I have read many articles this year wherein both payers and health systems have complained that the other side’s use of artificial intelligence (AI) provides an unfair advantage. Fitch Ratings recently described the battle this way: providers use AI to optimize coding and maximize reimbursement per encounter, while insurers respond with AI that flags aggressive coding patterns and builds the expected impact into pricing. Most hospitals feel that payers are winning the battle, but payers are trying to change the narrative.
The insurers’ latest shot across the bow came from the Blue Cross Blue Shield (BCBS) Association (BCBSA). Its analysis estimates that hospital AI coding tools added $942 million in costs for BCBS plans between 2023 and 2025. The share of inpatient cases billed as medically complex rose from roughly 37 to 40 percent, and about 70 percent of the added cost came from secondary diagnoses that moved claims into higher-paying categories. BCBSA’s evidence rests on a mismatch between diagnosis and treatment. It found anemia diagnoses (the white paper cites acute posthemorrhagic anemia, D62, which is classified as a CC/complication or comorbidity) after major bowel procedures increased without a matching rise in transfusions. BCBSA’s Luke Chalker put it bluntly, saying the disconnect suggests that AI is identifying more billable conditions, not sicker patients. However, as I discussed last week, inpatient acuity is rising because healthier patients are increasingly treated in the hospital outpatient setting.
Hospitals objected to this conclusion. The American Hospital Association (AHA) argued that the analysis insinuates conclusions without reviewing medical records to see whether the diagnoses were clinically supported and reportable. Coding vendors went further. Two executives told Healthcare Dive that the tools are capturing what providers or coders previously missed. One conceded that payer costs are rising, and the other noted that a fee-for-service system pays more when the clinical picture is complete. Meanwhile, HCA’s chief financial officer said in September at an investor conference that hospitals are lagging behind payers on AI-driven claims processing. Health system leaders describe payers automating claims review, downcoding, and issuing denials at a scale providers cannot match. So, are hospitals correcting years of undercoding, or optimizing the bill? Are payers protecting premiums, or protecting margins? Will either side win this war?
I suspect both stories are partly true. The payer side is not a beacon of integrity. Medicare Advantage (MA) plans benefit from every risk-adjusting diagnosis, and several have faced False Claims Act (FCA) allegations of upcoding to increase their payments from the Centers for Medicare & Medicaid Services (CMS). Kaiser Permanente agreed in January 2026 to pay $556 million to resolve such allegations without admitting wrongdoing. DOJ alleged that Kaiser mined patients’ records for unsubmitted diagnoses and sent queries prompting physicians to add them through addenda, sometimes months after the visit. Meanwhile, MedPAC estimates that coding intensity raised MA payments by about 4 percent in 2026. This is a double standard. When diagnoses benefit payers, some report unsubstantiated conditions; when diagnoses benefit hospitals, payers apply validation criteria that go well beyond industry standards and medical norms.
Perhaps there is hope for a ceasefire. Payers face limits when AI is used to deny care. Fitch notes that scrutiny of AI in prior authorization has already led regulators and lawmakers to trim authorization requirements for certain services. It also expects regulatory pressure for human involvement in care denials to persist, partly offsetting the efficiency insurers hoped to gain. In MA, that expectation is already CMS policy. MA plans may not base medical necessity decisions on an algorithm that ignores the individual patient, as KFF summarizes. In other words, a human must stand behind the denial. That requirement should extend to every denial that affects payment, including coding and clinical validation denials.
A recent healthsystemCIO analysis argues that audit trails showing which suggestions came from software and were confirmed by a clinician can help hospitals respond to payer allegations of wrongdoing. An audit trail can explain how a diagnosis entered the record, but it cannot defend a diagnosis the record does not support. Audit exposure grows when providers accept every suggestion, however weakly supported (or even when unsubstantiated).
An audit trail is not a big ask, and should be standard. The 2026 update to the ACDIS/AHIMA Guidelines for Achieving a Compliant Query Practice gives hospitals a reason to require one. The update adds a dedicated compliance framework for technology-generated queries, including prompts, alerts, and nudges (Section XI). It states that “humans remain responsible for ensuring that queries are compliant” and that organizations should “conduct structured auditing of technology-generated query output on a regular basis.”
Tracking the query agreement rate is something clinical documentation integrity (CDI) departments already do, so tracking provider agreement with automated suggestions is a natural extension. It also distinguishes the contribution of autonomous AI queries from that of queries with a human in the loop, which helps facilities measure the return on both their technology investment and their CDI team. A robust tracking process allows physician advisors to intervene when a provider routinely accepts AI-generated suggestions, even when the suggested diagnoses are not clinically supported.
This is also where CDI and coding professionals have something no tool can supply: professional judgment. They provide oversight by validating AI-generated suggestions. Although the ICD-10-CM Official Guidelines for Coding and Reporting can be built into AI logic, that does not mean AI will apply them the way an experienced professional would.
Most importantly, the Uniform Hospital Discharge Data Set (UHDDS) does not require treatment for a diagnosis to be reported on a claim. Section III of the Official Guidelines, which applies the UHDDS definition of other diagnoses, makes a condition reportable when it requires any one of the following: clinical evaluation, therapeutic treatment, diagnostic procedures, extended length of stay, or increased nursing care and/or monitoring.
Abnormal laboratory findings are not coded unless the provider indicates their clinical significance. Look again at BCBSA’s test. Did treatment follow the diagnosis? Payers continue to narrow the criteria used to validate diagnoses, as illustrated by BCBSA’s use of transfusion rates, among other measures, as evidence that an anemia diagnosis reflects real care – even though serial monitoring or iron supplementation, rather than transfusion, may be the clinically appropriate response.
The AHA has also made this point, noting that comorbidities can affect monitoring, nursing resources, and medication management without necessarily resulting in a transfusion or other easily identifiable service (AHA blog, Oct. 5, 2026). Writing in STAT, one resident physician observed that comorbidities act as an on/off switch for billing even though they affect patient care in nuanced ways. Most postoperative anemia is managed without transfusion, so a flat transfusion rate says little about diagnostic accuracy.
A postoperative hemoglobin that nobody evaluates, monitors, or treats is a lab value, not a diagnosis. When software flags it, the correct response is for a CDI specialist or coder to determine whether a reportable diagnosis may be indicated, and if so, to query. The diagnosis should not be an automatic addition to the health record. If the workflow is that the tool suggests, the physician accepts, and the coder reports, the human is not in the loop. We have merely added a signature.
Aetna’s chief medical officer reportedly declined to call this an agentic bot war. I agree, though perhaps for a different reason. A bot war implies humans are bystanders. Humans are the ones writing the notes, answering the queries, and deciding what the record says. If two machines are going to read the same chart, the only durable strategy is a chart that would survive either review.
So, is the battlefield level? Probably not, but the imbalance is less about whose algorithm is faster are about whose documentation can withstand scrutiny.


















