The first hospital function AI reaches is often the one whose value the hospital never learned to measure

Twelve nurses at Montefiore Medical Center in the Bronx were dismissed after receiving notices that their positions would be eliminated. They worked in utilization review, examining medical records and explaining to insurers why an admission, a medication change, or a specific care plan met medical-necessity criteria and should be covered. According to the nurses’ association, part of that work was transferred to AI-enabled software; Montefiore described the change as a nonclinical documentation program and said the union’s account misrepresented the project.

What is most intriguing is the function that was chosen.

For years, the public debate focused on whether AI would replace the physician who diagnoses, prescribes, or makes decisions at the bedside. Yet one of the first highly visible labor disputes involving AI inside a hospital emerged in a function classified as administrative: inpatient authorization and the review of insurance denials.

That classification deserves scrutiny because it comes from the place where the expense is recorded, while the work itself combines clinical interpretation, longitudinal review of the medical record, knowledge of contractual criteria, recognition of exceptions, and the ability to build an argument that protects continuity of care. One of the nurses affected had worked at Montefiore for 39 years, a span that represents accumulated knowledge about denial patterns, appeal language, insufficient documentation, and the situations in which an apparently objective rule must be challenged by the patient’s actual condition.

When that knowledge remains inside people’s experience instead of being structured and documented, the financial system sees a clearly defined payroll expense and leaves the value produced by the team scattered across the institution. A reversed denial appears in the insurer’s system; an avoided claim rejection appears in revenue cycle; a prevented delay appears in bed management; a preserved admission appears in the medical record; recovered revenue appears months later in financial reports. None of those systems easily attributes the outcome to the nurse who read the case, recognized the exception, and challenged the decision.

The economic case for automation is incomplete from the outset when it compares the full cost of a team with only a fraction of the value that team creates.

What is less obvious is that the first function reached by AI often combines two conditions: concentrated cost and poorly attributed institutional value. The more fragmented the results of that work are across the medical record, billing, audit, bed management, and payer relations, the less able the institution is to defend the function when cost reduction becomes the dominant criterion.

The selected function, therefore, is not always the easiest one to replace. It is often the one whose absence looks inexpensive before the consequences appear.

Automation has a legitimate role in this process when it locates documents, extracts contractual criteria, checks missing information, compares records, and organizes the evidence needed to support an authorization request. These tasks consume time, follow identifiable patterns, and allow objective verification. The risk rises when the system also closes the review, interprets exceptions, determines which facts deserve greater weight, and defends the decision before the insurer without a defined clinical owner, a record of disagreement, sample-based review, and follow-up on denials after implementation.

The distinction between support and replacement is not found in the software’s commercial description or in the project’s administrative label. It lies in the authority granted to the system within the workflow and in the institution’s ability to reconstruct each decision when the outcome is challenged. Which document was used, which rule was in force, which exception was considered, who reviewed the recommendation, which decision prevailed, and what happened to the patient afterward form the minimum record required for accountable automation.

In June, employers in the United States announced 45,849 job cuts, of which 14,029, roughly 31%, cited artificial intelligence as the stated reason. It was the fourth consecutive month in which AI led the explanations given by companies. The same report recorded 2,761 cuts in healthcare and health-related products during the month.

The figure does not show whether each substitution delivered the promised result, but it does show that AI is already influencing concrete workforce decisions. Inside hospitals, that pressure finds favorable conditions in activities with high volume, visible cost, and consequences dispersed across indicators that rarely communicate with one another.

Before removing professionals from a process, hospital leadership needs to reconstruct the institutional value of that function and establish a baseline that tracks denial rates, reversals, claim rejections, authorization time, prolonged length of stay caused by coverage delays, physician rework, recovered revenue, and cases escalated for human review. Without that measurement, the savings appear in the month of the dismissal, while the consequences move through different parts of the organization and lose any visible connection to the decision that produced them.

The question for hospital leaders comes before the purchase of any tool: which function inside the institution is making clinical decisions under an administrative classification and, for that reason, appears in full on the budget while remaining fragmented across the outcomes it sustains?