Hospital leadership has heard the automation promise before. Implement this platform, the pitch goes, and your revenue integrity team will finally get ahead of denied claims, underpayments, and timely filing lapses. And yet, according to the American Hospital Association, commercial insurers’ claim denial rates have roughly doubled over the past decade, and hospitals are still losing out on tens of millions annually to payment gaps they can’t fully see. The tech and tools changed, while the problem didn’t.
The failure has a root cause that the industry has been slow to confront directly: revenue cycle technology has never been designed around the actual structure of claims data. Until it is implemented that way, automation will keep underdelivering.
What rules-based automation actually does
The dominant paradigm in revenue cycle management automation this century has been robotic process automation and rules-based workflow tools. These systems do exactly what they promise, but only within tightly controlled parameters. They can execute a defined sequence of steps across a specific system when the inputs match a specific format.
That’s precisely where they break down in real hospital environments. Claims data is not structured, and lives across multiple EMRs, billing platforms, and payer portals that weren’t designed to communicate with one another. The data comes in via raw EDI files (highly formatted but semantically complex), and it reflects inconsistent coding practices, payer-specific contract terms, and a constantly shifting landscape of reimbursement rules. Rules-based tools need clean, predictable inputs, but hospital revenue cycle data rarely provides them.
The well-documented results are high exception rates, a lot of manual intervention to keep automation running, and revenue integrity teams putting way too much of their time on rework rather than pattern analysis. The HFMA has consistently found that the cost to rework a single denied claim runs from $25 to well over $100 depending on complexity (and that’s before accounting for claims that never get reworked at all).
Generative AI helped the wrong part of the problem
The wave of AI tools that have entered healthcare over the past two years made a genuine contribution in areas like drafting Letters of Medical Necessity, producing patient-facing summaries, and generating appeal language faster than any human team. Certainly real value there…but also, from a revenue integrity standpoint, the easy part. Writing a better appeal letter doesn’t identify why a claim was underpaid in the first place. Nor does it surface the pattern of a specific payer who is consistently reimbursing below contract rates for a particular DRG, or flag a filing deadline that’s 72 hours away on a high-dollar account. AI has made revenue cycle teams more productive writers but has stopped very short of changing the underlying intelligence problem.
An architectural shift that changes the equation
Rather than just executing predefined rules against structured inputs, agentic systems reason over unstructured data to determine what action is needed. Then it takes it end-to-end. Applied to revenue integrity, this distinction is significant.
The most meaningful difference is where in the data stack these systems operate. An agentic AI workflow built for claims can work directly at the EDI file level, ingesting raw transaction data holistically instead of waiting for normalization and loading into a downstream system. That means it does not require structured inputs or deep integrations to function. It can operate across multiple EMR environments at the same time, reconciling billing records, patient data, and payer information that have never before lived in the same space.
Beyond data access, the intelligence model itself is different. Implemented correctly, agentic systems can compare actual payments against contracted rates, identify coding errors before claims are submitted, track aging accounts against filing deadlines, and find underpayment patterns across payer relationships (without a human defining every rule in advance). In short, they learn from data rather than requiring data to conform to predetermined logic.
The stakes are too high for incremental progress
The financial pressure on hospitals has reached a level where one-step-at-a-time improvements to revenue cycle efficiency are no longer sufficient. Kaufman Hall reported that hospital operating margins remain very thin, with labor costs and reimbursement pressure compressing financial reserves at institutions that can least absorb the impact. So when a mid-sized hospital loses 2-5% of net patient revenue to denied and underpaid claims (a range that revenue integrity professionals consistently cite as realistic), that is not a rounding error. At an average net patient revenue of $242 million, that range represents $5 to $12 million annually in recoverable funds.
Among the hospital leaders I’ve spoken with over the past few months, the industry is clearly ready to rethink how this work gets done. Recent moves by major health systems to insource revenue cycle functions rather than outsource to traditional RCM vendors signal a broader dissatisfaction with a model that has prioritized volume over intelligence. The question now is whether AI can change the underlying economics, not just make existing processes marginally faster.
What health systems should be asking
The agentic AI category is still early enough that the gap between genuine capability and marketing language is wide. Health system leaders evaluating these technologies should push past the ‘agentic’ label and ask what the system actually does with messy data. Can it operate on raw claims without requiring clean structured inputs? Can it detect underpayment patterns rather than just execute predefined rules? Can it run alongside existing EMR infrastructure without a deep integration project? Those questions separate genuine capability from marketing positioning.
Revenue cycle automation has promised more than it has delivered for a long time. The architectural limitations were real, and the industry has been slow to name them. While agentic AI doesn’t solve every RCM challenge under the sun, it’s the first viable approach that addresses the right problem at the right level of the data stack. That distinction is worth understanding before the next wave of vendor pitches arrive.
About Brian Sathianathan
Brian Sathianathan is the Chief Technology Officer and co-founder at Iterate.ai. The company builds private AI solutions for enterprises across industries, including Generate for Healthcare. Previously, Sathianathan worked in product and emerging technology at Apple.
