AI in Healthcare Administration: Where the Savings Go

AI in healthcare administration is sold as the easy case. No diagnosis, no patient safety board, no regulator with a device pathway. Just forms, codes and claims, the part of medicine nobody defends. So it is worth stating what happened when the sector actually did it.

In January 2026 the Peterson Health Technology Institute put senior people from health systems, health plans, technology developers, investment firms and federal agencies in one room under the Chatham House Rule and asked what administrative AI had produced. The published conclusion is unusually blunt: as currently deployed, AI in healthcare administrative processes is likely to achieve only some of its goals, such as reducing manual effort, while simultaneously increasing healthcare costs. Applied on top of flawed workflows and incentive structures, it exacerbates the underlying issues.

The prize is not in dispute. Of roughly 350 billion dollars in annual US administrative waste, the same report attributes 266 billion to administrative complexity and 59 to 84 billion to fraud and abuse. What is in dispute is who ends up holding the saving, and the answer so far is nobody.

The task gets cheaper and the system gets more expensive

Three questions get collapsed into one whenever this topic comes up. Does the tool reduce manual effort? Does it reduce cost per claim? Does it reduce total spend? They have different answers, and only the first is currently yes.

The mechanism for the third is documentation. Better capture of a visit supports coding it at higher complexity, and higher complexity pays more. One multihospital health system deploying an ambient scribe across ambulatory and emergency settings recorded a 5 percent increase in Level 5 encounters and a 7 percent increase in Level 4 encounters for established patients, worth an average revenue increase of 1,004 dollars per provider per month. Nothing there is fraudulent. The visit may well have been that complex all along. But revenue up is cost up for whoever pays, and health plan participants described an immediate rise in coding intensity the moment the tools were switched on.

On cost per claim the honest position is that the number does not exist. There is no published evidence that faster prior authorization translates to a lower average cost per claim once the cost of the AI solution is included. Participants also reported rising transaction volumes and back-and-forth bot wars, which is the opposite of what a procurement model assumes.

Prior authorization is where the burden actually sits

If administrative AI has a headline target, this is it, and the burden it is aimed at is real. In a nationwide survey of 1,000 practising physicians, the American Medical Association found that physicians and their staff spend 13 hours a week on prior authorization for a single physician, 40 percent employ staff whose primary job is that task, 89 percent say the requirement increases burnout, and 75 percent report denials rising over five years.

The volume behind those hours is large and the review of it is thin. KFF data cited in npj Digital Medicine shows Medicare Advantage plans fully or partially denying 3.2 million of 50 million prior authorization requests in 2023, with the denied share climbing from 5.6 to 5.8 percent across 2019 to 2021 to 7.4 percent in 2022. One insurer's post-acute denial rate went from 10.0 percent in 2020 to 16.3 percent in 2021 to 22.7 percent in 2022, a period in which it was automating the process. Only one in ten denials is appealed.

That last figure is the one to hold onto, because the appeals data says the denials are frequently wrong. One study of Medicare Advantage plans found an overturn rate of nearly 82 percent, and in Affordable Care Act Marketplace plans denial rates reached 20 percent with fewer than 1 percent appealed and nearly half of those appeals reversed. A process that is wrong most of the time it is challenged, and challenged almost never, is not a process automation fixes by running faster.

Both sides bought the same tool, which is why nothing nets out

This is the part the vendor decks leave out. Administrative AI is not being deployed into a system. It is being deployed into a negotiation, by both parties.

On the payer side the adoption is near universal. The National Association of Insurance Commissioners surveyed 93 insurance companies across 16 states and found 84 percent using AI or machine learning across product lines, with individual major medical carriers reporting use or exploration in utilization management at 71 percent, prior authorization approvals at 68 percent, claims fraud detection at 50 percent and denying prior authorizations at 12 percent.

