The Denial That Never Arrives
Top 7 practices for handling overdue claims with AI agents
A Recovr case study · Pinnacle Services Corporation · September 2026
Every billing team has a queue for denials. Almost none has a queue for the claim that was never denied — the one a payer's front end rejected before adjudication, that produced no remittance, that still shows as "sent" in the practice-management system, and that quietly ages toward its timely-filing limit. Between late July and mid-September 2026, an AI agent pipeline running across a national collection of more than 12,000 physicians in some sixty specialties screened more than 22,000 silent claims, resent 1,401 of them, and captured the payers' verdicts within a day. What it found changes how the problem should be framed: the largest single cohort was a set of 214 Humana claims worth $456,245, rejected in their entirety because one laboratory line on each — typically a tenth of the claim's value — lacked a "detailed description of service." This is the story of what the agents did, what the humans still had to do, and the seven practices that fell out of the data.
Snapshot
Metric | Value |
|---|---|
Data set | National collection of more than 12,000 physicians, 60+ specialties, ~245,000 claims a month, ~700 payers |
Observation window | July 26 – September 16, 2026 (53 nightly runs) |
Silent claims with a status record by Sept 16 | ~22,200 (17,784 open, 4,460 closed) |
Claims resent unchanged by the agent | 1,401 (1,343 accepted by the clearinghouse) |
Median claim age at resend | 174 days from date of service; none under 94 days |
Payer front-end rejections on resent claims | 615 (55% of tracked resends), median verdict latency 23 hours |
Largest rejection cohort | Humana, 214 claims, $456,245, one NOS lab line per claim |
Charges moved from "silent" to "actionable" | >$920,000 (records carrying an amount) |
Cases handed to a human with a specific ask | 315 held cases, 159 approval cards, 128 open enrollment tasks |
Timely-filing denials in the book, trailing 12 months | 16,529 claims, $1.08M (CARC 29), run-rate up 5x year on year |
The problem: a rejection is not a denial
An 837 claim can fail in three places. It can be refused by the clearinghouse for syntax. It can be rejected by the payer's front-end edits and returned on a 277CA acknowledgment with a category code such as A3 ("returned as unprocessable — has not been entered into the adjudication system"), A6 ("rejected for missing information") or A7 ("rejected for invalid information"). Or it can be adjudicated and denied, which is the only one of the three that produces an 835 remittance with a CARC code.
Denial management tooling, staffing and metrics are built around the third case. The first two produce no remit, no CARC, no worklist entry. If the acknowledgment feed is not read line by line — and in a high-volume shop it rarely is — the claim simply sits. The payer's timely-filing window, which ranges from 90 days for a participating Cigna provider to 12 months at Blue Shield of California, keeps running. When the claim is finally noticed and resubmitted, the answer is often CARC 29: the time limit for filing has expired.
The Humana cohort shows how little it takes. Humana's front end requires a line-level description for "not otherwise specified" laboratory codes such as CPT 87798 (infectious-agent detection by amplified probe, NOS) and CPT 84311 (spectrophotometry, analyte not elsewhere specified). When the description is missing, Humana does not reject the line; it returns claim-level A3/21 ("missing or invalid information") and rejects the whole claim. In this book, 214 claims carrying $456,245 in charges were rejected that way. The flagged lines were worth $67,102 — 14.7% of the total, a median of 10.6% per claim. Of the thirty largest claims, the median carried 49 service lines, the maximum an 837P allows is 50, and a single 87798 line was enough to void all of them. One claim of $5,747.75 with 50 lines failed on one line billed at a few hundred dollars. None of these claims ever produced a denial to work.
The data and the setup
The data are drawn from Recovr's consolidated physician book: more than 12,000 physicians across some sixty specialties, billing about 700 payers nationwide. The pipeline described here runs nightly against every claim in that book with no remittance on file.
Practice 1 — Treat "sent" as a hypothesis, not a status
The first job is to stop trusting the practice-management system's word for it. Each night the pipeline takes every claim with no 835 on file, drops the ones another workflow owns, and asks the payer directly with a 276/277 claim-status inquiry — not on a fixed calendar, but once the claim is older than that payer's own 99th-percentile response time plus a 60-day buffer, so that a slow payer is not pestered and a fast one is not given the benefit of the doubt for months.
