How does AI medical coding work?
A model reads the clinical note, discharge summary or operative report, extracts the diagnoses and procedures, and maps them to standard code sets, integrated with the record system so suggestions appear in the coder’s queue. It is document AI specialized to one of the hardest vocabularies in business: tens of thousands of codes, specialty-specific language, and documentation written under time pressure. The current production pattern is assistive: the model drafts, a certified coder confirms, and the exceptions teach the next version.
How accurate is it really?
Better every year, and not yet trustworthy alone. Academic reviews of automated clinical coding describe genuine progress and call the task far from solved (PMC review, read 19 October 2026); reported scores on standard code prediction tasks remain well short of expert agreement, and performance swings with specialty and note quality. Fluency is the hazard: a wrong code arrives looking exactly like a right one. The honest deployment measures the model against your own coders on your own case mix before any autonomy.
Why is AI coding contested between providers and insurers?
Because the same capability reads two ways. Vendors and providers report AI finding legitimately billable care that manual coding missed; insurers argue the tools inflate reimbursable diagnoses and drive up spending, a dispute Healthcare Dive covered as billions in contested costs (Healthcare Dive, 28 September 2026, read 19 October 2026). A coding model optimizes whatever it is pointed at, and revenue capture and coding accuracy are not the same target, which is why payers are answering with automated claim review of their own.
What does safe adoption look like?
Like billing-grade intelligent automation. Codes above a risk threshold keep a human in the loop; every automated suggestion is logged with the note text that justified it, so an audit can replay the decision; sampling compares automated and human coding continuously; and upstream, cleaner documentation from ambient AI scribes raises the ceiling on what coding automation can do. Deploy it to make coders faster and claims defensible, and the revenue follows; deploy it to maximize revenue, and the audit finds you.