That 12 percent is smaller than the discourse suggests, and it is also not the whole mechanism. Payers are technically correct when they say AI does not deny care. What it does is decide which requests are straightforward enough to auto-approve and which get routed for additional scrutiny, and as one revenue cycle executive put it, technology determines which cases enter the denial risk pathway. The routing criteria are proprietary. So a provider chooses between loading extra clinical detail into every submission with no guarantee it changes the routing, or submitting standard documentation and accepting a higher chance of delay. Neither is a strategy.

Meanwhile plans are responding to provider-side coding gains with across-the-board downcoding, algorithms that flag outlier high-complexity codes, reduced reimbursement on selected modifiers and peer comparison against providers with similar patient populations. The effect on total cost of care is not yet known, and the reductions may fall hardest on providers that have not adopted AI at all. That is the shape of a cost incurred while every component works as designed, which is a category worth reading separately from the costs a health system pays while the tool works exactly as designed.

The automation that is genuinely working

None of this means the technology is useless here. It means the wins are narrower and more specific than the category name implies, and they share one feature: the underlying rule is deterministic.

The most valuable application is not faster authorization but no authorization. A panel convened by the Healthcare Financial Management Association heard that AI can rule out the need for authorization in about 40 percent of care episodes, which removes the transaction rather than accelerating it. Where authorization is required, the same approach pulls the relevant clinical evidence out of the record rather than making staff read medical policies that run 20 to 50 pages.

Real throughput numbers exist where the specialty is high volume and the criteria are objective. A revenue cycle company working with a national payer on touchless prior authorization, where neither side handles the request manually, reports a 92 percent success rate with responses inside an hour. That is a genuine result and it currently applies to radiology only.

On the drafting side the arithmetic is straightforward. A model-generated first draft reviewed and finalised by the treating physician can cut per-letter authoring from the 20 to 30 minutes typically reported to a review cycle of 5 to 10 minutes, roughly 7 to 14 physician hours a week across a practice averaging 43 weekly submissions. Note what that is: labour returned to the practice. It is not money returned to the system, and it is the honest version of the business case.

The limit is administrative precision, not clinical writing

Here is the finding that should reset expectations, and it is almost absent from the coverage.

When three frontier models were asked to produce prior authorization letters, the clinical prose was strong. The administrative scaffolding was not. Across 45 letters per model, CPT and HCPCS billing codes, which many payer systems require for routing and adjudication, appeared in 0 of 45 GPT-4o letters, 22 of 45 Claude Sonnet 4.5 letters and 23 of 45 Gemini 2.5 Pro letters. Authorization duration, a standard element specifying the requested coverage period, appeared in 0 of 45 for both GPT-4o and Gemini and 12 of 45 for Claude. Follow-up and monitoring plans appeared in 3, 4 and 12 letters respectively.

The explanation is in the training data. Clinical literature is abundant and payer form-filling conventions are not, so the models write persuasive medicine and omit the fields that make a submission processable. The architectural consequence is specific: a deployment-grade system needs a deterministic layer drawing on formulary databases, code repositories and payer requirement catalogues to supply what the model will not. The CMS 2024 electronic prior authorization rule, which mandates FHIR-based APIs at impacted payers, is what makes that layer buildable rather than bespoke.

The same lesson appears on the coding side. In a controlled comparison of architectures for medical coding and billing, generative-only systems were faster in raw processing but propagated systematic error in complex or ambiguous cases, while a human-in-the-loop design that escalated uncertain cases achieved higher accuracy and better audit readiness. Full automation, the authors conclude, is neither necessary nor desirable in a regulated administrative domain. Both findings point the same way, and they belong next to what the 2026 deployment evidence shows on the clinical side.

The rules arrived before the savings did

Anyone budgeting for this in a US market is now buying into a live regulatory question rather than a settled one.

At least 25 states have issued guidance based on the 2023 NAIC model bulletin as of early April 2026, establishing that consumer-facing decisions made or supported by AI must comply with existing insurance law and that regulators may inquire into validation, testing and ongoing audits. Statute has moved faster still. California's SB 1120 took effect on 1 January 2025 and reserves final medical necessity determinations to a licensed clinician competent in the specific issue. Texas SB 815 restricts payer AI to administrative support and fraud detection outright. Arizona, Maryland and Nebraska have all barred AI as the sole basis for an adverse determination, and Maryland additionally opens the tools to state inspection and audit.

