EviCore by Evernorth prior-authorization screening

What this models

This example runs on the EviCore-class delegated prior-authorization review network: 11 components and 24 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 2 assumed · 9 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• baseline

  D48-derived new org (Phase 6, clinical-decision-support). TOPOLOGY. Eleven nodes, all documented, none decorative. ONE model with THREE documented inputs - the submitted request, a criteria library the vendor writes itself, and the client's purchased configuration - because that third input is what makes this deployment what it is: an operating point bought by the paying principal is an input to the scoring layer. A worklist, because the governed quantity is the SIZE of the routed-review set: the score decides who is looked at rather than what is decided, and the routing threshold is the parameter the investigation found adjustable. Three record stores because the record documents three: the determination file, the criteria library, and the contract configuration. Two operator classes inside the pipeline plus the client-facing configuration function, and two reviewers who both sit OUTSIDE the vendor - the treating clinician on a peer-to-peer call and an appeal that runs through the client insurer. Absences are derived too: no enforcement node (the documented failure at the notice step is what the letter tells the member, carried here on the appeal channel's inbound edge, and a node would double-count it), no guardrail (no automated output screen is documented; the bound on this model is a rule about authority - a physician alone may deny - which is drawn as the topology, since the model has no pathway to a denial), no externalBoundary (no egress beyond the contracted pipeline is documented), no retriever, and no second model.

• baseline

  LEG RE-DERIVATION (2026-09-22, car LEG-2). This board was re-drawn at the coarsest granularity that still distinguishes every documented mechanism of the deployment, and it lands at eleven nodes and twenty-four edges. No node changed. Three edges were folded into survivors that already carry the same documented flow, and each folded fact is stated on its survivor. (1) The former direct pathway from the scoring layer to the physicians is carried by the file the score writes and by the escalation from the nurse tier, which is the order the record itself gives: below-threshold requests route to nurses, then to physicians. (2) The record write of an appeal outcome is stated on the overturn check, which is where the appeal's documented effect already sat; reach stays on the separate inbound pathway, as before. (3) The return write from a peer-to-peer call is stated on the inbound pathway that carries the call, with the asymmetry named there, because the record documents the offer and documents that nothing is published about what the calls change. Three edge kinds are drawn here in fewer instances than the source model this network is derived from uses; the source is unchanged, and the difference is a drawing decision recorded in this file's derivation comments.

• baseline

  NO operatorToModel EDGE, and the absence is the finding. Nothing in the record documents a reviewer's disposition, an overturn or an appeal outcome travelling back to the scoring layer, so no evaluate-and-confirm pathway is drawn - the corpus idiom for a human checking a model's output would assert a feedback channel this record does not show. The one human hand on the operating point is commercial rather than clinical, and it runs through a store: the account team writes a client's purchased posture into the contract configuration, and the configuration is what the scorer reads. That two-hop shape - nobody touches the model; the model reads what was sold - is the structural claim this network exists to make.

• baseline

  DEMAND 3 / CAPACITY 2. Demand 3: one review operation between about 100 million covered people and the care their clinicians asked for, for more than 100 client insurers across imaging, cardiology, oncology and other delegated lines, under state turnaround clocks that Connecticut's 196-file sample found missed, with a former reviewer's account of a required pace of at least 15 case decisions an hour. The annual request volume is quantified nowhere public. Capacity 2: with no scoring layer this is a licensed, credentialed utilization-review function doing the whole job - the denial is already reserved to a physician and the company's own position is that its algorithms accelerate approvals rather than make decisions - so the human counterfactual is real. It is 2 rather than 3 because the one closed enforcement loop found its violations in exactly that human channel (two denials without an appropriate clinical peer, one appeal decided by the initial denier, determinations past their clocks), and 2 rather than 1 because nothing here documents a hollowed-out desk. Nothing in the record measures the pre-scoring process, so a 3 would assert an unmeasured counterfactual.

• baseline

  BASELINES. Reinforcing pathways mirror the PAN org's evidence-derived edge widths as intensity rungs; the two CHECK rungs are compressed one band from that same mapping and keep PAN's ordering, because no correction rate of any kind is published for this deployment (the appeal check at 2 with its inbound reach at 1; the peer-to-peer check at 1). Three widths are load-bearing. The scoring pathway runs at 3 because every request is scored before any person reads it, which both sides of the dispute state. The client configuration runs into the scorer at 2 - deliberately the same rung as the two inputs nobody disputes, not wider - because the tuning finding belongs to the investigation and the companies dispute it, and a widest-edge treatment would convert an attributed finding into the network's loudest assertion. The model self-loop runs at 3 on undisputed facts: one scoring layer and one criteria library in front of care for more than 100 competing insurers and about 100 million people.