Payer's answer on the status channel | Claims | Share |
|---|---|---|
Finalized at zero payment, no remittance ever posted | 11,587 | 65% |
Status channel returned an error (E-category) | 2,221 | 12% |
Claim not on file (A4/D0) | 1,371 | 8% |
Rejected pre-adjudication | 645 | 4% |
Processing, pending, resent, enrollment, other | 1,960 | 11% |
Two-thirds of the "silent" claims were not silent at the payer at all: they had been adjudicated to a zero payment and the remittance had never reached the billing system's denial queue. That is a missing-remit problem, and it is worth surfacing because it is invisible to a denial dashboard. The 8% the payer could not find and the 4% it had already rejected are the population the rest of this case study is about.
Takeaway. A status inquiry costs a provider about $3.64 electronically against $13.80 by phone, and a phone inquiry takes 25 minutes on average, according to the 2024 CAQH Index. At those prices, sweeping every silent claim is cheap; not sweeping them is how a claim reaches its filing deadline unexamined.
Practice 2 — Read the status channel with error bars
A 277 response is evidence, not truth, and its categories mean different things. A4 or D0 across every entry means the payer has no record of the claim. An E-category response means the payer could not process the inquiry — usually because the inquiry itself was malformed. The two cannot be handled the same way, and a mixed response (one entry "not found," another "paid") must never be treated as missing.
Blue Shield of California made the distinction expensive. Its status endpoint requires the total claim charge on every inquiry and answered 397 of the eventually rejected claims with E0/178 ("submitted charges") — an error about the question, not the claim. Rather than treat those as unknowns, the pipeline used the claim itself as the probe: it resubmitted the unchanged claim specifically "to obtain the payer's full front-end response," and 529 of the 1,401 resends were logged with that rationale. The payer's 277CA on the resend became the definitive answer the 276 could not give.
The probe cut both ways. Among the claims resent after a "not found" or an error answer, the 835s that followed included CARC 18 duplicates — proof that the payer had the original all along and had simply never returned a remittance for it. That is not a failure of the method; it converted an ambiguous status into a documented fact and a duplicate denial that can be traced back to the original adjudication.
Takeaway. Classify every status response into "not on file," "inquiry error," "mixed," and "final" before deciding anything, and treat a controlled resend as a diagnostic instrument when the status channel is blind.
Practice 3 — Resend unchanged, but only through safety gates
Resending a claim is the cheapest recovery action there is, and the most dangerous to automate carelessly. The pipeline permits an unattended resend only for a claim whose 277 is declaratively "not on file," and only after three gates pass: no claim-edit history exists for the claim, so the agent cannot stack a second correction on an operator's first; billing history shows no 835 and no newer 837 to a different payer, failing closed if the history cannot be read; and eligibility on the date of service is not inactive. That third gate flagged 98 claims as "not covered" for a human to reroute rather than resending them into a certain denial.
A durable attempt marker is written before the live submission, and any unconfirmed attempt is held for 48 hours — past the clearinghouse's 24-hour idempotency window — so a network failure cannot produce a duplicate. Deterministic failures retire the claim instead of retrying it: 58 resends were retired that way, 29 because a MassHealth payer ID was not routable through the clearinghouse network, the rest for structural faults such as a missing DTP-03 on an L.A. Care coordination-of-benefits loop or a missing CLIA number on a Cigna laboratory claim. Each of those is now a rule or a credentialing task, not a nightly retry.
The outcome of the gated design: 1,401 claims resent across some 30 payers, 1,343 accepted by the clearinghouse on the first pass, a 4% mechanical failure rate, and, by construction, no second attempt inside the idempotency window.
Takeaway. Autonomy for an unchanged resend is defensible precisely because the claim's content is not being altered; the safety work belongs in the gates that decide whether to send at all.
