Underpayment research in revenue cycle software answers one operational question: was the payer obligated to send more money than the 835 remittance shows? The answer requires a machine-readable expected payment for every claim line, built from contracts, fee schedules, and coding rules. Then software compares that model to actual reimbursement and routes unexplained variance to an appeal or recovery queue. Organizations that only audit denied claims miss a larger population: claims recorded as paid but short of contracted rates. Source evidence indicates such discrepancies can persist unnoticed for months without a systematic baseline. This guide explains that workflow, the technical decision points that affect accuracy and recoverability, and where current documentation stops short of proof.

The Baseline Problem: Why Underpayments Stay Invisible

Denials announce themselves. A claim is rejected, a report triggers, and staff open an appeal. Underpayments behave differently. AAPC webinar commentary describes them as easy to miss unless software is checking daily against an expected payment amount, because the remittance still arrives and the account often closes with a zero balance [5]. Clear Health makes the same point in its product materials: without a model of what should be paid, an organization cannot know what is missing [4].

The severity varies by setting. Aspirion's white paper cites Medicare and Medicaid underpayments reaching $130 billion in a single year, with additional sums sitting in zero-balance accounts most health systems lack resources to revisit [3]. The limitation is that this figure comes from an advocacy-aligned vendor document rather than an independent audit. McKinsey survey data adds a structural observation: a majority of RCM leaders say payer-related factors are contributing to aging accounts receivable, though that does not isolate underpayments from broader payment delay [8].

Contract Digitization as the Detection Foundation

The first technical decision is how a platform turns payer agreements into structured rules. R1 describes a four-stage contract intelligence process: parsing PDF contracts into a searchable repository, extracting contract terms and remittance rate data with AI, validating through human review, and then applying those terms to identify payment variance [1]. That human-in-the-loop step matters because a single misinterpreted percentage-of-Medicare clause propagates false underpayment flags across every affected line.

OutcomeCatalyst's documentation points to rate volatility as a recurring source of error. The vendor notes that CMS set two 2026 physician fee schedule conversion factors, $33.57 for qualifying APM participants and $33.40 for everyone else, and applied a -2.5% efficiency adjustment to work RVUs for many non-time-based services [2]. A contract model using last year's fee schedule table shorts every affected line. Aspirion's white paper similarly emphasizes that contracts contain carve-outs and payment methodologies too complex for manual or spreadsheet review [3]. The evaluator should ask how often fee schedules refresh, not merely whether parsing succeeds on a sample PDF.

A Worked Example: Seasonal Contract Refresh Shortfall

Suppose a physician group's commercial contract pays 100% of the CMS physician fee schedule, and the payer fails to update its pricing table when the 2026 conversion factor takes effect. If the group uses an underpayment tool that models expected reimbursement from the payer's own last-loaded table rather than the current CMS file, the tool reproduces the payer's error and finds nothing [2]. If the tool re-models from the CMS source directly and compares it against the 835, it flags every non-APM line paid at the prior year's schedule rate instead of $33.40, provided the contract specifies which conversion factor applies [2].

This example illustrates two decision points: the platform's source of truth for fee schedules and its treatment of contract amendments. A tool that only compares claim to claim within one payer's history will miss a systemic underpayment affecting every line. A tool that refreshes CMS schedules centrally and maps payer-specific percentage terms can identify the whole class at once. The same logic extends to ASC lines affected by a 2.6% rate update or new covered procedures [2]. Evaluators should test whether the vendor's fee-schedule refresh process is automatic, and what timestamp evidence they can produce for a chosen lookback period.

Where the Recovery Work Actually Happens

Detection architecture splits into two patterns. The first is line-level variance detection at remittance arrival. Clear Health describes modeling expected reimbursement from digitized contracts, matching each ERA line as it arrives, flagging discrepancies with root-cause analysis, and routing them to a prioritized recovery queue [4]. The vendor claims underpayments are identified within 48 hours at ERA arrival and cites a 90-day example of $847,000 identified and $612,000 recovered at one regional health system, though these figures are vendor-attributed and unaudited [4].

The second pattern targets zero-balance accounts, where the claim has closed but payment fell short. Aspirion's white paper states that AI contract modeling analyzes zero-balance accounts at scale, comparing them against extracted contract terms to identify entire classes of underpaid claims, and reports that 79% of recoveries exceed $1,000 while the most common range is $1,001 to $2,500 [3]. Revecore positions itself in the same space, claiming recovery performance two to three times higher than variance-only approaches and cycles as low as 30 days for high-probability commercial claims, with the caveat that these are promotional performance statements [6].

Evidence Limits and the ROI Gap

The evidence base for underpayment software is heavily vendor-authored. R1 reports over $850 million recovered annually and a 90-day timeframe to positive ROI [1]. Revecore reports a 420% ROI on underpayment recovery [6]. Aspirion describes an AI-driven 27% success rate for DRG downgrades and claims lower-dollar claims become viable when processed at scale [3]. These numbers are directional marketing figures, not independently verified benchmarks, and they are not comparable across vendors because definitions of recovered revenue and baseline periods differ.

HFMA survey data shows why adoption remains uneven. Among 272 provider respondents, 57.1% cited cost and 54.9% cited IT resources as barriers to adopting AI-driven revenue cycle solutions [7]. Only 20.9% rated their organization's real-time analytics as effective for revenue cycle decision-making, with a further 57.8% saying somewhat effective [7]. McKinsey's 2025 survey adds that leaders have more measured expectations for near-term AI returns, while 60% plan to change vendor strategy within two years [8]. The practical implication is that underpayment research tools compete for the same internal analytics budget as denial management and coding automation, often against entrenched manual workflows.

A Suggested Evaluation Workflow for Underpayment Research

As an editorial suggestion, an organization could structure a pilot around one payer contract family with known fee-schedule complexity. The first step is extracting twenty contract clauses covering carve-outs, stop-loss provisions, and percentage-of-Medicare pricing, then checking whether the platform flags the same anomalies a senior analyst identifies manually. The second step is loading six months of historical 835 data and measuring false-positive rate on line-level flags. The third step is comparing zero-balance account findings against a manual re-audit of the highest-dollar closed claims.

This structure reflects the documented workflow rather than a generic feature checklist. OutcomeCatalyst notes that if underpayments are reviewed once a year or only by an outside firm after the fact, the gap is people and data, not features [2]. AAPC webinar commentary similarly cautions that underpayment visibility requires software that is constantly updated, and advises due diligence when vetting technology companies [5]. The evaluation should name the unit of measurement before any software is selected: recovered dollars net of pursuit cost, detection latency, or false-positive burden per analyst hour.