• assumed

  THE ASSUMED SPLIT, DECLARED. Every width on the pathways out of the scorer - to the nurse tier, to the account team, and onward through the escalation to the physicians - inherits one quantity that is published nowhere: the share of requests approved outright versus sent to review. The vendor's claim that roughly two thirds of decisions render in real time and 90 percent of approvals complete within one business day is a company figure with no independent verification, and the investigation states plainly that the article does not quantify the split. These rungs therefore rest on the routing MECHANISM, which is common ground, and not on a measured share. The PAN org carries the same declaration on the same parameters.

• baseline

  THE TWO LATENT CHECKS ARE THE GOVERNANCE FINDING. Both sit on the governed parameter itself and both are drawn at 0 because the record documents them as not performed: no independent read of the scoring layer exists in the public record, and no channel reconciles a client's configuration against the determinations it produced. What the record does show is checking that lands elsewhere - a market conduct examination that read 196 FILES and made no finding about any threshold, a 2018 federal audit that reached this vendor only through one of its insurer clients, and state-published denial data that an investigative team, not a regulator, analysed. The 2018 audit is also the proof that the missing check is possible: an outside reader did adjudicate this vendor's denials against its criteria, once, at one client.

• baseline

  THE HUMANS ARE THE DENIAL PATH, NOT THE CHECK, and the diagram says so on purpose. In most networks in this catalogue the operator classes carry the correction. Here the two operator classes inside the vendor are the pathway a routed case travels toward a possible denial - only a physician can deny, so sending more cases to people is what produces more denials - and both inhibiting edges come from OUTSIDE the company: the treating clinician on a peer-to-peer call before the determination, and an appeal that runs through the client insurer afterwards. The appeal is the correction the record credits in this decision class and the one whose doorway Connecticut measured failing: 77 of 196 sampled letters omitted the notice that tells the member the external appeal exists. Per-item effect and reach are drawn as two different edges and are never averaged into one.

• baseline

  WHAT THIS NETWORK REFUSES TO CARRY. (1) No sentence anywhere in this org says the scorer denies care: both the investigation and the company state that it cannot, and the two-step mechanism - score, then route - is preserved on every edge. (2) The Connecticut consent order is drawn as what it is, an admitted procedural utilization-review matter with a $16,000 fine, and never as an algorithm finding; no regulator on this record examined the routing threshold. (3) The dial, the 15-percent denial-increase sales claims, the high-touch client requests, the 3-to-1 return pitch, the risk-contract savings capture, the Arkansas analysis, the Vermont figures, the sentinel-effect declines, the 15-cases-an-hour pace and the 2018 audit finding are the ProPublica / Capitol Forum investigation's findings, carried with the companies' published dispute attached. (4) The Arkansas figure of almost 20 percent is that investigation's ANALYSIS of state-published data, not a finding Arkansas published, and the roughly 7 percent Medicare Advantage comparator covers a different population and programme - an order-of-magnitude contrast, never a controlled comparison. (5) Ownership is in motion: on 30 April 2026 Cigna announced a strategic review of alternatives for eviCore, with no transaction underway per the company, so nothing here assumes stable ownership going forward.

• baseline

  WHAT ERROR MEANS HERE. On this channel an error is a routing output misaligned with what the clinical record supports - a supportable request sent into the denial-capable path, or an unsupportable one approved outright. It is never an accuracy figure about a person, and nothing measures it: no deployment error rate is published, and the denial-rate signals in the record (almost 20 percent in the Arkansas analysis since 2021, 6.1 to almost 15 percent in Vermont Medicaid's quarterly figures) sit downstream of human review and cannot be decomposed into a model error. The federal evaluation that orders the error term for this decision class examined Medicare Advantage insurers, not this vendor, and nothing from it is attributed here.

• assumed

  Served people are not in the dynamics. No authorization outcome, no medical-necessity judgement, no health outcome and no financial outcome for any person is computed from anything drawn here, and no score over a person is authored anywhere in this network. The denial rates on this record are properties of request streams, analysed by an investigative team from state-published data, and one state's 196-file sample - recorded, never derived. The record's disparities are between service lines, between states and between payers, not between subpopulations of served people, and this network does not convert a book-of-business rate into a subpopulation rate.

