Author:
Hospitals lose an estimated 3% to 5% of revenue to leakage each year, according to HFMA. Healthcare doesn't have an AI shortage: vendor after vendor claims more capable models, more automation and a growing number of AI agents operating across clinical, administrative and financial work. Yet denials and underpayments persist and grow. Health system employees work valiantly to get their “piece” of the cycle right, but no one has full visibility of the whole system. Revenue leaks through those gaps, often discovered only after a claim has stalled or been denied.
We have spent years improving individual tasks without always examining the full process those tasks serve. Walking around HLTH, you’ll notice that access to AI technology is no longer the limiting factor; how work moves through the health system is.
A faster task can still leave a broken process
Consider an AI tool that flags a likely authorization or coding issue within seconds. What happens next determines whether that speed changes the outcome: the information must reach the right person, carry enough clinical and financial context to be useful, and move into the next system. The technology has done its job by identifying the issue. Without those next steps, however, it has accelerated detection while leaving the bottleneck downstream.
A documentation gap becomes a coding question. The coding question delays a bill. The delay creates follow-up work, and the missing context may not become fully apparent until a denial or appeal—far from where the issue began.
Process intelligence reveals how work really moves
Traditional process maps show the intended path. Process intelligence exposes the path that work actually takes, including detours, repeated exceptions, missing information and unnecessary handoffs.
Much of this knowledge already exists inside healthcare organizations. Experienced coders recognize documentation patterns that create rework. Authorization teams know which cases are likely to stall. Denial specialists can often see recurring failure points before they appear in enterprise reporting. The challenge is turning that experience into intelligence the organization can apply consistently.
Connected workflows need a common operating system
A revenue operating system provides a common environment for data, workflow, decisions and governance. It allows each workflow across the cycle to operate with the same context and gives the organization a way to capture what happened after a recommendation or action.
That feedback matters. If a documentation prompt prevented a denial, the system should retain that lesson. If an automated action created rework, leaders should be able to trace the result to its source and adjust the workflow. Each deployment should leave the organization smarter about what worked, what did not and what should happen when the same situation appears again.
Rethink the work before scaling the technology
Healthcare leaders can start by examining one end-to-end workflow, such as prior authorization or denial prevention, rather than the performance of a single tool. Trace it from the first input to the final outcome. Where does information change hands? Where do teams recreate context? Which exceptions return week after week? Name an owner for the full workflow and establish a baseline before changing it. That is a better test of AI's value than a task that was already predictable and highly automated.
The human role also needs to be designed upfront. Some situations call for an AI recommendation. Others support an automated action within clear guardrails. Complex or high-risk exceptions require experienced people who can understand context the technology may not have. Human oversight should be part of the workflow design, not a control added after deployment. Among respondents familiar with agentic AI in new IDC research sponsored by R1, 38.5%* reported difficulty determining when humans versus AI should act—underscoring the need for defined guardrails and human-oversight design.
Measurement must follow the same end-to-end view. Task speed matters, but so do denial rates, revenue yield, cash acceleration, employee effort, rework and cost to collect. Improving one measure while shifting work or risk elsewhere is not system improvement.
The real test of AI in the revenue cycle isn't how many tasks we can automate. It's whether the system gets better at preventing the problems that create those tasks.
That's the standard Phare OS is built to meet. See how it connects data, intelligence and work across the revenue cycle- https://www.r1rcm.com/phare-os
*IDC White Paper, sponsored by R1, The Revenue Operating System: From Automation to Orchestration, doc # US54991926, November 2026