I sit through vendor demos constantly, and the word “agentic” gets attached to almost anything with a modern interface. If you cannot tell the difference in the room, you will buy a rules engine at an AI price. Then you will spend a year wondering why your denial rate never moved.
So before another health system signs another contract, let us agree on the words.
As part of my new book “The Healthcare Revenue Cycle AI Playbook,” I outlined why you should think of artificial intelligence as the umbrella term. It covers any system built to simulate human thinking, reasoning, or decision making. Machine learning is one approach underneath it, where systems learn patterns from data instead of being told every rule; denial prediction models and fraud detection tools are the versions most revenue cycle leaders have already seen.
Generative AI produces text, summaries, and drafts when prompted. Someone asks, and it answers. In our world that looks like drafting an appeal letter or summarizing a batch of denials. It waits for a person to start it.
Agentic AI completes multistep work from end to end. It reasons through a workflow, makes decisions, adapts when something unexpected shows up, and takes the next step on its own, with a person overseeing the work rather than driving it. In the revenue cycle, that looks like eligibility verification, claim status checks, and prior authorization, carried from start to finish.
That is the whole distinction. Generative AI answers a question. Agentic AI finishes a job. Robotic process automation, by contrast, follows fixed, programmed rules. It has been running in healthcare for more than twenty years, and it does the exact same thing the exact same way every time. That works beautifully right up until something changes, and in this industry, something always changes.
Why rigid logic breaks in revenue cycle
None of this makes rules-based automation obsolete. RPA and traditional rules engines still do real work in steady, structured processes, and when a process holds still, a bot is cheaper and more predictable than anything an agent will do for you. The question worth asking before any purchase is whether the process holds still. If it does, automate it the cheap way.
But payer portals do not hold still. A field moves. A form adds a required box. A layout changes on a Tuesday with no notice, and a bot built to click in the same place every time simply breaks. I have watched a team celebrate an automation win in January and quietly go back to manual work by March, and nobody wrote that part down.
Automation with lipstick
Some vendors market slightly fancier RPA as agentic AI. I call that automation with lipstick. The front end looks modern, and underneath it is the same brittle, rules-based technology carrying every one of RPA’s failure points. It will break in the same places for the same reasons, on your dime.
The question that flushes this out every time: ask what happens when a payer changes a portal. Listen closely for who fixes it and how long that takes. If the honest answer involves a support ticket and a services team standing by, you are buying RPA with a new coat of paint.
There is a second, quieter version of the same problem. Some solutions present as AI driven while a team of people does the work by hand behind the interface. The output arrives looking automated because somebody typed it, and at demo volume that holds up fine. Then month end lands, volume spikes, and the whole model stalls, because manual labor does not scale on demand. Ask how many people touch a transaction, and ask for a straight-through rate measured end to end, in writing.
The 90/10 rule
Here is where real agentic AI earns its keep, and where I think every revenue cycle leader should set expectations. About 90 percent of a workflow can run without human intervention, including much of the exception handling inside it. Almost all of that 90 percent is repetitive clicking, data extraction, manual follow-up, and portal navigation, the work that drains hours and drains morale without requiring anyone’s clinical or coding judgment.
The remaining 10 percent belongs to your people, and it should stay there. Complex appeals that need a nuanced clinical argument. Peer-to-peer conversations with a medical director. Interpreting payer behavior that does not make sense yet. No agentic system should be handling that work alone, and no vendor should be telling you otherwise.
The technology takes the volume. Your experts take the complexity. That is the trade worth asking for in every vendor meeting you take this year, and it is the only version of “agentic AI” worth paying for.
Vanessa Moldovan is the author of The Healthcare Revenue Cycle AI Playbook and Head of RCM Strategy at Magical. To learn how health systems are using agentic AI to close care gaps and reduce administrative burden, visit getmagical.com.
The post Most of What’s Sold to You as “Agentic AI” Isn’t Really Agentic AI appeared first on Becker's Hospital Review | Healthcare News & Analysis.