The denials nobody works

Change Healthcare's analysis of roughly $3 trillion in submitted claims found about 9% denied on first pass, some $262 billion a year. More recent revenue-cycle benchmarking has tracked initial denial rates closer to 11%. From the same body of work comes the finding that matters more: roughly two-thirds of denied claims are never reworked at all, despite a majority being recoverable. Reworking one costs somewhere between $25 and $118 in staff time depending on complexity. Do that arithmetic against a claim worth a few hundred dollars and the write-off is the correct business decision.

The starkest version of the pattern is on the payer side. KFF's analysis of HealthCare.gov transparency data found insurers denied roughly 19% of in-network claims, and that consumers appealed fewer than 1% of those denials. Almost nobody pushes back, on either side of the transaction.

The second problem is that the winning argument is almost always the payer's own policy language, and that language lives across PDFs, portals, and bulletins that change quarterly. An appeal citing the payer's current criteria and the specific chart pages that satisfy them performs very differently from one that argues in general terms. Finding that provision is research work, and research work is exactly what a specialist with a queue longer than the day cannot do.

An agent team that treats the write-off pile as the target rather than the overflow.

• Triage reads each denial and classifies it by actual root cause, which is frequently not what the denial code says, because payers miscode denials routinely. Coding error, documentation gap, medical necessity, eligibility, timely filing, authorization, bundling, or payer error.
• Policy retrieval locates the governing payer provision and pins it to a document and effective date, so an appeal argues against the criteria in force on the date of service.
• Documentation assessment checks whether the chart actually supports the argument and names the specific gaps when it doesn't, which is often more valuable than the appeal itself.
• Appeal construction drafts the letter, citing policy provisions and referencing chart pages by location.
• Root cause analytics clusters denials back to the upstream defect, such as a service line, an ordering provider's documentation template, or a payer behavior change, so the same denial stops recurring.
• Orchestration ranks the queue by expected recovery against appeal deadlines, and escalates anything approaching a filing limit.

An appeal that asserts a clinical fact not in the record isn't an error. It's fraud.

So the appeal agent is architecturally forbidden from generating clinical content. It may quote and cite what is documented; it may not infer that a condition existed, however strongly the surrounding record implies it. A verification pass confirms that every clinical assertion in a draft maps to a specific chart location, and a draft with an unsourced assertion fails rather than ships. Where the record doesn't support the argument, the correct output is “this documentation gap blocks the appeal,” not a more persuasive sentence.

The same discipline applies to codes. The system proposes; a certified coder approves. Nothing about upcoding exposure is worth automating away.

Denials management has a property most business processes lack: the payer grades your work. The remittance advice comes back in a structured, well-specified format and tells you whether you were right automatically, every day. That means the system can be measured continuously against ground truth rather than against a human's opinion of its output, and it improves on evidence rather than on intuition.

We measure it that way deliberately. Overturn rate on agent-prepared appeals, tracked per payer and per denial category, is the number that matters, not how good the letters look.

Coders approve every code change. Staff review every appeal before submission. The system never touches clinical decision-making; it operates entirely on the reimbursement side of the business, which is a line we drew explicitly and documented, because it's also what keeps a system like this clear of medical-device territory.

Claims that were previously written off unexamined now get worked. In the first three months, the system prepared appeals on 2,100 denials below the old economic threshold, with an overturn rate of 41% and $185,000 recovered against a baseline of zero. Root cause clustering has since reduced missing-prior-authorization denials by 28%.