Why the Same RCM Denial Code Means Three Different Things
RCM teams sort denials by CARC, but the same code routes three different ways at three different payers. The standard is real. The interpretation is not.
TL;DR. Mid-market RCM teams sort denials by CARC, but the same code routes three different ways at three different payers. CO-16 is meaningless without the RARC. CO-50 reflects whichever medical-necessity playbook the payer's reviewer is using. CO-97 follows bundling logic that Medicare publishes and commercial payers do not. A denial work queue grouped by CARC alone is fast to build and easy to misread.
A medical billing director at a $40M revenue cycle firm shows you the weekly denial report. CO-16 sits at the top with 22 percent of denied dollars. She tells you the team is "working CO-16 cases." Within the next thirty minutes, you watch three specialists open three different claims with the same CO-16 code and route them three different ways. One is a missing rendering provider NPI. One is a missing prior-auth number. One is a member ID that does not match the eligibility file. Same code. Different team. Different fix. Different turnaround.
This is what every mid-market RCM operator knows and every AI denial-management demo misses.
The denial code is standardized. The action is not.
CARC stands for Claim Adjustment Reason Code. The list is standardized by X12, the body that maintains the 835 electronic remittance advice format, and CMS publishes additions and deactivations three times a year. There are 358 active CARCs. Every payer that processes claims under HIPAA uses the same code set, by law.
That is where the standardization ends.
Three codes drive most of the denied dollars at any mid-market RCM firm: CO-16 (claim lacks information), CO-50 (not medically necessary), and CO-97 (service bundled into another payment). Each one means something different at different payers, and the operational implications are not subtle.
CO-16 is the empty box. The code says "claim/service lacks information needed for adjudication." It says nothing about which information. The Remittance Advice Remark Code (RARC) paired with the CO-16 is what tells you the actual problem. CO-16 with RARC N290 means missing or invalid ordering or referring provider NPI, which is a registration data fix, routes to the front-end team, and resolves in a corrected claim. CO-16 with RARC MA27 means missing patient identification, also registration, also a corrected claim, but the fix touches a different field. CO-16 with RARC M127 means missing patient medical record, which is a documentation pull, routes to the clinical team, and takes two or three days because someone has to chase the chart. Three different RARCs, same CO-16. Three different teams, three different SLAs, three different recovery rates.
"A clean CO-16 with no remark code is the worst kind of denial. The system tells you exactly nothing. You log into the payer portal, find the original claim, read the EOB image, and reverse-engineer the actual reason. That is a forty-five-minute task to do work that should have been five."
Most mid-market denial dashboards group by CARC. That weekly report showing CO-16 at 22 percent of denied dollars is a fiction unless it also shows the RARC breakdown.
CO-50 is the payer's medical-necessity reviewer making a judgment call. The code says "not deemed a medical necessity by the payer." That is a clinical decision made by a person, usually a nurse reviewer applying a published criteria set. Most large commercial payers use MCG or InterQual. Some use proprietary criteria. Medicare uses NCDs (national coverage determinations) and LCDs (local coverage determinations) that vary by MAC. The same procedure on the same patient with the same documentation can pass CO-50 review at Aetna and fail at UnitedHealth, because the underlying criteria are different.
The operational fix is not "appeal CO-50 denials." The fix is "appeal CO-50 denials with payer-specific clinical documentation aligned to that payer's published criteria set." A peer-to-peer review with an Aetna medical director that cites MCG criteria has a different success rate than the same letter sent to UnitedHealth, which is applying its own internal policy. The HHS Office of Inspector General has found that overturn rates on CO-50 medical-necessity denials can exceed 80 percent when the appeal documentation is built for the right payer. They drop to under 20 percent when the appeal is generic.
CO-97 is the bundling decision. The code says "service bundled into another payment." Medicare's bundling logic is published. The National Correct Coding Initiative (NCCI) maintains edits showing which procedure pairs are bundled and which modifiers can unbundle them. You can look up the edit. You can appeal with the right modifier.
Commercial payers have their own bundling rules and they do not publish them. UnitedHealth has Optum-derived bundling logic. BCBS plans bundle differently by state. A CO-97 denial from Cigna is not appealed the same way as a CO-97 from Medicare, because the underlying edit is not knowable from outside the payer's claims system. The shop has to figure out the rule by pattern recognition across hundreds of denials.
Three codes. Three different operational paths. Three different teams.
Why your work queue is lying to you
Open the denial management module in any major RCM platform, whether Waystar, Availity, Change Healthcare, athena, eClinicalWorks, or NextGen, and sort by reason code. The work queue is fast to navigate, the dollar volume is visible at a glance, and a denial specialist can knock through a list of "CO-16 denials" in a morning.
The problem is that the morning's work was not actually homogeneous. The specialist who worked thirty CO-16 cases routed twelve to registration (different RARCs, different missing fields), seven back to coding (modifier-related missing data), six to the chart-pull queue (missing documentation), four to eligibility (member ID mismatches), and one to the appeal team because the original RARC was wrong. The CARC was the same. The work was not.
This shows up in two places. The first is staffing ratios. The MGMA 2024 Cost and Revenue Report found that the average initial denial rate hit 11.8 percent in 2024. Experian Health's 2025 State of Claims survey found that 41 percent of providers see denial rates of 10 percent or higher. The conventional staffing ratio is roughly 0.5 FTE of denial work per 1,000 claims per month. At that ratio, a $40M RCM firm processing 800,000 claims a year ends up with somewhere between 30 and 50 FTEs working denials. If those FTEs are routed by CARC instead of by what the work actually requires, you have eligibility specialists working chart-pulls and coding specialists doing front-end registration fixes, and turnaround stretches.