What this example does not show

• The adjustable routing threshold — the dial — and everything that hangs off it is the ProPublica and Capitol Forum investigation's finding, built on internal documents, corporate data and interviews including five former employees, and eviCore and Cigna dispute the characterization, stating the algorithms exist only to accelerate approval of appropriate care. Asserted here as undisputed: the scale and ownership, the scoring-and-routing mechanism with its physician-only denial rule, the Connecticut consent order, and the April 2026 strategic review. Nothing here says the algorithm denies care; both sides agree it cannot.
• The Connecticut consent order is small, procedural and admitted: a $16,000 fine over appeal-notice language, missed deadlines, peer-review assignment and file documentation, in a 196-file sample from calendar 2021. It says nothing about the algorithm, the threshold or any denial rate, and no regulator on this record has examined a routing threshold. Presenting it as algorithmic enforcement would overstate the only enforcement loop that ever closed.
• There is no litigation centered on this screening algorithm as of 28 August 2026. The malpractice action reported in the investigation dropped both the insurer and the vendor as defendants; a 2025 ERISA class action over behavioral-health medical-necessity criteria names a sibling unit of the same corporate family, not this vendor, and pleadings are allegations rather than findings.
• The almost 20 percent denial figure is the investigation's ANALYSIS of data Arkansas published about this vendor's book of business, not a finding Arkansas itself published, and the roughly 7 percent Medicare Advantage comparator covers a different population and programme — an order-of-magnitude contrast, never a controlled comparison. The Vermont range of 6.1 to almost 15 percent is one state's Medicaid quarterly reporting. Staffing counts, the real-time decision share and the one-business-day approval figure are company claims with no independent verification, and the share of requests approved outright is quantified nowhere at all.
• Served people are not modeled. Authorization outcomes, the care that follows them and the costs that follow that are boundary quantities recorded in the case file; the Lab models institutional propagation through the operator network and computes no outcome for any person. The record's disparities are between service lines, states and payers, never between subpopulations of served people, and this scenario does not manufacture one. Ownership is also in motion: the April 2026 strategic review contemplates partnership or combination, with no transaction underway per the company, so the parameter's owner may change hands.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

• EviCore by Evernorth (eviCore healthcare MSI, LLC), a Tennessee-domiciled utilization-review entity owned by The Cigna Group since 2018, performs delegated prior-authorization review for more than 100 client insurers — including UnitedHealthcare, Aetna, Blue Cross Blue Shield plans, Medicare and Medicaid contractors, and its own parent — covering about 100 million people, roughly one in three insured Americans. The mechanism is common ground between the company and its critics: an artificial-intelligence-backed algorithm scores each submitted request with a probability of approval, requests above the operating threshold are approved with no clinical review, requests below it route to in-house nurses and then to physicians, and only a physician may issue a denial. The algorithm approves or routes; it denies nothing. On October 23, 2024 ProPublica and The Capitol Forum (T. Christian Miller, Patrick Rucker, David Armstrong; co-published by CNN on November 7), working from internal documents, corporate data, and interviews including five former employees, reported that this routing threshold is adjustable and that insiders called it the dial: 'If EviCore wants more denials, it can send on for review anything that scores lower than a 95%,' one former employee said, and a former executive said of the review rate, 'We could control that. That's the game we would play.' The same investigation reports strictness sold as a product feature — a marketed return of three dollars of medical spend avoided per dollar of fees, sales staff touting denial increases of up to 15 percent, insurers including Aetna and Cigna requesting 'high touch' configurations that send more cases to review, and 'risk' contracts under which the vendor keeps or splits what it saves below a client's baseline spending target. EviCore and Cigna dispute this characterization, stating that the company 'uses the latest evidence-based medicine' and that its algorithms exist 'ONLY to accelerate approval of appropriate care and reduce the administrative burden on providers.' The dial and everything hanging off it are the investigation's findings, resting on internal documents and largely unnamed former employees, and are not adjudicated fact.

  empirical
  • Investigative Miller, T. C., Rucker, P., & Armstrong, D. (2024, October 23). 'Not Medically Necessary': Inside the Company Helping America's Biggest Health Insurers Deny Coverage for Care. ProPublica / The Capitol Forum https://www.propublica.org/article/evicore-health-insurance-denials-cigna-unitedhealthcare-aetna-prior-authorizations
  • Advocacy Coalition to Strengthen America's Healthcare (2024, November 4). ICYMI: Denials for Dollars: ProPublica Reveals How EviCore Helps Corporate Insurers Deny Care https://strengthenhealthcare.org/icymi-denials-for-dollars-propublica-reveals-how-evicore-helps-corporate-insurers-deny-care-to-increase-their-bottom-lines/