Practice 4 — Listen for the 277CA within a day, and taxonomize it by payer
The resend is not the result; the acknowledgment is. A dedicated listener polls the clearinghouse's 277CA feed, checks for a matching 835, and, for payers with no acknowledgment route, falls back to a real-time 276 after thirty days. Verdicts arrived fast. The median 277CA rejection landed 23 hours after the resend (interquartile range 22 to 29 hours, maximum 72). Where a claim went on to adjudicate, the 835 arrived in a median of seven days.
Of the 1,123 tracked resends, 615 came back rejected on the 277CA, 121 were adjudicated and denied on an 835, 2 were paid outright and 383 were still in flight on September 16. The rejections sorted cleanly by payer and code:
277CA code | Meaning | Runs | Dominant payer |
|---|---|---|---|
A3/21 | Missing or invalid information | 266 | Humana (205), L.A. Care (30) |
A7/18 | Entity received claim but returned invalid status | 128 | Blue Shield of California (101) |
A3/54 | Duplicate of a previously processed claim/line | 107 | Blue Shield of California (107) |
A7/88 | Not eligible for benefits for dates of service | 17 | Cigna |
A3/26 | Entity (patient) not found | 15 | Blue Shield of California, Aetna Better Health |
A7/97, A3/33, A3/32 | Subscriber or eligibility not found | 17 | Cigna, AARP, UMR |
A3/562, A7/562 | Provider NPI invalid | 11 | Arizona Blue, Humana, Simply |
The value of the table is that each row is a different piece of work. Humana's A3/21 is a line-note omission with a deterministic fix. Blue Shield's A3/54 is a front-end duplicate edit that, on inspection, was not always a true duplicate: in one case the agent found professional and technical component claims — modifier 26 and TC on the same pathology codes, same date of service — flagged as duplicates of each other, and in most others it could find no matching claim anywhere in the patient's billing history. L.A. Care's A3/21 arrived with the free-text reason "R1 — unrecognized segment ID," which names no segment and can only be resolved by asking the payer's EDI desk. Cigna's A7/88 is a genuine eligibility question. None of these should be routed to the same queue, let alone the appeals queue.
Takeaway. A 277CA rejection is a denial with a 24-hour turnaround and no CARC. Capture it, decode it per payer, and route it by root cause; the median claim in this cohort had already waited 174 days for anyone to ask.
Practice 5 — Fix at the line, resubmit the claim, then codify the fix
When the rejection-fix agent opened the Humana cohort, the root cause was the same on every claim: a service line for CPT 87798 or 84311 (151 and 99 flagged lines respectively, with a handful of 87899) carried no line-level note. The fix is a single NTE segment on the affected line — NTE*ADD*DETECT AGENT NOS DNA AMP — with every other field untouched, resubmitted as an original claim. The agent validated the edit against the clearinghouse in test mode, then released the corrected claim under the billing team's approval policy: 51 claims were resubmitted and accepted by the clearinghouse that way.
The other 163 exposed the limit of a generic fix. Their lines already carried the generic descriptor, and Humana's edit wanted the specific organism or analyte tested — information that lives in the laboratory's order, not in the 837. The agent did not guess. It held each claim and issued a follow-up to the operator naming the CPT, the line and exactly what was needed ("Provide organism/analyte detail for CPT 87798 line"). On July 28, two days after the first wave of rejections, the pattern was written into an autorun rule: any 87798 or 84311 line missing a note gets the default text; lines that already carry a note are left alone; anything outside A6/306 or A3/21 is held for a human rather than improvised.
The economics of the pattern are what make it a practice rather than an anecdote. For 124 of the 214 claims, the flagged line was under 15% of the claim's value, and those claims carried $252,714 in charges. A ten-dollar omission held five-thousand-dollar claims hostage, and the information that would have released them was never requested by anyone until an agent asked for it by name.
Takeaway. Edit the offending line, never the claim; validate before you send; and write the fix down as a rule the same week, because the payer's edit will fire again next month.
Practice 6 — Keep a human at the send button, and make the human's part measurable
Humana's specific-organism rejections, Blue Shield's false-positive duplicates and Cigna's eligibility questions have one thing in common: the right answer is not in the claim data. An agent that guesses at a member ID, deletes a modifier to get paid, or resubmits a duplicate to "see what happens" creates compliance exposure faster than it creates revenue. The pipeline therefore runs on a hard rule: nothing modified is sent without a human's approval, and an agent that cannot find evidence for an edit holds and asks.
