By Caitlin Walker | Coronis Health
Why healthcare denial management keeps getting more expensive even as platforms get smarter. The answer is rarely in the workflow leaders are trying to fix—it’s in the RCM technology underneath it, which most specialty groups never asked to inherit.
Healthcare denial rates have moved from a long-standing benchmark of around 8 percent to roughly 11 percent in the past two years. Healthcare denial management is now its own cost center: a recent Premier survey of 516 hospitals put the cost of overturning denied claims at $19.7 billion in 2022. Most of the public conversation around those numbers focuses on payers; narrower medical-necessity policies, AI-driven adjudication, tighter prior-authorization rules. The piece IT and revenue cycle leaders can control sits one layer down. A large share of specialty groups are still running RCM platforms whose rule engines were architected for primary care and general medicine, and most of the denials those groups pay to manage trace back to a platform doing exactly what it was built to do for a kind of practice it was not built to support.
In over a decade of working alongside multispecialty groups and health systems as their RCM partner, the pattern is consistent. Anesthesia, radiology, oncology, cardiology, ABA, ASCs— different clinical worlds, same architectural mismatch. These platforms aren’t broken. They were built around a billing model that primary care still uses and that specialty practices have outgrown. Most of the denials we work through with clients trace back to that gap, not to anything the workflow can fix on its own.
Where ‘Generic’ Rule Engines Actually Break
Anesthesia is one of the clearest examples. Anesthesia billing runs on a temporal axis: base units, time units, modifying units and qualifying circumstances, layered on top of concurrency rules that determine whether a clinician is personally performing the procedure, medically directing or medically supervising at any given minute. Generic engines rarely calculate those time-based algorithms dynamically, and they routinely drop qualifying circumstances such as controlled hypotension, extreme age or total body hypothermia. The pattern is consistent across anesthesia clients running generic platforms: qualifying-circumstance modifiers get stripped on edge cases because the rule was written for elective outpatient procedures, claims deny, coders spend twenty minutes finding what got cut and revenue arrives weeks late. Multiply that across a month of claims and the cost is meaningful.
Radiology has its own version of the same problem. Claims split into a technical component and a professional component, with Modifier 26 distinguishing the two, and small configuration
All information contained in this document is proprietary and confidential. Any reproduction and/or dissemination of this proposal without the express consent of a Visual Identity Group representative is prohibited.
Errors trigger immediate duplicate-billing rejections. The No Surprises Act added Independent Dispute Resolution workflows on top of that, with tight statutory deadlines and heavy documentation requirements. Generic platforms typically lack centralized IDR tracking, which extends accounts receivable days and pushes more work onto already short-staffed billing teams.
Ambulatory surgery centers compound both problems. ASC facility claims live on the CMS 1500, require POS 24 and often need the SG modifier in the first position. CMS publishes an ASC-approved procedure list, and anything off that list is denied on contact. A Becker’s ASC Review denial study found that 38 percent of ASC denials traced back to non-covered services, often because the patient’s commercial plan simply does not cover the procedure in an ambulatory setting. Terminated procedures introduce another modifier surface: Modifier 73 for procedures terminated before anesthesia, Modifier 74 for those terminated after anesthesia or after the procedure begins and Modifier 52 for discontinued radiology each carry different documentation and payment consequences. Treating commercial payers like Medicare in an ASC, particularly around bilateral procedures and separately billable biological implants, is one of the faster ways to bleed margin from a roll-up.
Cardiology runs into the same problem through prior authorization on advanced imaging. Oncology runs into it through drug coding and infusion authorization. ABA therapy through authorization caps and treatment-plan documentation. Plastic and reconstructive surgery through reconstructive procedures getting misclassified as cosmetic. The clinical details look different in every case, but the platform breaks for the same underlying reason: it was built around an average claim that very few specialty cases actually look like.
Healthcare Denial Management Beyond CARC Codes
Even when denials happen, generic systems make it hard to learn from them. Payers communicate decisions through Claim Adjustment Reason Codes (CARCs) and Remittance Advice Remark Codes (RARCs), but CARCs only describe what the payer’s system concluded, not what went wrong inside your operation. A CO-197 marked as “authorization absent” can mean four different things: the front desk never requested it, the payer denied the request, the wrong CPT was authorized or the RCM engine failed to map a valid authorization number onto the EDI 837. Each one has a different upstream fix and the payer code doesn’t tell you which one you’re looking at.
When a denial doesn’t fit a clean bucket, generic platforms drop it into a generic “other” category and reviewers prioritize cash recovery over root-cause coding because they’re measured on resolution speed. The data needed to prevent the next denial gets lost as soon as the appeal goes out. HFMA has been direct on this: organizations that enforce a structured internal taxonomy and assign a root cause at the time of initial review make measurably better upstream changes than those relying on CARC data alone. Across the client portfolios we work through, most denials cluster into three categories—end-user error, technical setup issues or patient requirement failures, and a real taxonomy lets you separate them.
