Crushing Fraud Starts With Getting the Findings Right

CMS is asking how its next generation of fraud defense contractors should operate. Its draft Statement of Objectives proposes a greater CMS role in developing leads, setting priorities, and directing investigations, with an emphasis on defensible actions, coordination, and measurable outcomes. Arclight welcomes that direction. The Fraud Defense Contracts request for information offers an opportunity to improve how Medicare identifies fraud and how it evaluates the people doing that work.

In Arclight's Medicare audit and appeal work, we have encountered fragmented claims, changing denial rationales, and reviews that repeat earlier conclusions without meaningfully addressing clinical evidence. When an unsupported finding moves downstream, the next reviewer, the provider, and sometimes the patient inherit its consequences. Each additional level of review can become another place to repair a problem that should have been caught earlier.

Our paper, Upstream Denials, Downstream Costs, examines how those burdens can escape conventional measures of program integrity performance. That experience informs the five reforms I would prioritize, beginning with accountability for the accuracy of the initial finding.

1. Make accuracy consequential from the first finding

Every consequential finding should remain linked to what happens afterward: administrative action, appeal, enforcement, correction, and final disposition. CMS should be able to determine whether the original finding was supported, whether its rationale survived independent review, and what it cost to resolve.

A final decision upholding a denial does not necessarily validate every reason the originating contractor gave. If one rationale fails and a different rationale sustains the result, the performance record should show both. Otherwise, a contractor can appear accurate while its initial reasoning was defective and the provider had to respond to a moving target.

We recommend an issue-by-issue record showing the original finding, supporting evidence, assessed amount, and disposition: sustained, rejected, not reached, or replaced. Reversal reasons should distinguish initial error from new evidence or changed policy. Contractors should be compared using similar case populations, rather than a raw overturn rate that treats every reversal as the same event.

CMS already recognizes the value of this feedback. Its Program Integrity Manual directs Medicare Administrative Contractors to examine denials routinely reversed on appeal and consider changes to policy, procedures, outreach, or review strategies. The next contract model should make that principle measurable and consequential across the investigative process. CMS Program Integrity Manual, Chapter 3, Section 3.7.4.

Because appeals take time and cover only contested cases, CMS also needs prompt, independent checks of initial work, including unappealed actions and leads closed without investigation. Reviewers should ask whether the record was read, the applicable policy was understood, contrary evidence was addressed, and the conclusion was supported. Recurring errors should lead to verified corrections, retraining, and contractual consequences.

2. Pay for quality and verify the savings

Performance standards matter more when they affect compensation. A contractor should benefit financially from producing accurate, timely, defensible work, including a well-supported decision to close an allegation that does not warrant further action.

Fixed pricing can encourage efficiency, but the payment unit still matters. Paying per review can reward additional reviews. Paying per referral can reward additional referrals. Neither establishes that the work was necessary or useful.

Arclight recommends sustainable funding for required capacity, fixed prices for bounded work, and a meaningful quality-linked payment component. Initial payments should depend on independent acceptance of the work. A downstream component should be reconciled as appeal and enforcement outcomes mature, distinguishing contractor errors from decisions or delays outside its control. Incentives can encourage internal quality assurance, but independent verification remains necessary.

Savings also need careful definitions. GAO's 2026 review of Medicare fraud analytics distinguishes overpayment recoveries from projected cost avoidance associated with suspensions and revocations. These are different measures, and contracts should preserve that distinction. GAO's Medicare fraud analytics report.

Temporary payment holds and projected recoveries should not be presented as realized savings. Any savings bonus should require independently verified attribution, account for refunds, and deduct attributable investigation, appeal, and remediation costs. Estimates of prevented future losses can still inform decisions, but they should be reported separately with transparent assumptions.

3. Connect the case across every handoff

Centralization could improve oversight substantially if it gives CMS a continuous view of each matter. Changing who receives an assignment will accomplish less if claims, records, and decisions still become disconnected afterward.

CMS already has Unified Case Management infrastructure supporting investigations and coordination. Our recommendation is to connect that foundation with claims, appeal records, and corrections so a matter remains traceable throughout its life. CMS's description of Unified Case Management.

Each claim should remain linked to its parent case, evidence, policy issues, review history, and individual outcome, even when claims are split across appeal dockets or transferred between organizations. Every handoff should identify a responsible owner, an acknowledgment of receipt, missing materials, and unresolved decisions. CMS should commit to timely data access and approvals alongside the deadlines it imposes on contractors.

A secure provider portal would make this improvement tangible. Providers and authorized representatives should have one place to submit reusable records, receive filing receipts, and see permitted status information across review levels. Sharing records must preserve independent adjudication and protect restricted investigative information. A shared system should also distribute corrections, so an overturned finding does not continue circulating as an established fact.

4. Get the clinical review right before escalating

A medical review needs the complete relevant clinical record and someone qualified to understand it. A narrow claim sample should not produce a fragmented view of the patient's treatment.

In wound care, for example, the significance of a treatment may depend on prior conservative care, serial measurements, treatment changes, and the patient's response over time. An isolated note may not answer the question the reviewer is trying to resolve. Reviewers also need the coverage policy applicable to the jurisdiction and dates of service.

CMS should require specialty expertise where the clinical question warrants it, individualized findings with specific evidence, and consideration of contrary facts. Missing information or ambiguous policy should prompt clarification where appropriate. Documentation deficiencies, coverage disagreements, and substantiated fraud require different responses. Education or correction may resolve a compliance problem without sending it through an expensive appeals process.

Targeted review should begin with a defined question and expand when evidence warrants expansion. Prepayment controls need reassessment and exit criteria. Successful appeals should trigger correction of associated risk flags and controls unless separate evidence supports them.

Beneficiary access belongs in this assessment. When a suspension or revocation threatens ongoing care, CMS should coordinate outreach, clinically appropriate alternatives, and help transferring records. These safeguards should accompany timely action on credible fraud and urgent patient risks.

5. Put a human in the lead of AI/SI

Advanced AI/SI offers opportunities to find suspicious networks, assemble longitudinal records, detect duplicate work, and prepare evidence for qualified review. Persistent agents could update leads as information arrives and flag inconsistencies before a weak finding travels downstream. Future superintelligence capabilities should be evaluated as they become available and reliable.

These systems need explicit objectives for accurate detection, fair review, and protection of legitimate care. Historical denial labels should not automatically become examples of confirmed fraud. Training and evaluation should incorporate independently reviewed outcomes, legitimate outliers, false positives, and missed risks. This is the design problem explored in Arclight's Design Before Deploy.

An independent challenge process should examine alternative explanations, verify policy and evidence, and preserve disagreements. A second agent may help with that work, but agreement between models cannot establish that a conclusion is correct.

A human in the lead must have the expertise, time, and authority to reject, revise, or stop a recommendation before consequential action. CMS should require testing of the complete workflow before deployment, monitor performance afterward, and define when a system must be paused or corrected.

CMS also needs control of case histories and sufficient rights to retain and transfer contract-developed workflows, documentation, and evaluation materials. Replacing a vendor should not mean losing institutional knowledge. Portability should be demonstrated through transfer and restoration tests, with replacement models validated again.

CMS's proposed direction gives these reforms a practical opening. Arclight welcomes greater coordination and a clearer focus on outcomes. The strongest version of that reform would follow each finding through its consequences, pay for independently verified quality, and use what the system learns to improve the next investigation. That would help concentrate fraud defense resources where the evidence supports action while reducing avoidable costs for providers, beneficiaries, and government.