The record shows the restraint working. Of 577 rejection-fix runs, 159 ended as approval cards awaiting an operator's click, 315 were held with a specific operator follow-up, 80 produced a validated correction that was resubmitted, and in 32 the agent concluded the claim was not recoverable and documented why. In one Blue Shield duplicate case the agent compared the rejected claim line by line with its sibling, established that one carried modifier 26 and the other TC, refused to alter either — a standing compliance rule forbids removing a component modifier to obtain payment — and asked the payer to override a false-positive edit. A second-model reviewer reads the held cases; on the 169 it reviewed it concurred 161 times and bounced eight back for more investigation.
The same record shows the cost of the human side of the loop. "Humans have to do their part for the process to work," is how Vitali Khvatkov, Recovr's founder, puts it, and the September 16 snapshot is the evidence. The 159 approval cards had a median age of 11 days. The 315 held cases had a median age of 32 days, and 254 of them had reached the two-attempt cap, meaning the agent would not touch them again until a person did. The 163 held Humana claims — 154 of them asking for an organism or analyte — had been waiting a median of 50 days; the requests went out in late July, and the laboratory that holds the answers had not yet been asked as a batch. A further 128 payer-enrollment tasks, many for Medicare Part B contractors and opened on July 21, were still open eight weeks later, and dozens of claim-edit submissions to those payers were refused at routing in the meantime. The agent's loop closes in 23 hours; the human loop, unmanaged, closes in weeks, and the timely-filing clock does not pause for either.
Three findings follow. First, the agent's request has to be specific enough to act on without opening the claim, and in this pipeline it is — each follow-up chip names the CPT, the date of service, the identifier and the exact fact needed. Second, identical requests need to be batched and routed to whoever holds the answer: 154 near-identical Humana asks are one email to a laboratory, not 154 tickets. Third, human turnaround needs the same instrumentation the agent has — an age on every card, a weekly review of everything at the attempt cap, and an owner for enrollment tasks — because a queue nobody measures is a queue nobody works.
Takeaway. Human-in-the-loop is the right design for modified claims, and it only works if the human loop has a service level. Measure the age of every card and every ask; that number, not the agent's accuracy, will decide the recovery rate.
Practice 7 — Run the timely-filing clock as a first-class variable
Every claim in this cohort was already old when the agent found it: a median of 174 days from date of service, with 58% between 90 and 179 days and 41% between 180 days and a year. Against the windows the payers publish — 90 days for a participating Cigna provider and 180 out of network; a 180-day floor for contracted L.A. Care practitioners; 90 days under several Humana contracts and 365 for Medicare Advantage; 12 months at Blue Shield of California — that means the Cigna claims, and any Humana claims under a 90-day contract, were already outside their initial-filing windows on the day they were resent, the L.A. Care claims were straddling theirs, and the Blue Shield claims still had months of runway.
The book-level consequence is visible in the remittance data. CARC 29 denials — "the time limit for filing has expired" — totaled 16,529 claims and $1.08 million in the twelve months to mid-September 2026, and the monthly run-rate rose from about $40,000 in September 2025 to $217,000 in August 2026. Across the book's recoverable denials, 77% of the recoverable claims have already passed their refile deadline: $8.6 million of billed charges classified as lost against $3.4 million still current.
Two design choices follow from the arithmetic. The age gate for the status sweep is set from payer response times, which is right for detecting silence; prioritization within the sweep should be set from the filing calendar, so that a Cigna claim at day 80 outranks a Blue Shield claim at day 200 regardless of charge. And every resubmission of an aged claim must travel with its proof of original timely submission — the clearinghouse acceptance and the payer's own earlier acknowledgment — because the payer manuals accept exactly that evidence when a CARC 29 is disputed, and this pipeline retains all of it. The 45-day escalation that fired on 176 resent claims with no payer outcome exists for the same reason: an aged claim that goes quiet a second time cannot wait for someone to notice.
Takeaway. Know each claim's filing deadline at the moment it goes silent, rank the sweep by it, and keep the acknowledgment trail that turns a timely-filing denial into a winnable dispute.