• The one enforcement action on this vendor's record is procedural, small, and admitted. Between September 19, 2023 and January 29, 2024 the Connecticut Insurance Department conducted a market conduct examination of 196 of eviCore's calendar-2021 Connecticut files. Its report and stipulation and consent order of February 5, 2024, Docket MC 24-15, records 77 adverse-determination letters that omitted the mandatory notice of the 120-day external appeal; urgent-care, appeal, and retrospective determinations issued past their statutory 48-or-72-hour, 30-day, and 60-day clocks; three files whose documentation could not support regulatory review; two denials not reviewed by an appropriate clinical peer; and one appeal decided by the same physician who made the initial denial. The violations are of Conn. Gen. Stat. sections 38a-591b and 38a-591d and Regulation 38a-591-8. eviCore admitted the allegations, paid a $16,000 fine, and undertook to file a corrective-action report within 90 days. These are utilization-review compliance findings and are stated here as fact because they were admitted. Their scope is load-bearing and is stated with them: the examination read FILES, it made no finding about the algorithm, the routing threshold, or any denial rate, and presenting the fine as algorithmic enforcement would overstate the only enforcement loop on this record that ever closed. No regulator on this record has examined a routing threshold. (ProPublica's shorthand of '77 violations found in a review of 196 files' compresses the order: 77 is the external-appeal-language category alone, and the total findings span seven categories.)

  empirical
  • Government State of Connecticut Insurance Department (2024, February 5). Market Conduct Report of eviCore healthcare MSI, LLC d/b/a eviCore healthcare, with Stipulation and Consent Order, Docket MC 24-15 https://portal.ct.gov/cid/-/media/cid/1_stipulation/mc-24-15-evicore-healthcare.pdf
  • Investigative Miller, T. C., Rucker, P., & Armstrong, D. (2024, October 23). 'Not Medically Necessary': Inside the Company Helping America's Biggest Health Insurers Deny Coverage for Care. ProPublica / The Capitol Forum https://www.propublica.org/article/evicore-health-insurance-denials-cigna-unitedhealthcare-aetna-prior-authorizations

• Three measured signals sit around the contested threshold, each with its own limit. First, ProPublica's ANALYSIS of data that Arkansas publishes about eviCore's book of business found prior-authorization requests turned down in full or in part almost 20 percent of the time since 2021, against roughly 7 percent for Medicare Advantage plans overall in 2022. The ~20 percent is the reporters' analysis and not a finding Arkansas published, and the ~7 percent comparator covers a different population and a different programme, so the pair is an order-of-magnitude contrast and never a controlled comparison. Second, Vermont Medicaid materials showed quarterly denial rates under this vendor's review ranging from 6.1 percent to almost 15 percent, and a 2019 Vermont presentation recorded advanced-radiology requests falling 16 percent, to 3,629, and cardiology requests falling 38 percent after review began — the demand suppression the industry markets as the 'sentinel effect', whose governance property is that a request never submitted appears in no denial statistic taken by anyone. Third, a former eviCore physician reviewer, maternal-fetal medicine specialist Gail Miller, told the investigation she was required to decide at least 15 cases an hour, one every four minutes, frequently outside her own specialty, and left after nine months. All three are the ProPublica/Capitol Forum investigation's reporting, disputed in characterization by eviCore and Cigna. None of them is a measurement of the algorithm: every denial rate here sits downstream of human review and cannot be decomposed into a model error, and no deployment error rate for this system is published anywhere. Neither is the share of requests approved outright rather than routed to review, which is quantified nowhere public at all. The HHS OIG evaluation of Medicare Advantage prior-authorization denials and the Senate Permanent Subcommittee on Investigations majority staff report of October 17, 2024 document insurer-side prior-authorization automation raising denial rates, but both examine INSURERS and not this vendor; they are domain context, and nothing in them is attributed to eviCore.

  empirical
  • Investigative Miller, T. C., Rucker, P., & Armstrong, D. (2024, October 23). 'Not Medically Necessary': Inside the Company Helping America's Biggest Health Insurers Deny Coverage for Care. ProPublica / The Capitol Forum https://www.propublica.org/article/evicore-health-insurance-denials-cigna-unitedhealthcare-aetna-prior-authorizations
  • Government evaluation U.S. Department of Health and Human Services, Office of Inspector General (2022, April 27). Some Medicare Advantage Organization Denials of Prior Authorization Requests Raise Concerns About Beneficiary Access to Medically Necessary Care (OEI-09-18-00260) https://oig.hhs.gov/oei/reports/OEI-09-18-00260.asp
  • Government U.S. Senate Permanent Subcommittee on Investigations (Committee on Homeland Security and Governmental Affairs), Majority Staff Report, Refusal of Recovery: How Medicare Advantage Insurers Have Denied Patients Access to Post-Acute Care, 17 October 2024 https://www.hsgac.senate.gov/wp-content/uploads/2024.10.17-PSI-Majority-Staff-Report-on-Medicare-Advantage.pdf

