A composite analysis of Recovr's production customer base — 3.38 million claims and 1.76 million denials processed over 12 months — reveals which denied dollars are truly recoverable and what AI agents do about them. Every dollar figure is stated at expected payer fees — what payers actually reimburse — not gross billed charges. This July 2026 update adds the live production record: 4,099 AI denial analyses, 2,396 reconsiderations filed through payer APIs, and 252 denials already reversed.
This case study draws on the collective experience of Recovr's production customer base: clinical laboratory and pathology networks processing roughly 3.4 million claims per year against a payer mix spanning Medicaid managed care plans, the Blues ecosystem, and national commercial carriers. The figures that follow are aggregated across all customers over a trailing 12-month window ending April 2026.
These are high-volume laboratory operations where denial rates are structurally elevated. Routine chemistry panels, surgical pathology, and molecular assays draw frequent coding, bundling, and medical-necessity denials that individually are small-dollar but in aggregate dwarf most physician practices' entire revenue. What the data reveals is not one lab's problem — it is an industry pattern.
This update pairs that 12-month book diagnosis with the live execution record from Recovr's flagship production deployment: every agent analysis, payer reconsideration, and corrected-claim submission logged between December 1, 2025 and July 20, 2026, read directly from the production system on July 22, 2026. Where the April edition projected what agents would do, this edition reports what they did.
Note: All dollar figures in this analysis are stated in expected-fee terms — billed charges converted at the customers' observed 20.7% dollar-weighted paid-to-billed ratio — because payers reimburse contracted fees, not gross charges. All labour costs are benchmarked at $10 per hour fully loaded — the floor rate for offshore RCM operations — to demonstrate that even in the most cost-optimised manual scenario, the economics of AI-driven recovery remain compelling.
Across Recovr's customer base, the same interrelated failures appeared before deployment:
1.76 million denied claims in 12 months — roughly 6,700 per business day — against finite billing staff. No scalable way to tell which denials are worth working.
Denials were addressed one claim at a time, without insight into the CARC code, payer, and CPT patterns generating 80%+ of the loss.
74.6% of denied claims were never touched again after the first 835 posted. Only 13.6% received more than one rework attempt.
Even at offshore rates, the 25.4% of denials that were reworked consumed $3.4M in staff time — and much of that was wasted on structurally unrecoverable claims that staff had no way to distinguish from real opportunities.
46% of denied dollars ($54M at expected fees) are structurally unrecoverable — categories where the insurer is simply not the payable party for the setting billed, and no amount of rework changes the outcome. Without automated line-level triage, staff spend hours working claims that were dead on arrival.
Recovr's first deliverable is not a chat interface — it is a diagnosis. By normalizing and correlating every 835 CAS segment against the originating 837, the platform surfaces the systemic loss patterns that no individual claim reviewer is in a position to see.
The largest truly recoverable CARC code. These are mechanically correctable: the claim was submitted without a required data element. Once identified and supplied, each is a refile candidate. This single denial pattern alone represents 13.2% of total denied dollars.
CO-29 (timely filing) denials run $1.4M per year at expected fees — claims rejected because payers say the filing window has closed. Conventional wisdom treats these as dead losses. Recovr's agents discovered otherwise — and have now tested that discovery in production at scale.
Recovr agents found that payers such as UnitedHealthcare were incorrectly assigning a 90-day timely-filing limit to Medicare Advantage plans where the contractual limit is actually 365 days. Acting on that discovery, agents filed 932 CO-29 reconsiderations through the Optum API between February and July 2026 — claims every conventional workflow had written off as un-appealable. Of the 463 decided by July 20, 48 (10.4%) came back overturned or revised, with 177 more awaiting a decision. Agents separately discovered 229 CO-29 claims that had in fact already been paid. This is exactly the kind of pattern that is invisible at the claim level but obvious at scale when an AI agent reads the fine print — and files the paperwork.
Denials are concentrated: the top 8 payers represent 62.8% of denied dollars. Four of them are Medicaid managed care organisations with well-documented, written resubmission and appeal policies. Recovr's payer-policy knowledge base indexes those policies on ingestion, giving the agent direct citations to use in every subsequent action.
