Accuracy Is the Common Ground in the AI Coding Debate

By Missy Harbert, Solution Advisor, Revecore 

A new Blue Cross Blue Shield Association analysis estimates that rising inpatient coding complexity added about $942 million in costs for Blue plans from 2023 through 2025. The share of cases billed as complex climbed from roughly 37% to about 40%. BCBSA ties much of that growth to secondary diagnoses such as anemia and low sodium, which AI-assisted tools can detect from a single lab value. 

When Healthcare Brew asked for my perspective, I started with a principle I think payers and providers already share. Nobody in the revenue cycle should defend a code the clinical record doesn’t support, and accuracy has to be the standard on both sides of the claim. 

What claims data can and can’t show 

BCBSA’s approach looks for diagnoses that appear without the treatment you’d expect. At hospitals with the fastest coding growth, anemia diagnoses after major bowel procedures ran 38% higher than at peer facilities, while transfusion rates among those patients were lower. That pattern is worth examining closely. 

The analysis is built on claims, though, and claims can’t tell a diagnosis that was added inappropriately from one that was always present but previously missed. Many clinically meaningful conditions never lead to a separate procedure like a transfusion. They change how a patient is monitored, how long they stay, or their risk of complications. The AHA raised a similar point in its response, arguing that the central question is whether the medical record supports the diagnoses. Brown University health economist Christopher Whaley offered a balanced read to CNBC: some added diagnoses are legitimate conditions that simply weren’t captured before, while others may have little bearing on the patient’s care. 

Both sides of the claim deserve attention 

During the same period, payers scaled their own automation to review, deny, downgrade and underpay claims. For many hospitals, better documentation technology arrived as a response to that pressure. McLaren Health Care’s CFO made this point to The New York Times, noting that hospitals are reviewing the same charts insurers already scan for denials. 

The costs on this side are harder to see because they never appear as a premium line item. Hospitals spent nearly $18 billion overturning claim denials in 2025, according to the AHA. A Health Affairs study of Medicare Advantage claims found 17% of initial submissions were denied and 57% of those denials were ultimately overturned. When a denial that would likely be reversed goes unappealed because a hospital lacks the staff, that revenue is lost, with considerable consequences for services and access. 

We now have detailed analyses of how hospital AI affects coding, but very little public data on how payer algorithms make denial and payment decisions or how often those decisions are overturned. Both deserve the same scrutiny. 

Where AI is already improving efficiency 

In our work, the clearest wins come from taking administrative work out of the system while keeping clinicians in every decision that matters: 

• Machine learning ranks denied and underpaid claims by their likelihood of recovery, so teams can focus their time where it will make a difference.  This ranking also systematically sets threshold limits based on propensity for payment rather than setting a threshold that is the same for all claims. 

• AI summarizes long medical records and surfaces the clinical evidence that supports an appeal, and a nurse reviews every appeal before it’s submitted. We’ve seen this cut appeal preparation time by more than half. 

• Root-cause analytics spot patterns, such as a payer repeatedly downgrading a certain DRG or denying a specific service. This year, that work helped one health system catch a mammogram denial trend across several payers early enough to correct it at the source. 

Of the three, root-cause work has the most potential to save money for payers and providers alike, because a problem fixed upstream never turns into a dispute. 

What a practical truce could look like 

Leaders from Cigna and Emory Healthcare recently shared a stage to discuss easing the coding wars, and Emory’s CEO described a future of competing bots endlessly negotiating as his nightmare. In practice, a truce could mean both sides keep their AI tools and agree on a few ground rules for using them: 

• Transparency: Hospitals explain how AI shapes coding, and payers explain how AI shapes denials and payment. 

• Human oversight: No code is escalated and no claim is denied on the basis of an algorithm alone. BCBSA’s David Merritt has called for appropriate human involvement as well, which suggests real common ground. 

• Shared clinical definitions: Much of the conflict centers on sepsis, malnutrition and anemia. If both sides agreed on criteria up front, I suspect much of the downstream conflict would fade. 

• Earlier resolution: Settling a claim months after the fact, through appeals and audits, is the most expensive path available. 

Coding and reimbursement should reflect the care delivered. That’s the standard we apply when helping hospitals prevent denials and recover underpayments, using technology to find patterns and clinical expertise to confirm the evidence behind each claim. As AI becomes more embedded in revenue cycle workflows on both sides, I expect that standard to carry even more weight.