Where AI agents fit in health system operations right now
There’s a lot of noise about AI agents in healthcare right now. Some of it is real. A lot of it is ahead of what the tools can safely do inside a hospital.
Here’s my plain view of where they fit today.
By an AI agent, I mean software that can read information, take a series of steps, and produce a piece of work. A drafted letter, a filled-in form, a sorted queue, a flagged exception. The useful ones today do the prep work. A person still makes the call.
That setup, where the agent prepares and a person approves, is where I’d start every conversation.
Good places to start
The best early uses share a few traits. The work repeats. The inputs are mostly documents and system data. The output is something a trained person can check quickly. And a mistake gets caught before it reaches a patient or a payer.
That points to the back office.
Prior authorization prep. Pulling the right clinical notes, matching them to a payer’s criteria, and assembling the packet takes staff a lot of time. An agent can draft the packet and flag what’s missing. A person reviews it and submits. This area is changing on the payer side too. Since January, the CMS prior authorization rule has required Medicare Advantage, Medicaid and federal-exchange plans to decide urgent requests within 72 hours and standard ones within 7 days, and to give a specific reason for denials. Electronic prior authorization APIs are due by 2027. As payers build those, the prep work on your side becomes more structured, and a better fit for this kind of help.
Denials. An agent can read a denial, pull the related records, sort it by likely root cause, and draft an appeal letter for the cases worth fighting. Your appeals staff spend their time reviewing and deciding instead of gathering.
Coding queues. Agents can pre-review charts, suggest codes, and flag documentation gaps for a coder to confirm. Coders stay in charge. The queue moves faster, and the hard cases get more of their attention.
Supply chain. Matching invoices to contracts, catching price differences, and finding spend that should have gone to a contracted item is tedious, rule-heavy work. Agents handle that kind of work well. A buyer approves any change.
Where I’d wait
I’d hold off on anything where the agent acts directly on a patient’s care, or sends something binding to a payer or a patient without a person reviewing it first. These tools make mistakes, and they can make them with confidence. In the back office, a reviewer catches those. In clinical decisions or patient messages, a miss costs far more.
The most public test of AI in prior authorization this year is on the payer side, and it shows the risk. WISeR is a pilot that adds AI-assisted prior authorization review to traditional Medicare. CMS records published by EFF, an advocacy group, show two vendors denied 5,944 requests in the first three months, and one request went unanswered for 83 days. Whatever you think of the program, the lesson for your own agents is the same. Speed and volume are not the goal. Accurate work, with a person accountable for each decision, is.
I’d also be careful with any pitch that promises to replace a team. The realistic gain today is that the same team handles more volume, with less time spent gathering and more spent on judgment. That’s a real benefit. It’s also a different business case than cutting headcount, and you should plan your budget that way.
How to bring one in
Treat an agent like a new hire on probation.
Start with one narrow task in one department. Define what good output looks like. Have a person review every piece of work at first, and track how often they change it. That edit rate is the best single sign of whether the agent is ready for more. When reviewers stop changing much, you can start sampling instead of checking everything.
There’s now an outside yardstick for this kind of oversight. In June, the Joint Commission launched a voluntary certification for responsible use of AI. It looks at how you govern, monitor and train people on AI, not at any single tool. Even if you never apply, what it looks at makes a good checklist.
Keep a record of what the agent did and why. If an auditor or a payer asks how something was decided, you need to be able to show them.
Watch what the agent can see. Agents need to read a lot to be useful. Give each one access to what its task needs and nothing more, and have your security and compliance teams review it the way they would any new system that touches patient data.
Then measure it like any other change to operations. Time per item, items handled per day, errors, and dollars. Run it alongside the current process for a while so you have a fair comparison. If it doesn’t beat the old way on those numbers, it isn’t ready, no matter how good the demo looked.
My suggestion is to pick one back-office queue where your staff spend most of their time gathering information rather than deciding. Prior auth packets and denial appeals are good candidates. Pilot an agent there with a person approving every output, and judge it after 60 days on time saved and errors caught.
That’s a modest start. It’s also the kind that turns into real capacity, without betting patient trust on a tool that isn’t ready yet.
Discussion