The second place is reporting. The denial trend report that the RCM director shows the practice owner each month says "CO-16 dropped from 22 percent to 18 percent." That number can move because the front-end team fixed a registration workflow, because the coding team caught a modifier issue, or because a clearinghouse upgrade started passing more claims clean. The reason matters more than the number. A drop in CO-16 driven by front-end fixes is durable. A drop driven by a clearinghouse change can reverse the next time the payer updates its edit logic.
The fix in the work queue is mechanical, and most firms have not done it: group by CARC plus RARC plus payer. Aetna CO-16 with RARC N290 is a queue. UnitedHealth CO-50 with documentation-gap RARCs is a different queue. Medicare CO-97 NCCI bundling is a third. The dashboard suddenly has thirty queues instead of ten reason codes, and the staffing ratios start to map to what each queue actually requires.
Most mid-market firms have not built this view because the underlying analytics tooling does not surface RARC alongside CARC, or because the team report has been the same shape for ten years and nobody wants to break it. The cost is roughly a percentage point of net collection rate, distributed across slow turnaround, mis-routed claims, and dropped appeals.
Where the AI demos stall
This is where AI denial-management pilots get into trouble. The vendor demo shows an LLM reading the ERA, classifying the denial, drafting the appeal letter, and routing it. The training data is CARC-labeled, because that is the only label that comes pre-tagged in the 835 file. The model learns "CO-50 means appeal with clinical documentation." It does not learn "Aetna CO-50 with MCG criteria means appeal with this specific clinical pattern; UnitedHealth CO-50 means appeal with a different clinical pattern; Medicare CO-50 means appeal with NCD or LCD citations."
The pilot looks good on a slide deck. It works on Medicare claims because Medicare publishes its rules. It does not work on commercial payers, and the practice ends up with appeal letters that get rejected at higher rates than what the specialist team was doing manually. The vendor blames the data. The data is fine. The model never had the payer-specific context.
There is a real role for AI in denial management. It is not the one in the demo. The useful version reads the ERA, pulls the RARC, looks at the payer, pulls the prior twelve months of that-payer-with-that-CARC-and-RARC outcomes from the firm's own history, and surfaces the appeal pattern that worked. The model is not a denial classifier. It is a pattern-matcher against the firm's institutional memory. That is harder to build than what the vendor demo showed, and it requires the firm's denial history as training data, which most firms do not have in clean, structured form.
The same pattern shows up across mid-market AI adoption, not just in RCM. We wrote about why AI pilots stall at the same point every time, and the gap between the vendor demo and the actual operational complexity is where most pilots die. RCM denials are an especially clear example because the operational complexity is structured (CARC plus RARC plus payer) and the demos consistently ignore it.
The mid-market RCM firms that get value from AI on denials start by fixing the work queue first. They group by CARC plus RARC plus payer. They map the historical outcomes. They give the AI a smaller, cleaner problem to solve: "draft the appeal letter for this Aetna CO-50 with this clinical context" rather than "automate denials." The pilot ships, the metric moves, the team trusts the tool. Then it expands.
FAQ
Does this mean we should not use a CARC-based work queue at all? You can start there because that is what most platforms give you out of the box. But the production view needs to be CARC plus RARC plus payer. The CARC tells you the category, the RARC tells you the specific defect, and the payer tells you the appeal pattern. Any of those three alone misroutes the work.
How many RARC variants matter in practice? For a mid-market RCM firm, the answer is usually thirty to fifty RARCs across the three or four CARCs that drive most denied dollars. Beyond that, the long tail can sit in a general queue. The first time a firm builds this view, the surprise is usually that two or three RARCs make up the bulk of CO-16 volume, and those two or three are fixable upstream.
What about payers that do not pair RARCs with CARCs cleanly? Some commercial payers report partial RARC data or use proprietary remark codes that map roughly to the standard set. The CAQH CORE Payment and Remittance Rule requires alignment with CORE-defined business scenarios, but enforcement is uneven. When a payer routinely sends ambiguous remits, the only fix is a payer-specific playbook built from the firm's own outcomes data.
Where does this fit in a broader RCM operations rebuild? The denial work queue is one of several systems that mid-market RCM firms outgrow at the $30M to $50M revenue band. We covered the playbook for cutting denials in How $40M RCM Firms Cut Claim Denials From 18% to 8%, and the clearinghouse layer in Waystar vs Availity vs Change Healthcare for $40M RCM.
If your denial work queue is sorted by CARC alone, the next step is not buying an AI denial tool. It is rebuilding the view around CARC plus RARC plus payer and getting the FTE routing right. The broader denial-workflow rebuild lives in How $40M RCM Firms Cut Claim Denials From 18% to 8%. If you want help wiring the queue and the trend reporting into your existing platform, book 30 minutes with Granular. Fixed price, four-week delivery, working tool.
Keep Reading
- How $40M RCM Firms Cut Claim Denials From 18% to 8% - The operational playbook for cutting denial rates at a mid-market revenue cycle firm, from front-end registration through appeals workflow.
- Why AI Pilots Stall at the Same Point Every Time - Why mid-market AI pilots reliably get to a demo and then die before production, and how to design around the failure pattern.