Federal policy is pulling in two directions at once. CMS requires Medicare Advantage medical necessity decisions to rest on individual circumstances and be reviewed by an appropriately qualified professional, while the agency's own WISeR pilot runs AI-assisted prior authorization across roughly 20 outpatient CPT codes in traditional Medicare, six states, six years, with vendor shared savings tied to speed, overturn rates and burden reduction. At the same time a national AI policy framework proposes preemption that could nullify state consumer protections governing AI in prior authorization and claims review. Underneath all of it sits a public that has not been consulted: 77 percent say they are concerned about the privacy of personal health information given to AI tools. The governance questions this raises for an operator are set out in AI governance in healthcare.

What to ask before signing an administrative AI contract

The evidence supports a short list, and none of it is about model quality.

Does the tool remove the transaction or accelerate it? Removal compounds and acceleration invites a counter-response. Is the underlying rule deterministic, like an ejection fraction threshold, or does it hinge on whether a patient is in significant pain? The second category is where the panel expert placed the difficulty, and it is where automation stalls. What is the measured cost per claim including the licence, not the time saved per task? What happens to your coding intensity after deployment, and what has your largest payer done in response? Is there a human decision point of record that survives an audit? And what does your state require this year, given that five have legislated on it since 2025?

Every one of those questions has a different owner. Coding intensity is finance, routing asymmetry is contracting, the human decision point is compliance, the state rule is legal, and the returned physician hours are clinical operations. That is the reason the EX Future Summit runs Health Tech alongside Finance and Government rather than in a separate room, and it runs 18 to 20 November 2026 as a single continuous thirty-hour broadcast between Las Palmas and Bali, twelve hours apart, with online attendance free for verified researchers and students. Before any of it is budgeted, it is worth knowing what the tool costs once it is running.

FAQ

Does AI actually reduce administrative costs in healthcare?

It reduces manual effort on specific tasks, and that part is measurable. It has not been shown to reduce the average cost per claim once the price of the software is included, and the Peterson workshop's conclusion was that as currently deployed it increases total healthcare costs. Treat labour returned and money returned as separate line items, because so far only the first one has evidence behind it.

Can a health insurer legally use AI to deny a claim?

It depends where. CMS requires Medicare Advantage medical necessity denials to consider the individual's circumstances and to be reviewed by a professional with relevant expertise. California, Texas, Arizona, Maryland and Nebraska have each legislated against AI as the sole basis for an adverse determination, with Texas going furthest by confining payer AI to administrative support and fraud detection. A pending federal preemption push could unsettle those protections, so the answer may not be stable through 2027.

What is the CMS WISeR model?

Wasteful and Inappropriate Service Reduction, a six-year pilot in six states applying AI-assisted prior authorization to roughly 20 outpatient CPT codes in traditional Medicare, a programme where prior authorization historically did not apply. Participating technology companies earn shared savings against metrics including decision speed, accuracy as measured by overturned denials, and reduction in provider burden. It has drawn objection from physician groups and companion bills in both chambers aimed at stopping it.

Does AI-assisted coding count as upcoding?

Not on its own. If the documentation genuinely supports a higher complexity level, the code is correct and the payment is owed, which is why plans are responding with downcoding policies rather than fraud enforcement. The unresolved question is whether the richer note reflects the encounter or reflects the tool, and neither side currently has the claims data to settle it.

Do providers know which prior authorization requests a payer's AI will flag?

No. Payers generally treat the routing logic as proprietary, so a practice cannot tell in advance which submissions will auto-approve and which will be pushed into review. That information asymmetry is why provider-side investment in more thorough documentation is a gamble rather than an optimisation, and why transparency demands from provider organisations centre on approval, denial and overturn rates broken out by whether a request was routed.