When Your AI Agent Goes Slowly Crazy: What Revenue Cycle Leaders Need to Know Right Now
There is a phrase that should be hanging on the wall of every revenue cycle director in America right now. It came from Jeff Gautney, CIO of Rush University Medical Center, during a panel at Becker's 16th Annual Meeting this spring.
"An interface breaks, it's binary, it's on or off. AI agents go slowly crazy usually."
Rush is about 18 months into its agentic AI journey. They are not a cautious, wait-and-see organization. They moved early, they moved deliberately, and they still ran into something nobody warned them about: the quiet drift. Not the catastrophic failure that sets off alarms. The slow, barely-visible deviation that operations leaders do not notice until the damage is already done.
That is the story of agentic AI in the revenue cycle right now. And if you are an RCM executive evaluating your next technology investment, this is the most important thing you will read this week.
What Agentic AI Actually Is, and Why It Is Different
Most of the AI tools that health systems deployed in the first wave were single-function. Ambient documentation. Denial prediction. Eligibility verification. You turned them on, they did a job, you measured the output.
Agentic AI is different in kind, not just degree. An agent does not wait for a prompt. It runs continuously, makes decisions, takes actions, and consumes resources around the clock whether or not anyone is actively using it. It can schedule appointments, pull records, initiate outreach, flag claims, and trigger workflows, all without a human hand on the wheel.
That is the promise. The problem is what happens when the wheels start to drift.
Unlike traditional software, which fails cleanly, an agentic system can begin producing subtly wrong outputs without any obvious alarm. A claim gets filed with stale eligibility data. A patient outreach message goes to the wrong address. A prior authorization gets triggered on a procedure that no longer matches the payer's current criteria. None of these failures are loud. They accumulate quietly in your A/R, your denial rate, and your patient satisfaction scores.
Rush discovered this firsthand. Their agentic tool, deployed to surface accurate schedule information, started surfacing inaccurate schedule information. Not because the AI was broken. Because the clinic's process for updating its own hours involved an email to marketing, which emailed a website vendor, which introduced a two-to-three day lag. The agent grabbed whatever was in the system.
"What looked like a hallucination," Gautney said, "was actually bad data and bad process behind the bad data."
The Revenue Cycle Is Not Ready for This. Most Systems Are Not.
Here is the uncomfortable truth that most RCM technology conversations skip over: agentic AI does not create data quality problems. It exposes them.
Every eligibility error you have been papering over with manual workarounds. Every payer contract term that lives in someone's head instead of a structured data field. Every clinic hours update that sits in an email chain for 48 hours before it hits your scheduling system. Every charge description master entry that nobody has audited since 2021. Agentic AI finds all of it, and then it acts on it.
In RCM 2030: Strategy and Survival for Revenue Cycle Leaders, I make an argument that I now believe more strongly than when I wrote it: the revenue cycle is the hospital's financial engine, and the quality of your data is the fuel. Bad fuel does not just slow the engine. It sends it in directions you really do not want it to go.
That is not a metaphor. Gautney used almost exactly that language at Becker's. "Agentic AI immediately exposes all of the bad data in your organization and all of the bad process," he said. "If you have bad fuel, the engine stops working or goes into directions you really don't want it to go."
The Financial Risk Nobody Is Talking About
There is a second problem with agentic AI that is flying under the radar in most health system finance departments, and it is not a clinical problem. It is a cost model problem.
Traditional software runs on licensing fees. Predictable. Budgetable. Agentic AI runs on token consumption. Every action the agent takes, every query it processes, every decision it makes costs tokens. And it costs them whether anyone is using the system or not, because the system is always on.
Gautney flagged this directly: "When you turn on an agent, it's on all the time. It is consuming tokens all the time, and your pricing model very quickly — it's easy to see how the agent becomes more expensive to operate than the people that it's replacing."
Most health systems have not built the financial operations discipline to manage this. They are evaluating agentic AI the way they evaluate a software license: what does it cost per year? The right question is what does it cost per transaction, per decision, per hour of continuous operation, and how does that scale as we add more agents, more use cases, more data?
CFOs who are not asking that question right now will be asking it later, in a much less comfortable setting.
What Operations Leaders Actually Need to Do
The temptation when you read something like this is to slow down. To wait for the technology to mature. To let someone else be the cautionary tale.
That is the wrong instinct. The systems pulling ahead right now are not the ones who paused. They are the ones who moved fast and built the right oversight structures around the speed.
Here is what that looks like in practice.
First, treat your data infrastructure as a prerequisite, not a parallel track. Before you deploy an agent into any revenue cycle workflow, map the data it will consume. Where does that data come from? How often is it updated? Who owns the update process? If the answer to any of those questions is "I'm not sure," you have found your first project.
Second, assign operational ownership for AI agent performance to the operations team, not IT. This is the most important cultural shift in the Rush story and the one that most organizations are not making. IT can tell you when the system is down. Only operations can tell you when it is drifting. The people closest to the workflows are the ones who will notice first that the agent's outputs do not match what they know to be true. They need to be empowered to flag it, escalate it, and stop it.
Third, build a financial model for token consumption before you sign a contract. Know your baseline cost per claim, per eligibility check, per patient outreach attempt. Know how that scales. Know what your off-ramp looks like if the math stops working. The health systems that are renegotiating vendor contracts right now, insisting on one-year out clauses and shorter terms, are doing so in part because they understand that AI pricing models are still being figured out by everyone, including the vendors.
Fourth, invest in identity management now. As health systems deploy agents from Epic, Workday, Salesforce, UiPath, and others simultaneously, the question of which agent has access to what data becomes a governance problem with real compliance and security implications. Gautney called identity management the next essential infrastructure element for getting through what he called the "cloudy period." He is right.
The 2030 Horizon
None of this is an argument against agentic AI. It is an argument for going in with your eyes open.
By 2030, the revenue cycle functions that are still manual today will be fully automated. Eligibility will run continuously in the background. Prior authorizations will clear at scheduling. Appeals will generate themselves. The question is not whether agentic AI will be running your revenue cycle. It is whether the data, processes, and oversight structures underneath it will be solid enough to let it run well.
The organizations that get there first will not be the ones with the biggest technology budgets. They will be the ones that treated data quality and process integrity as the foundation, not the afterthought.
The gate is already wide open. The question now is what you are putting inside it... because it’s way too late to close it.
Want to know where your revenue cycle actually stands on AI readiness? The RCM AI Readiness Scorecard walks you through five domains: denial prevention, coding and documentation, patient financial engagement, cash forecasting, and governance.
