programs in musculoskeletal, cardiovascular, gastrointestinal, sleep, diagnostic imaging and sepsis care, and products across utilization management, payment integrity, appeals, care management, claims operations and quality.

The Humana case study supports prior authorization across all 50 states for Medicare Advantage and commercial members; no published validation is broken out by specialty or by inpatient and outpatient review. The buyer is the health plan, and providers use the product without choosing it.