AI-augmented Coding Does Not Fix This
The 2026 AMA CPT set formally recognizes AI in medical coding through Appendix S, which splits services into assistive, augmentative and autonomous tiers. New AI-augmented codes cover coronary plaque assessment, perivascular fat analysis (0992T, 0993T), digital pathology pattern recognition and retinal imaging, among others. That’s a real step forward, and it’s also a new surface for denials if the platform underneath the AI isn’t built to handle specialty logic. An AI-augmented coder running on top of a rule engine that doesn’t understand the specialty won’t solve the denial problem. It will just process the same bad claims faster. Before relying on AI-augmented coding for specialty CPT sets, operators should be able to answer five questions:
- Medical necessity linkage. Is the AI tying ICD-10 to CPT according to the specific payer’s local coverage determination, not a generic mapping?
- Documentation alignment. Does the clinical note explicitly document the physician’s review and final interpretation of the AI’s output? Payers are auditing this, and minor omissions deny.
- Confidence thresholding. Does the system assign a confidence score to each chart and route lower confidence or higher risk cases, such as multi system reconstructions or unusual modifier combinations, to human coders before straight through processing?
- Explainability. Does the AI cite the exact EHR passage it relied on, with one-click human correction that feeds back into the model? Black-box outputs are not compatible with healthcare compliance.
- Specialty fit. Was the model fine-tuned on specialty-specific operative notes, not general discharge summaries? Dermatology lives on lesion sizes. Cardiology lives on device placement. Generic training data quietly under-codes both.
The principle behind those five questions is what already shapes how revenue cycle teams think about AI: narrow models built for one thing tend to beat generalist models trying to do everything. The platform underneath the AI works the same way. A specialty-aware substrate gives AI augmentation something useful to do; a generic one just gives it wrong answers to deliver faster.
What to Evaluate Before Trusting a Specialty RCM Stack
When we work with clients to evaluate or rebuild an RCM stack for a specialty group, four areas tend to separate platforms that hold up from ones that demo well.
The first is interoperability and front-end capture. Bidirectional EHR, PACS/RIS and practice management integration; direct payer APIs or modern clearinghouse connections for real-time eligibility and prior authorization; and dynamic payer mapping that updates for 2026 CPT additions and AI-augmented service crosswalks. If the platform can’t ingest current code sets in something close to real time, the rest of the stack inherits that lag.
The second is prior authorization and clinical documentation integrity. AI-driven form population, EHR-native auth workflows and operative-report auditing that flags missing implants, blood loss or complications before a coder finalizes the claim. Inpatient teams should also have DRG validation checklists. This is where most upstream prevention happens, and it’s the layer generic platforms tend to under-invest in.
The third is IT security and AI governance, which is where most specialty groups need the most outside help. The list is long: PHI classification on AI outputs, retrieval-augmented generation pipelines scoped to minimum necessary data, encrypted and audited vector stores, shadow-AI detection on clinical networks, business associate agreements that address training opt-outs and inference-layer breach notification, and prompt and response logging tied to your SIEM. Healthcare is still catching up to compliance expectations that defense and financial services have treated as baseline for years. Getting it right now is much cheaper than retrofitting it later.
The fourth is reporting that’s actually actionable. Clean Claim Rate in the 95 to 99 percent range, First Pass Resolution Rate, Net Collection Rate near 100, accounts receivable days, denial rate sliced by root cause and payer. If you can’t pivot any of those by specialty, payer and provider, the reporting layer is not specialty-grade.
The Bottom Line
Generic RCM technology isn’t failing because vendors built bad software. It’s failing because the logic underneath it doesn’t match the clinical and regulatory reality that specialty groups now live in. Another bolt-on rule pack won’t close that gap. Neither will an AI coder running on top of a substrate that was wrong to begin with. What closes it is harder work. Specialty-aware architecture under the platform. AI augmentation on top of a substrate worth augmenting. Denial taxonomies built to catch what CARC codes were never designed to show. Partners who can see the pattern across specialties rather than retrofitting the same generic engine one practice at a time. Real medical billing denial management is architecture work, not a workflow problem to hand off. The groups that recognize that in 2026 will spend less time appealing denials and more time taking care of patients. The ones that don’t will keep paying for the gap, one $450 medical-necessity denial at a time.
This article is for informational purposes only and is not intended as legal, financial, coding or compliance advice. Practices should consult qualified advisors and internal compliance resources regarding payer requirements, documentation standards, billing rules and regulatory obligations.