The single most valuable thing Recovr does for a book this size is tell staff which denials to stop touching. In pathology and lab settings, nearly half of denied dollars are structurally unrecoverable — categories any seasoned RCM professional recognizes on sight, where the insurer is not the payable party and no rework changes the outcome. The hard part was never knowing the categories; it is enforcing that judgment claim by claim across 1.76 million denials a year. Recovr segments the entire denial pool automatically, separating the $55M that can actually be recovered from the $54M that cannot.
| Motion | Claims | Dollars | % |
|---|---|---|---|
| Correctable / Refile | ~452,000 | $25.1M | 21.5% |
| Appealable | ~211,000 | $14.0M | 12.0% |
| Reroute / COB / Discovery | 165,347 | $14.5M | 12.4% |
| TFL Appeal (CO-29) | ~18,000 | $1.4M | 1.2% |
| Under Review | 185,628 | $8.5M | 7.2% |
| Terminal — structurally unrecoverable | ~931,000 | $53.6M | 45.8% |
After fencing off the structurally dead categories — and reclassifying CO-29 timely-filing claims that agents proved were wrongly denied — $55.0M across ~846,000 claims sit in a recoverable motion. 46% of the book is dead on arrival. Knowing which half to work was never the hard part; enforcing that judgment on every remit, at line level, is. Live production bears the segmentation out: of the 4,099 denials agents have analyzed end-to-end, 71% were classified recoverable, and 96% of recommendations carry a confidence score of 0.80 or higher (mean 0.91).
Recovr's specialised AI agents execute the recovery actions themselves, reducing the human role from operator to reviewer. For each recoverable denial, the agent performs root-cause analysis, selects the recovery method, populates field-level corrections, drafts appeal letters with payer-policy citations, submits via API or portal, uploads documentation, and monitors resolution. Since December 2025 that model has been running in production — the results below are counts read from the live system, not projections.
Between December 1, 2025 and July 20, 2026, Recovr agents analyzed 4,099 denials end-to-end, classified 71% as recoverable, and executed the recovery motions themselves: 2,396 reconsiderations filed programmatically through the Optum API, 123 corrected claims built through the clearinghouse pipeline — 76 live submissions, 98.7% accepted — and 21,136 automated payer interactions — claim-status checks, eligibility verifications, coverage discovery, and coordination-of-benefits evaluations — that no human had to perform.
The blended 20.4% production win rate carries a deliberate skew: agents take shots no human team would staff. Nearly half of all reconsiderations filed are CO-29 timely-filing appeals — a high-volume, long-odds motion on claims the industry writes off at 100% loss, where every win is found money. On motions with documentation attached, the rates look very different: 51.3% of medical-records reconsiderations and 39.4% of corrected-claim reconsiderations decided to date were overturned or revised in the provider's favor. And because filing costs the agent minutes rather than the industry-benchmark 45 of staff time, expected value stays positive even on the long-odds queue.
Applying conservative industry win rates per recovery motion to the full recoverable pool:
| Motion | Recoverable $ | Win Rate | Expected Recovery |
|---|---|---|---|
| Correctable / Refile | $25.1M | 55–65% | $13.8M – $16.3M |
| Appealable | $14.0M | 40–55% | $5.6M – $7.7M |
| Reroute / COB / Discovery | $14.5M | 30–45% | $4.4M – $6.5M |
| TFL Appeal (CO-29) | $1.4M | 10.4% observed | $0.1M |
| Total Expected Recovery | $55.0M | 43–56% | $24M – $31M |
| Pillar | Outcome |
|---|---|
| 1. Insights | Fenced off $52.0M in structurally unrecoverable denials automatically at line level; identified $15.5M in missing-info correctable claims and $14.5M in payer-routing errors; and proved in production that "dead" CO-29 timely-filing denials are appealable at scale — 932 reconsiderations filed on claims payers had stamped expired, 48 reversed to date (10.4% of decided), 177 still in flight. |
| 2. Recoverability | Segmented 1.76M denials into a funnel where 47.0% of denied dollars ($55.0M) are truly recoverable and 45.8% ($53.6M) are terminal — each tagged with method, confidence score, and cited sources so staff never waste time on dead claims. Live production confirms the discipline: 71% of 4,099 agent-analyzed denials classified recoverable, 96% of recommendations at ≥0.80 confidence, and 416 "denials" exposed as already paid. |
| 3. Agent Execution | In production Dec 2025–Jul 2026: 2,396 reconsiderations filed programmatically via the Optum API and 76 live corrected claims submitted via clearinghouse (98.7% accepted); 252 denials reversed — 20.4% of decided, rising to 51.3% on medical-records motions; 21,136 automated payer interactions; 98.9% operator approval across reviewed recommendations, with the human role reduced to a 2-minute review. At book scale: $2.8M annual labor avoided even at $10/hr offshore rates; $24M–$31M annual expected-fee revenue potential from truly recoverable denials. |
Data source — book analysis. All volume, dollar, and CARC figures are computed directly from customers' HIPAA EDI 835 remittance and 837 claim records loaded into Recovr's data warehouse. Adjustments exclude group code PR (patient responsibility) and reversed service-line records.