• The scoring layer and the human reviewers both work against clinical criteria that eviCore authors itself, and the record contains one dated instance of that arrangement failing. Per the investigation, citing the audit, a 2018 CMS audit found that outdated eviCore cancer guidelines led to inappropriate denials for 30 patients at the Blue Cross insurer HCSC; eviCore retrained staff in response. Two properties of that event matter more than its size. It reached this vendor only through one of its insurer CLIENTS rather than through any regulator of the vendor itself, and it is the single documented demonstration that an outside reader can adjudicate this vendor's determinations against its own criteria — the check that the public record otherwise shows nobody performing. No audit cadence for the criteria library is published, and no independent technical evaluation of the scoring layer exists in the public record. An adjacent settlement suggests the incentive belongs to the market structure rather than to one firm: Carelon, formerly AIM and the utilization-management arm of Elevance, paid $13 million in 2022 over wrongful-denial techniques, without admitting fault.

  empirical
  • Investigative Miller, T. C., Rucker, P., & Armstrong, D. (2024, October 23). 'Not Medically Necessary': Inside the Company Helping America's Biggest Health Insurers Deny Coverage for Care. ProPublica / The Capitol Forum https://www.propublica.org/article/evicore-health-insurance-denials-cigna-unitedhealthcare-aetna-prior-authorizations

• The vendor's published position is the other side of this record and is vendor-tier evidence throughout. An Evernorth article updated October 22, 2024 — the eve of the investigation's publication — states 'We aren't in the denial business, we're in the approval business', and reports more than 500 board-certified physicians across more than 60 specialties, more than 1,200 nurses and other clinical specialists, roughly two-thirds of medical decisions rendered in real time, and 90 percent of approvals completed within one business day, with peer-to-peer consultations framed as educational. Every one of those figures is a company claim with no independent verification, and the auto-approve versus routed-to-review split they imply is published nowhere. Two later company statements are operator-tier and are dated rather than characterized. On its Q1 2026 earnings call, held April 30, 2026, Cigna reported removing hundreds of tests, procedures, and services from prior authorization altogether and cutting medical prior-authorization volume by about 15 percent, citing the 2025 industrywide voluntary prior-authorization commitments; on the same call, incoming chief executive Brian Evanko announced a strategic review of alternatives for eviCore, contemplating partnership or combination with other industry participants, citing scalability, management attention, and industrywide standardization, with no transaction underway per the company. Any reading of this deployment should therefore avoid assuming stable Cigna ownership going forward: the owner of the adjustable parameter may change hands.

  empirical
  • Vendor Evernorth (The Cigna Group) (2024, October 22). How prior authorization brings value to patients and their health plans https://www.evernorth.com/articles/evicore-prior-authorization
  • Vendor The Cigna Group Q1 2026 earnings call (30 April 2026), transcript via The Motley Fool: strategic review of alternatives for eviCore announced https://www.fool.com/earnings/call-transcripts/2026/05/01/cigna-group-ci-q1-2026-earnings-call-transcript/

Where this connects

Institutional pressures in this domain

• Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
• Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
• Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
• Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
• Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.

All of them in context on the Clinical decision support & deterioration alerting domain page.

Levers available here and the patterns behind them

• Upgrade model — Improve the model
• Review the riskiest first — Risk-tiered oversight
• Mark AI-written records — Provenance labeling
• Gate record entries — Human-in-the-loop write gating
• Pause AI on alarms — Deployment circuit-breaker
• Escalate checks — State-feedback vigilance
• Assign a challenger — Structured dissent
• Understand the system — Understand the system
• Check with a second model — Cross-model verification
• Check copied records — Reconcile copied records
• Review on schedule — Oversight cadence & retrospectives
• Gate vendor updates — Vendor quality gate

Documented case histories

• EviCore by Evernorth: the review threshold
• TREWS sepsis early-warning system
• Advance Alert Monitor (AAM) deterioration model
• Sepsis Watch deep-learning detection system
• Proprietary EHR sepsis model (external validation)
• nH Predict Utilization Review
• Cost-Proxy Care Stratification
• CA-CDS Child Abuse Alerting
• IDx-DR Autonomous Screening
• Viz.ai LVO Stroke Triage
• IBM Watson for Oncology
• OPTN eGFR Waiting-Time Correction
• Practice Fusion Pain CDS
• UBH Level of Care Guidelines (Wit v. UBH)
• Cigna PxDx