Healthcare administration is the AI use case with the clearest numbers

Most talk about AI in healthcare centers on diagnosis and clinical decision support, where evidence standards are high and the stakes are patient safety. The administrative side of a practice or health system is different: the work is repetitive, rule-bound within each organization, and already measured. That makes it the more tractable place to start, and this piece deliberately stays on the non-clinical side.

How big the administrative problem is

The CAQH Index, the industry's annual benchmark of administrative transactions, was published on February 19, 2026, covering 2024. It estimates that electronic transactions and improved data exchange avoided about $258 billion in administrative cost in 2024 and that a further $21 billion could be saved by fully automating transactions that are still manual or partially manual ^[1]. Medical administrative spend fell 9% and dental 4% ^[1].

The remaining $21 billion is the part that automation, including AI-assisted tools, can still reach. It is small relative to the $258 billion already captured, which is the honest framing: much of the easy electronic work is done, and what is left is the awkward, variable, document-heavy residue.

Prior authorization: the burden in practice

The American Medical Association's survey of 1,000 physicians, reported in April 2025, found practices completing an average of 39 prior authorization requests per physician per week, consuming 13 hours of physician and staff time combined, and 40% of physicians said they have staff who work exclusively on prior authorizations ^[2].

Source: AMA ^[2]. The article reports no AI findings, so these numbers describe the baseline rather than any tool's effect.

Why turnaround matters

Prior authorization has a clock on both sides: practices want fast answers so care is not delayed, and payers need complete requests to decide. For provider organizations, the bottleneck is often preparing complete requests and tracking them to a decision, which is where administrative tooling can help.

Where AI fits, and where it does not

CAQH reports that more than half of health plans and 25% of provider organizations use AI tools in administrative workflows ^[1]. The gap suggests payers are automating their side of the exchange faster than providers are automating theirs.

Task types that suit AI assistance in an administrative setting, because the work is document-heavy and reviewable, include:

• Assembling request packets: pulling the relevant fields and documents together for a human to check before submission.

• Reading payer correspondence: extracting the denial reason, deadline and requested information from letters and portal messages.

• Tracking status: flagging requests that have gone too long without a response.

• Drafting appeal text: producing a first draft from the denial reason and chart references, to be edited by staff.

Task types that need caution are those where an error changes someone's care or coverage. A model should not make the decision on a request or an appeal; it should prepare material for the person who does.

A sensible starting sequence

1. Measure first. Count requests, touches per request and turnaround time for one payer or one service line. The AMA figures ^[2] give you a benchmark to compare against.

2. Pick one document-bound step. Reading denial letters or assembling packets is easier to check than anything involving judgment.

3. Keep a person on every outbound submission. Review is cheap relative to a wrong submission.

4. Track error and rework rates, not just time saved, and compare against the baseline you recorded.

5. Prefer tools that can use structured payer responses as payers make them available, rather than depending on screen-scraping portals.

For a practice or health system, the practical implication is that administrative AI should be sized against a measured baseline of requests and hours, kept behind human review, and scoped so that no tool decides coverage or care.