Expected-fee normalization. Payers reimburse contracted fees, not gross charges, so every dollar figure in this study is stated as expected payer reimbursement. Billed charges are converted at the customers' observed dollar-weighted paid-to-billed ratio of 20.7%, measured from 3,182 adjudicated-and-paid claims returned by live 277 claim-status checks through July 20, 2026 ($344K paid on $1.67M billed; median per-claim ratio 21.7%). At billed charges, the underlying denied book is $567.2M on $1.55B submitted; at expected fees it is $117.1M. Labor figures are actual costs and are not scaled; claim counts are unaffected.
Data source — production results. All production execution counts (agent analyses, reconsideration tickets, corrected-claim submissions, eligibility/status/discovery/COB volumes) are read directly from Recovr's live operational data store via its MCP interface, pulled July 22, 2026. Win rate is computed per operational convention as (Decision Overturned + Revised) ÷ (Overturned + Revised + Upheld); in-progress tickets, tickets closed without a decision label, and submission failures are excluded from the denominator. Reconsideration figures cover the UnitedHealthcare/Optum programmatic channel, currently the highest-volume API motion in production. Operator approval rate = approved ÷ (approved + rejected) among human-reviewed recommendations.
Windows. Book analysis: trailing 12 months (April 2025 – April 2026). Production execution: agent analyses December 1, 2025 – July 20, 2026; Optum reconsiderations February 4 – July 20, 2026; clearinghouse corrected claims May 25 – July 20, 2026. July figures run through the 20th.
Rework signal. Measured as the number of distinct 835 remit IDs posted against each denied claim's ACN. More than one remit is used as a proxy for a rework attempt.
Labor benchmarks. 45-minute manual rework time per denial (15 min research + 20 min drafting + 10 min submission). All labour costs in this analysis use $10/hr fully loaded — the floor rate for offshore (India) RCM operations — yielding $7.50 per claim reworked. For reference, the HFMA 2022 US benchmark is $43.84 per claim at $45/hr fully loaded.
Recovery win-rate benchmarks. Refile 55–65% (AAPC); appeal 40–55% with clinical justification (HFMA); reroute/COB 30–45% (industry composite). Observed production rates at Optum through July 20, 2026: medical-records reconsiderations 51.3% (142/277 decided), corrected-claim 39.4% (26/66), patient eligibility 20.0% (3/15), coordination of benefits 18.4% (9/49), timely filing 10.4% (48/463), reimbursement rate 6.8% (8/118), policy 6.1% (13/213); blended 20.4% (252/1,237).
Structural exclusions. $52.0M in denied expected-fee dollars sits in categories that experienced laboratory RCM teams treat as terminal on sight — services billed in settings where the insurer is not the responsible payer. These well-understood categories are excluded from the recoverable pool on ingestion; the platform's contribution is enforcing that exclusion automatically at line level across 1.76M denials, not rediscovering it. Conversely, CO-29 timely-filing denials ($1.4M), traditionally classified as terminal, are reclassified as recoverable after Recovr agents demonstrated in production that payers incorrectly apply 90-day TFL limits to Medicare Advantage plans with 365-day contractual windows — 48 of 463 decided CO-29 reconsiderations (10.4%) overturned or revised as of July 20, 2026, with 177 still pending. Net adjusted recoverable pool: $55.0M at expected fees.
Conservatism. All headline numbers use the low end of benchmark ranges. Upper-bound figures are presented as ranges, never as single point estimates.