AI helps billing—but with limits

Medical billing is the part of healthcare that turns a visit into a claim, and revenue cycle management is the longer chain that runs from patient registration to the final insurer or patient payment. The current argument is not whether artificial intelligence can help, but which links in that chain are repetitive enough for software and which still depend on human judgment. (ahima.org) (hfma.org) That question got louder because denials are rising. Medical Economics, citing Experian Health and Healthcare Financial Management Association data, says 41% of providers now report that more than one in 10 claims is denied, initial denial rates reached nearly 12% in 2024, and up to 65% of denied claims are never reworked. (medicaleconomics.com) (hfma.org) The most useful version of artificial intelligence in billing looks less like a robot biller and more like a spell-checker before you hit send. Medical Economics says the strongest use cases are presubmission checks such as real-time eligibility verification, coordination-of-benefits review, demographic checks, and payer-rule matching before a claim goes out. (medicaleconomics.com) That is a shift from fixing errors after a denial to preventing errors at the front desk, in the chart, or in the claim scrubber. In the Medical Economics piece, Purnendu Bala argues that the operational win is fewer downstream touches, not replacing billing staff with a general-purpose model. (medicaleconomics.com) The trouble starts when artificial intelligence moves from checking forms to interpreting medicine. Healthcare IT Today used hierarchical condition category coding as the example, which is the Medicare risk-adjustment system that groups diagnoses into categories tied to expected future costs. (healthcareittoday.com) (cms.gov) (aapc.com) On paper, hierarchical condition category coding looks perfect for automation because software can scan charts faster than a person can. In the April 9, 2026 Healthcare IT Today commentary, Ritwik Jain says many tools instead flood clinicians and coders with low-quality suggestions that miss clinical context, forcing humans to sort noise from signal. (healthcareittoday.com) That failure is expensive because risk adjustment pays on documented disease burden, not on what a model guesses was probably true. The Centers for Medicare & Medicaid Services says its hierarchical condition category model depends on diagnosis coding and model software, which means unsupported suggestions can create audit risk as easily as missed revenue. (cms.gov) So the dividing line is not “billing” versus “coding.” The dividing line is whether the task has a clear right answer, like a missing subscriber number, or requires clinical nuance, like whether a diagnosis is documented with enough specificity to support a risk-adjustment category. (medicaleconomics.com) (healthcareittoday.com) Both pieces land in almost the same place. Use artificial intelligence where you can measure fewer manual touches, cleaner claims, faster eligibility checks, and better audit trails, and keep humans in the loop where documentation logic and payer interpretation still decide whether money is paid or clawed back. (medicaleconomics.com) (healthcareittoday.com)