What health systems need to get right before agentic AI can run the revenue cycle

By Noah Breslow, CEO, Revecore 

Agentic AI is generating more attention in revenue cycle circles than any technology in years, and for good reason. McKinsey's latest research on the topic projects that AI enablement could cut a health system's cost to collect by 30 to 60 percent, a number large enough to reshape how CFOs think about revenue cycle investment for the next decade. The industry has settled on a term for where this is heading: a touchless revenue cycle, one where claims move from charge to cash with minimal human intervention.

That vision is real, but it is still a long way off for most organizations. The revenue cycle carries decades of payer-specific rules, fragmented data, and institutional knowledge that lives in people's heads rather than in documented workflows. While agentic AI can improve reimbursement and reduce administrative burden, whether it actually does comes down to how an organization approaches implementation, long before the first agent goes live.

Why the revenue cycle resists easy automation

The revenue cycle is one of the most rules-heavy, exception-riddled processes inside a health system. Every payer runs its own prior authorization requirements, coding edits, and appeal timelines, and those rules shift on a rolling basis, often with little notice. McKinsey puts the average claim denial rate at nearly 20 percent, and finds that as many as 60 percent of denied claims are never appealed at all — work that simply falls through the cracks of a manual process.

A lot of what keeps that work moving today lives outside any documented workflow. Experienced billing staff know which payers require a phone call instead of a portal submission, which modifiers a specific plan tends to reject, and which documentation gaps trigger an automatic denial. That knowledge rarely sits in a system, and agentic AI cannot learn what it has never been shown.

The groundwork that has to happen first

To realize the benefits of agentic AI, organizations will need to do some heavy lifting that has nothing to do with AI at all. Decision-making that has always lived in an experienced biller's head needs to get written down: which payer gets escalated and when, which denial codes warrant an appeal versus a write-off, which documentation gaps get routed back to coding versus clinical staff. Agentic AI can only replicate a process once it can actually see it.

Data quality is equally important. An agent trained on incomplete or inconsistent claims data will make confident, incorrect decisions at scale, which can lead to a worse outcome than a slower manual process. And before deploying anything, organizations need a clear-eyed view of where their bottlenecks form. A denials queue that backs up because of a specific payer's turnaround time calls for a different fix than one that backs up because documentation keeps arriving incomplete.

How to measure whether it's working

Revenue cycle leaders should track a small set of measures that actually reflect what agentic AI is supposed to change. Reductions in manual touches per claim show whether staff time is genuinely freed up rather than simply redirected. Turnaround time on denials and underpayments shows whether work is moving faster, not just differently. Reimbursement outcomes, including denial overturn rates, days in accounts receivable, and net revenue realized, show whether the effort is translating into more money collected rather than activity completed.

HFMA's February 2026 survey of revenue cycle leaders found that only about 7 percent describe their teams as very prepared for the revenue cycle of the future, with 27 percent deploying AI at scale and 53 percent still running pilots in select areas. That gap between activity and readiness is exactly why measurement discipline matters. It is easy to launch a pilot, but often hard to prove it is paying off.

From what’s real today to touchless tomorrow

Most initial agentic AI deployments in healthcare today come from third-party vendors working specific parts of the revenue cycle, not from health systems running integrated agents across the full life of a patient visit or a claim. A touchless revenue cycle, one with minimal human involvement across scheduling, documentation, coding, billing, and collections, is still years away for most organizations.

But that doesn’t mean health systems should wait to get started. Organizations holding out for a mature end-to-end solution are the ones most likely to still be manually working claims a few years from now, while the rest of the industry moves on without them. Denials management and underpayment identification are already common, well-scoped starting points, and organizations that prove out a result there before expanding are the ones building real momentum.

While agentic AI is not going to run the revenue cycle by itself in the next year or two, what it can do right now, on a well-defined set of high-volume back-end tasks, is real. And the health systems treating it as a discipline built on documented processes, clean data, and clear measurement will capture its full value — today and tomorrow.