Healthcare is racing to automate administrative work. But the cases that don’t fit neatly into the workflow may tell us where experienced clinicians will matter most.
By Susan McDonnell, RN, BSN | August 2026
Healthcare is rapidly automating the administrative work surrounding patient care. Billing, scheduling, insurance verification, documentation, prior authorization, appeals, and denial management are all receiving attention from AI developers—and for good reason. These processes consume enormous amounts of time and money, frustrate clinicians, and frequently delay care.
As an RN who has spent much of my career in utilization review, case management, coding, revenue cycle, prior authorization, and denial management, I understand the appeal immediately. But I think we’re asking the wrong question.
The question isn’t simply, “Can AI do this work?” Increasingly, the answer is yes. The more important question is: What happens when the case doesn’t follow the rules?
That is where healthcare gets interesting.
A straightforward authorization can be remarkably structured. Retrieve the record. Identify the diagnosis and requested service. Find the relevant documentation. Compare the clinical information against established criteria. Determine whether the requirements have been met. Document the result and move the case forward.
AI is extraordinarily well suited to portions of that workflow. It can find information buried in a medical record far faster than a human reviewer. It can summarize documentation, identify missing elements, compare information against criteria, populate forms, and draft correspondence.
The infrastructure around that work is changing as well. CMS’s current prior-authorization rule requires affected payers to comply with certain operational provisions beginning in 2026, with major API requirements generally beginning in 2027. Those APIs are designed to support electronic prior-authorization requests and responses, documentation requirements, approvals, denials, and requests for additional information.
So automation isn’t coming to this work.
It’s here.
But healthcare workflows look wonderfully orderly when they’re drawn on a PowerPoint slide.
Patients aren’t.
A patient doesn’t always have the right diagnosis documented in the right place. The physician’s note may not clearly explain why standard treatment failed. A payer’s criteria may not neatly accommodate a patient’s combination of conditions. Records may be incomplete. Something clinically important may have happened at another facility. An authorization may technically fail a criterion while an experienced reviewer immediately recognizes that the case deserves another look.
That is the difference between processing information and understanding an exception.
The Research Is Beginning to Show the Difference
A 2026 benchmark called HealthAdminBench evaluated AI agents on 135 realistic healthcare administrative tasks involving prior authorization, appeals and denial management, and durable medical equipment. One system achieved an impressive 82.8% success rate on individual subtasks. Yet the best-performing system successfully completed only 36.3% of entire end-to-end tasks.
CHI-Bench went further, testing AI agents on complex workflows involving provider prior authorization, payer utilization management, and care management. These weren’t simple question-and-answer exercises. They involved policies, multiple healthcare applications, handoffs, peer-to-peer interactions, patient outreach, and other long-horizon work.
The best-performing agent successfully resolved just 28% of complete tasks. When all tasks were executed continuously in a single session, performance fell to 3.8%.
Those numbers will improve—probably quickly.
But they reveal something important:
Doing the steps is not the same as managing the case.
And That Distinction Matters for Jobs
I don’t believe we should reassure every administrative nurse that AI will simply become a helpful assistant while staffing remains unchanged. If automation allows one experienced nurse to oversee substantially more cases, healthcare organizations may eventually need fewer people performing routine reviews.
We should be honest about that.
But I don’t believe the answer is to protect repetitive work simply because nurses currently perform it, either.
If AI can gather 200 pages of records, locate relevant clinical information, populate routine fields, and identify cases that clearly satisfy established requirements, I’m not convinced an experienced RN should spend her day doing those things manually.
Let the machine do the routine work. Then put the nurse where the machine becomes uncertain.
That may be the real opportunity.
Imagine utilization review in which AI processes straightforward cases while experienced nurses handle clinical exceptions, ambiguous criteria, complex denials, appeals, peer-to-peer preparation, quality auditing, and cases where a patient’s circumstances don’t fit neatly into an algorithm.
That isn’t simply keeping a “human in the loop.” It is a fundamentally different staffing model.
And it requires us to acknowledge an uncomfortable reality: healthcare may need fewer nurses performing routine administrative review while simultaneously placing greater value on sophisticated clinical expertise at the point of exception.
The Workforce Question We Should Be Asking Now
Healthcare leaders should be asking more than how many hours AI can save.
What work are we automating? What happens when the AI is uncertain? Who reviews its mistakes? Who recognizes when technically correct criteria produce a clinically questionable result? Who is accountable when automation affects a patient’s access to care?
And perhaps most importantly:
Are we retraining the people who already understand these workflows before we automate their current jobs?
Experienced utilization review, denial-management, revenue-cycle, and case-management nurses possess institutional knowledge that took years to develop. They know where documentation fails. They know which cases routinely become appeals. They know where handoffs break down. They know which policies look perfectly reasonable on paper and create chaos in actual practice.
That knowledge may become more valuable as routine processing becomes automated—but only if healthcare organizations recognize its value before those employees walk out the door.
I’m enthusiastic about what AI can remove from healthcare. We desperately need less administrative friction.
I’m considerably more cautious about what we remove with it.
The goal shouldn’t be preserving every task because a human has always performed it. Nor should it be automating every task simply because technology now can.
The goal should be designing a system in which machines handle what machines do exceptionally well—and experienced clinicians are deliberately positioned where judgment, context, advocacy, and accountability matter most.
Because eventually every beautifully automated healthcare workflow encounters a patient who doesn’t fit the workflow.
And that patient is the reason the exception matters.
References
- CMS — Interoperability and Prior Authorization Final Rule (CMS-0057-F): https://www.cms.gov/initiatives/burden-reduction/overview/interoperability/policies-regulations/cms-interoperability-prior-authorization-final-rule-cms-0057-f
- HealthAdminBench (2026) — Healthcare Administrative AI-Agent Benchmark: https://arxiv.org/abs/2604.09937
- CHI-Bench (2026) — End-to-End Healthcare Workflow Benchmark: https://arxiv.org/abs/2605.16679
Originally published in Aging, Care & the Fine Print, my LinkedIn newsletter exploring healthcare, aging, advocacy, technology, and the systems that shape how people receive care.
Susan McDonnell, RN, BSN
Founder, Brightway Aging Advocacy