R1 Phare OS

AI agents do the core work in each Phare module, and the platform around them carries value of its own. Phare Claim's coding, Phare Flow's denial appeals and payer calls, and the medical necessity reviews in Phare UM all come from models that read medical records, while orchestration agents pass work between those specialist agents and R1 staff.

Other parts would stand without the models: Payer Atlas, the connectivity layer with more than 1,500 payer connections, the data platform that unifies EHR and financial data, and the Command Center dashboards. Phare is also built to orchestrate labor alongside agents, and R1 presents it as the system that runs its own revenue cycle teams as well as a product for health systems, so the models are the engine of the modules rather than the whole offer.

Phare is designed for autonomous operation, and R1 describes when work leaves the agents. The system uses R1's analytics to decide when a task must be routed to a person, keeps a full audit trail and history of each decision, and runs under what R1 calls expert human oversight and supervision. Orchestration agents pass work between specialist agents and human staff, and the Command Center lets revenue leaders see human and AI operations and intervene.

The routing threshold is the missing part: no confidence level, dollar amount or case type is named that sends a claim, code or appeal to a person. Phare Flow is described as appealing any denial type autonomously when given a complete record, and Phare Claim as producing production ready code lists, so a buyer cannot tell which outputs reach payers without human review.

Phare runs on large language models, voice models and computer use models, according to R1's chief product officer, with specialist agents trained on R1's historical decisions and orchestration agents routing work among them. Phare Claim is described as context aware, clinically guided AI producing complete code lists with sequencing, principal diagnosis and supporting evidence rather than matching terms.

None of the models is identified for the product, and no versioning or update practice is given beyond a self learning loop. R1's lab, R37, has written about post training NVIDIA's Nemotron model for long context reasoning on denials, which shows the kind of work underway without saying which model runs in Phare today.

Phare is built by R1's own lab on large language, voice and computer use models, and no outside model provider, cloud host or subprocessor is named for it. The specialist agents are trained on R1's own historical data, which shortens the chain for the coding and appeal logic, while the underlying models and their hosting stay unidentified.

R37's published work on post training NVIDIA's Nemotron points to open models R1 can run itself, without stating that Phare does so in production. The trust center's documents sit behind an access request, and no subprocessor list or terms on retention and training with model providers can be read beforehand.

Production figures come from R1 alone, with no method, sample or period behind them. At launch in October 2025, R1 said Phare Claim reaches autonomous coding accuracy as high as 97 percent for emergency and physician office visits and automates more than 30 percent of inpatient coding, and that Phare Flow fully automates more than 20 percent of payer documentation requests by dollar volume. The Phare OS page adds more than $250 in revenue lift per discharge for Phare Audit.

Named customers stand behind the platform without outcome data: Providence says Phare OS supports care delivery across its 51 hospitals in seven states, Singing River Health System signed an exclusive agreement at launch, and UF Health's finance chief cites R1's AI capabilities. How accuracy was measured, against which reference coders and on how many charts, is not given.

Customer data is the training base by design. R1 says its specialist agents are trained on historical decisions, actions and outcomes, two decades of revenue cycle work and some 280 million transactions a year, and that Phare keeps capturing decisions and outcomes in a self learning process that improves performance over time.

R1 refers to an AI governance framework that sets guardrails for data use, compliance and transparency, without saying what it allows. No retention period, de identification practice or limit on pooling one health system's records to train models used for another is stated, and nothing describes safety engineering, such as how a wrong code or appeal is caught before it reaches a payer, beyond routing uncertain work to people.

HIPAA appears among the compliance items in R1's trust center, and no business associate position for Phare is set out without signing in. The trust center lists HIPAA beside SOC 1, SOC 2, HITRUST, PCI DSS and the California privacy laws, with the documents themselves released through an access request.

R1 handles patient records for hundreds of providers and accesses more than 670 million patient records, which makes it a business associate in the ordinary course. How its agreement covers Phare's models, which subcontractors it binds, including any outside model providers, and whether module customers sign the same agreement as full outsourcing partners all go unstated.

