Where RCM breaks today

Costs are rising

Coding errors have always been expensive. What has changed is the scale, the visibility, and whose problem it now is.

Data cited through HFMA from MDaudit found that coding-related denials have increased 126% over the past three years, outpacing authorization-related denials. Inpatient denials rose 220% over the same period, reaching an average of $10,000 per claim. External audit volumes doubled. A February 2026 survey from Fierce Healthcare found that most RCM teams are now spending between 51 and 75 hours per week managing denial-related work, and nearly half reported losing between 3% and 4% of net patient revenue to denials, underpayments, and missed timely filing windows.

This has moved denial management from an operational nuisance to a boardroom conversation. More than a third of RCM executives in that same survey said denials are now discussed at the executive level of their organizations.

The issues in RCM

Denials get most of the attention, but they are often the downstream symptom of problems that started much earlier. There are three places where revenue tends to quietly bleed out.

The first is in the clinical documentation itself. When documentation is incomplete or ambiguous at the point of care, coders are left making judgment calls that should not be judgment calls. They may code what is clearly supported and leave out what is only implied. Across thousands of encounters, that conservative interpretation adds up to significant undercoding, not from error but from caution.

The second is in inconsistency. The same clinical note, coded by two different people, can produce different results. There is no malice in that. It is just the nature of a complex, rules-heavy task done under volume pressure by people each applying their own interpretation to guidelines that were not designed for easy application at scale. For health systems navigating tighter payer scrutiny, that inconsistency is a compounding liability.

The third is in what happens after a denial. Many denied claims are never resubmitted and not because they were legitimate denials, but because appeals require time, coordination, and documentation that most teams do not have. Revenue that was coded correctly but denied anyway can disappear. Hospitals and health systems are now estimated to spend close to $19.7 billion annually managing denied claims, according to a Premier survey.

We need more sophisticated tools

The industry's response to coding pressure has largely been more tools layered on top of the same process.

Legacy NLP systems and dictionary-based CAC tools speed up pattern matching, but they cannot reason over clinical text. They surface everything that could potentially be coded, leaving coders to sort through false positives and apply the judgment the software cannot. The output requires review. The queue still exists. The workload shifts without shrinking.

Generic LLMs are the popular next step for teams trying to build coding into their products. Integration overhead is low, and recall on common codes is reasonable. But there is no embedded rules logic. A generic model does not know what a trained coder knows about what to include or exclude given the clinical context. It cannot distinguish between a condition that should be coded because it drove care decisions and one that should not because it was incidental. The result is output that looks complete but is not auditable, and not reliable under payer scrutiny.

The deeper issue is structural. Most of these tools are designed to help coders do the same work faster. They are productivity plays, not accuracy plays. And in an environment where prior authorization requirements have increased by 30% over the past three years and payers are using increasingly sophisticated audit tooling, speed without accuracy does not move the needle on what actually costs health systems money.

The payer side of the equation

It is worth being honest about something the RCM conversation often underplays. Denials are not just a provider-side problem. Payers have gotten considerably more aggressive. A 2026 survey found that RCM leaders now view payer behavior, not internal operational gaps, as the primary driver of revenue risk. That is a shift from prior years, when internal optimization was seen as the main lever available.

This matters for how you think about what the coding function needs to do. It is not enough to code accurately in a general sense. The output needs to be defensible at the code level, with documentation support that can survive an audit. That is a different bar than simply getting the right code most of the time.

From documentation to code

The bottleneck in RCM today is not a staffing shortage that more hires will fix, and it is not a speed problem that faster tools will solve. It is a documentation-to-code translation problem that compounds at scale: inconsistent inputs, rules-heavy judgment applied unevenly, and output that often cannot be defended when challenged.

The organizations that are starting to get ahead of this are the ones treating coding accuracy as a systems problem rather than a coder performance problem. That means structuring the workflow so the hard judgment calls are the exception, not the default. It means output that is auditable at the claim level, not just accurate in the aggregate.

That is where AI-augmented coding earns its place. Not as a faster version of the same broken process, but as a different architecture entirely.

Next in this series: What does “good” looks like in automated medical coding? Defining the standard for accuracy, auditability, and speed.