Results scoreboard: what moved in seven weeks
Stage | Outcome |
|---|---|
Detection | ~22,200 silent claims given a payer-confirmed status; 65% turned out to be adjudicated-at-zero with no remittance posted |
Gated resend | 1,401 claims resent unchanged across 30 payers; 1,343 accepted by the clearinghouse; 98 held as not covered; 58 retired with a named deterministic cause |
Verdict capture | 615 payer rejections captured at a median of 23 hours; 121 adjudicated denials and 2 payments captured on 835s at a median of 7 days; 383 still in flight |
Diagnosis and fix | 577 rejections diagnosed to a root cause; 80 corrected claims resubmitted with clearinghouse acceptance; the Humana NOS pattern (214 claims, $456,245) reduced to one rule plus one batched information request |
Human handoff | 159 approval cards, 315 held cases with a specific ask, 128 enrollment tasks — each now visible, dated and owned |
Dollars | >$920,000 in charges moved from silent aging into a state where a defined next action exists |
The honest caveat is that conversion to cash is the next chapter, not this one. Two resent claims had paid by September 16; the rest of the money sits in approval cards, information requests and in-flight adjudications whose outcome depends on the human turnaround described in Practice 6. What the seven weeks demonstrate is the part that had never happened before: the claims were found, the payers were made to answer, the answers were decoded, and every one of them was turned into a specific, dated task.
What we would tell a peer
Sweep every claim with no remittance against the payer's own clock, and treat the practice-management status as a claim to be tested. Decode status responses before acting on them, and use a controlled resend as the probe when the status channel cannot answer. Automate the unchanged resend behind gates that check edit history, billing history and eligibility, and retire deterministic failures instead of retrying them. Listen for the 277CA the next morning and sort rejections by payer and code, because each is different work. Fix at the line, validate, resubmit, and write the rule the same week. Keep a person at the send button for anything modified, and put a service level on that person's queue. And carry the filing deadline and the acknowledgment trail with every claim, because the clock is the only variable in this process that no agent can stop.
Methodology and sources
Data. Claim-status, resend, acknowledgment and agent-run records were read from Recovr's production record on September 16, 2026; nightly run summaries from the pipeline's run logs; denial and timely-filing figures from Recovr's denial analytics for the same book, September 18, 2025 through September 17, 2026. Charge figures for the Humana cohort come from the claim-level charge on each 277CA; the ">$920,000 moved" figure sums submitted amounts on the 866 of 1,221 tracked records that carry one and is therefore a floor. Resend counts (1,401) come from the clearinghouse audit log; verdict shares are computed on the 1,123 tracked resends with a current status record.
Code sets. 277CA claim-status category and status codes are as defined by X12 (A3, A6, A7 reject before adjudication; A4/D0 not on file; E0–E2 inquiry errors; F1/F2 finalized). CPT descriptors are paraphrased from the AMA code set.
External references. CAQH Index 2024 (claim-status inquiry cost and time); Experian Health, State of Claims 2025 (54% of providers report rising denials; about half still review claims manually); Cigna, "When to File" (90 days participating, 180 days out of network); Blue Shield of California Independent Physician and Provider Manual, Section 4 (12 months from date of service); L.A. Care, Submitting a Claim (180-day floor for contracted practitioners); Humana Healthy Horizons Indiana claims guide (90 days) and Muni Health summary of Humana limits (365 days Medicare Advantage; clearinghouse acknowledgment as proof of timely filing); Therabill, status code 306 (line-note requirement for non-specific codes).
About Recovr
Recovr, by Pinnacle Services Corporation, is an AI claim-recovery platform for laboratories, physician groups and revenue-cycle companies. It reads the 835 and 837 files a billing system already produces, finds the claims that were denied, underpaid or never adjudicated, and either fixes and resubmits them under the billing team's approval policy or hands an operator a specific, dated task. It runs as a dedicated instance per customer, offers three levels of adoption — Assisted, Human-in-the-Loop and Rules-Driven — and is live within 48 hours of receiving EDI files.
Contact: pinnacleservice.co · Book a demo