R1 runs a SafeBase trust center listing SOC 1 and SOC 2 reports, HITRUST, PCI DSS and alignment with the NIST Cybersecurity Framework, with the reports themselves behind an access request. No report types, assessment versions or dates are shown, so currency and scope cannot be checked before asking.

The same center outlines endpoint protection, disk encryption, endpoint detection and response, firewalls, security event monitoring, incident response and penetration testing, and names Ascension, Intermountain Healthcare, Lifepoint Health, Providence, Quorum Health and Sutter Health as customers. Whether Phare, launched in October 2025, falls within the scope of the current reports is not stated.

Phare makes no device claim; it is revenue cycle software for coding, authorizations, utilization review and collections. Its clinical reach is the medical record: Phare Intelligence reads documentation to support medical necessity determinations, codes and appeals, and Phare UM brings payer risk intelligence to utilization review on the provider side.

Those outputs are billing and coverage decisions governed by payer rules and coding guidelines, not care decisions. The regime that bites is billing compliance, which R1 addresses through its compliance and integrity program, so the administrative positioning matches what the product does.

R1 names an AI governance framework and a compliance and risk program, and shows no evaluation behind them. The framework is described as providing guardrails for data usage, compliance and transparency with human oversight, and the trust center lists a compliance and risk program document behind its access gate.

No results show how coding accuracy, automation rates or appeal success vary by payer, service line, patient population or hospital type. For a system trained on twenty years of past billing decisions, the risk a buyer would weigh is that it repeats patterns in that history, and nothing describes how the models are tested for it.

Every Phare decision carries an audit trail and history, which lets a health system trace a code, appeal or authorization back to how it was produced. Uncertain work is routed to people, and the Command Center shows where agents and staff acted.

Nothing commits R1 to stand behind Phare's output for module customers: no coding accuracy guarantee, liability for payer recoupment after an incorrect code, or remediation commitment. R1 promises appeals it calls airtight and codes it calls precise, and none of that is backed by a stated remedy. Its outsourcing partnerships may carry performance terms, and none are described, so a health system buying a module has only its contract to rely on.

Integration is claimed broadly, with no record system named for Phare. R1 says Phare OS is fully integrated across all major EMRs and banks, that its data platform unifies real time clinical and financial data from EHRs and beyond, and that it connects to more than 1,000 payers, with Payer Atlas listing more than 1,500 payer connections and 600 million transactions a year.

The customers named run large EHR estates, and R1 helped Intermountain Health through an EMR implementation across three states, but no EHR vendor, marketplace listing, interface type or write back path is described for Phare itself. A buyer cannot tell whether Phare reads and writes in a health system's own record system or works from extracts.

Where Phare runs and where patient data rests goes unstated. The platform is presented as secure by design and governed by default, with a data platform that pulls clinical and financial data from health system EHRs, and no hosting provider, cloud region, tenancy model or on premises option is named.

R1's trust center lists a document titled Geographic Location of Data, released only through its access request, so the answer exists for a buyer willing to ask. R1 also runs large global operations teams that work alongside the agents, and where those teams sit relative to the data is not stated either.

No price, fee structure or unit of charge is offered for Phare OS or its modules, and every path leads to a personalized demo. The selling model is visible: Phare OS is modular, so a health system can start with one module, such as prior authorization, coding audit or denials, and add others, or take the platform inside a broader enterprise partnership in which R1 also runs revenue cycle operations.

Physician groups and critical access hospitals have a separate entry point. Whether modules are licensed per claim, per discharge, as a share of collections or as part of an outsourcing fee is left to the contract, and no implementation cost or term is stated.

Hospital and health system revenue cycles are the setting, from patient access through coding to collections, with named systems behind part of it. Providence runs Phare OS across 51 hospitals, Singing River Health System on the Mississippi Gulf Coast is implementing it, and R1 says it works with 95 of the top 100 US health systems, including Ascension, Sutter Health and Intermountain Health, mostly through its outsourcing business.

Service line coverage is specific where figures exist: the coding claims cover emergency department and physician office visits, with inpatient coding partly automated across specialties. Physician groups and critical access hospitals are named as markets with their own offer, without a named Phare deployment in either.