R1 at HFMA: AI in Revenue Cycle Is a Routing Problem, Not an Automation Problem

Lee Kupferman, co-CEO of R1's innovation lab, stepped to the mic at HFMA Annual 2026 and said something most AI vendors would never say out loud: the technology didn't deliver what everyone promised two years ago. That kind of honesty is worth paying attention to — because the framework he offered in its place is one of the most useful things I've heard at a conference in years.

The Hype Hangover

Cast your mind back to 2023 and 2024. The investor thesis was simple: large language models were going to automate clinical coding within twelve months. It was treated less like a prediction and more like a scheduled delivery. Vendors positioned accordingly. Health systems started penciling in headcount reductions. The future had arrived — it was just waiting on implementation.

It didn't happen. Not on that timeline, not at that scale. And at HFMA 2026, Kupferman said so plainly.

This matters. Healthcare has a long history of hype cycles — EHRs were going to eliminate administrative burden, robotic process automation was going to end denials, and machine learning was going to make prior authorization painless. Each wave brought real progress alongside real disappointment. AI in RCM is no different. The question isn't whether the technology works at all — it demonstrably does in specific contexts. The question is whether we're being honest about where those contexts end.

Where AI Is Actually Winning Right Now

Here's the Kupferman framework, distilled: AI works best on high-volume, high-consensus work.

Think about a simple inpatient encounter — a straightforward pneumonia admission, a routine hip replacement, an uncomplicated delivery. If you put fifty experienced coders in a room and gave them the same chart, they'd all come out with the same answer. There's no gray area. The documentation is clean, the coding guidelines are unambiguous, and the payer rules are consistent.

That's where you let AI run. Autonomously, at scale, without a human in the loop for every single encounter. The ROI is real, the error rates are manageable, and you free your most experienced staff to work on cases that actually require judgment.

This isn't a small opportunity. High-volume, routine cases often represent 60–70% of a health system's total claim volume. Getting AI to handle that block reliably creates significant capacity — both in cost and in human attention — that can be redirected where it matters most.

Where AI Still Falls Short

The honest counterpart to that insight is this: as soon as you move away from clean, consensus cases, AI performance degrades — sometimes sharply.

Complex encounters are the problem. Longer documentation. Conflicting clinical notes. Encounters that span multiple conditions, comorbidities, or procedure combinations. Cases where payer rules diverge meaningfully from one plan to the next. Kupferman was direct: most AI models still struggle here.

That's not a knock on the technology — it's an accurate description of a hard problem. Clinical documentation is messy. It was written for clinicians communicating with other clinicians, not for automated parsing systems. And payer contracts, with their hundreds of carve-outs, modifiers, and plan-specific rules, represent exactly the kind of structured complexity that current LLMs handle inconsistently.

The risk isn't that AI gets these cases wrong occasionally. The risk is that AI gets them wrong confidently — without surfacing the uncertainty that would prompt a human review. That's where the guardrails Kupferman referenced become non-negotiable. You need systems that know what they don't know.

The Real Problem: Routing

So if AI crushes simple, high-volume work and still struggles with complex encounters, the operational challenge shifts. It's no longer a question of whether to use AI — it's a question of which work goes to AI and which work goes to humans.

That's a routing problem. And most RCM organizations aren't built for it.

The mental model most teams have is binary: either AI does the coding, or a coder does the coding. What's actually needed is a triage layer — something that evaluates each encounter for complexity, confidence, and payer-specific risk, then routes accordingly. Simple case? AI handles it. Documentation ambiguity, unusual payer rules, or high-dollar encounter? It goes to your best coder.

Getting that routing right is where the real work is. It's not glamorous. It doesn't make for good vendor slide decks. But it's the difference between an AI deployment that saves money and one that generates denials you spend months chasing.

The Fragmentation Problem That's Blocking Everything Else

Kupferman's second major point was harder to hear — because it's a structural problem that no single AI tool can fix.

Healthcare's revenue cycle is deeply fragmented. Not just somewhat fragmented — hundreds of narrow point solutions that don't communicate with each other. Your coding team operates in near-total isolation from your prior authorization team. Your denial management workflow has no real-time feed from your eligibility verification system. Your contract management tools don't talk to your claim scrubbing engine.

The consequence of that isolation is predictable and painful: denials that could have been caught at point of care trigger weeks of rework downstream. A prior auth that should have flagged at scheduling becomes a denial at claim adjudication becomes a write-off at 180 days. The inefficiency compounds because the data that could prevent it is siloed in a different system, owned by a different team, with no integration between them.

This is, in Kupferman's telling, the single biggest obstacle to AI delivering on its promises in RCM. It's not model quality. It's not compute cost. It's that the tools aren't connected, so efficiency gains in one silo don't propagate through the system.

What This Means for Your RCM Team Today

If you're responsible for revenue cycle operations, Kupferman's framework suggests a few concrete things:

First, do your own complexity analysis. What percentage of your claim volume is genuinely high-consensus? If you haven't segmented your encounter types by coding complexity and documentation quality, you don't have a baseline for evaluating what AI can absorb. That analysis is the prerequisite for a rational deployment strategy.

Second, build the routing layer before you expand AI scope. A lot of RCM teams are trying to increase AI automation percentages without first investing in the triage logic that determines what should be automated. That's backwards. The routing intelligence is what makes higher automation percentages safe.

Third, audit your integration gaps. Where is your coding team flying blind because they don't have visibility into prior auth, eligibility, or denial history? Those gaps are multiplying the downstream cost of every error. They're also the places where connecting your tools — even imperfectly — could generate the fastest returns.

Fourth, hold your vendors accountable for honest benchmarks. Any AI vendor who can't tell you exactly what encounter types their model performs well on — and where confidence scores drop — is selling you a vision, not a product. Demand the specifics. Where does the model's accuracy degrade? What's the confidence threshold for autonomous processing vs. human review? What happens at the edge cases?

The Collaborative Window Is Real — Don't Waste It

There was one more signal from Kupferman worth flagging. He noted something that would have been remarkable two years ago: payers and providers are now showing genuine willingness to work together on the payments process.

His framing was sharp: "Everybody is in violent agreement about what the problem is — they're just trying to figure out the best way to solve it."

That's a significant statement. For most of the last decade, payer-provider dynamics in RCM have been adversarial by design. Payers optimized for denial rates. Providers optimized for appeals. Both sides built increasingly sophisticated systems to beat the other. The administrative cost of that arms race — conservatively estimated in the tens of billions annually — is a drag on everyone.

If there's genuine momentum toward collaboration, that changes the calculus for how AI gets deployed. Tools built for an adversarial environment look different from tools built for a cooperative one. Prior auth workflows designed around fighting denials look different from workflows designed around preventing them upstream through shared data.

The alignment on the problem — the fragmentation, the denials, the administrative burden — is real. What's still being figured out is the path from diagnosis to fix. RCM leaders who engage in that process now, while the collaborative window is open, are better positioned than those who wait for the solution to arrive pre-packaged.

Kupferman's message at HFMA wasn't pessimistic. It was calibrating. AI in revenue cycle is real, it's working, and it's going to work better over time. But the teams that will extract the most value from it are the ones who've done the unglamorous work: segmenting their claims, building the routing logic, closing the integration gaps, and demanding honest benchmarks from every vendor in the stack.

The automation didn't arrive on the promised timeline. The opportunity is still very much here.