Scope, not category: healthcare AI ROI at mid-market in 2026

"We fax all the time. It's insane." That is a former United Healthcare executive talking to STAT News on 27 April 2026, on a prior-authorisation workflow the rest of the economy retired three decades ago. It is also a fair description of the procurement gap that hospital CFOs are working through this year. Real money is landing from healthcare AI in 2026. It is landing in narrower places than the demo decks suggest, on a tighter timeline than the regulatory environment will allow, and at health systems whose data engineering reality looks nothing like the 200-bed regional hospital the typical reader of this post runs.

This post is written for a specific buyer. A hospital COO, CFO, or CMIO (chief medical information officer) at a 200-800 bed US system or an NHS Trust at equivalent scale, looking at three or four AI proposals on the desk, asked to defend a procurement decision to a board that has read the Becker's Hospital Review headlines and wants the same numbers. The Becker's numbers are real. They are also miscalibrated for the buyer reading this.

The dollars are real, but they live in narrow scope

Three reference points anchor the 2026 AI-ROI conversation in US hospital finance. CommonSpirit Health published "north of $100 million" in annual savings from AI and robotic-process automation. UNC Health attributes $6 million to ambient scribing, $5 million to AI infusion scheduling, $3-4 million to automated prior auth, and $5.4 million to AI-driven nursing retention. Highmark Health reports $27.9 million in value from a Google Cloud AI assistant in 2025. The Becker's Hospital Review roundup dated 1 December 2025 has the numbers in one place.

Underneath those headlines, the cleanest annual baseline is ONC Data Brief No. 80, published September 2025 (ONC is the Office of the National Coordinator for Health IT). US hospitals' use of AI to "automate billing procedures" jumped from 36% to 61% in one year. "Facilitate scheduling" rose from 51% to 67%. Predictive AI integrated with the EHR (electronic health record) sat at 71% adoption. Administrative use cases are the fastest-growing category and the most-deployed category. The dollar wins reflect where AI is actually being deployed.

What the headlines do not say plainly is the size of the operator behind each number. CommonSpirit is a 140-hospital system. UNC Health is a 16-hospital academic-medical-centre network with an internal data engineering team. Highmark is a $27 billion integrated payer-provider. Intuition Labs' August 2025 analysis of US hospital AI adoption places small independent hospitals at 31-37% AI usage and small hospitals under 100 beds at 53-59%, against 81-86% at multi-hospital systems. The published ROI lives in the top tier.

Narrow scope, not categorical

The temptation at this point is to draw the line between operational AI and clinical AI, then recommend the operational side. That line does not hold under the 2026 evidence.

The radiology AI ROI numbers are stronger than the operational ones if you actually read the studies. The Journal of the American College of Radiology (JACR, March 2024) found that an AI platform in hospital radiology workflow delivered 451% ROI over five years, rising to 791% when radiologist time savings were counted. The Lancet Digital Health published a December 2025 evaluation of Brainomix 360 Stroke across 107 NHS hospitals: thrombectomy rates doubled from 2.3% to 4.6%, with estimated annual savings of £30 million if scaled to all of England. The AMA's 2026 CPT (Current Procedural Terminology) code set moved three clinical-AI applications (CT-FFR, AI diabetic retinopathy detection, AI coronary plaque assessment) from Category III into Category I, with the new CPT 75577 for AI coronary plaque paying $950.50 per case under the 2026 Hospital OPPS Final Rule. That is real reimbursement, on the clinical side, in 2026.

The operational side has its own failures. Olive AI shut down in October 2023, and Olive was an operational, admin-AI play. Epic's first Sepsis Predictive Model, despite being clinical, is the canonical pre-2024 "AI fails in production" story: JAMIA Open (Journal of the American Medical Informatics Association, November 2024) published an external validation showing 14.7% sensitivity in the six-hour window, median advance warning lead time of zero minutes. The split that matters is not operational versus clinical. It is narrow scope tied to a measurable line item versus broad-platform AI looking for a use case. Olive was a broad-platform operational play. Epic Sepsis v1 was a broad-platform clinical play. Both failed.

The argument the rest of this post makes is plain. Match the scope of what you buy to the line item it will move, and price three frictions that the showcase deployments hide.

Three frictions to price before procurement

1. The prior-auth savings pool is contracting

Prior-auth automation has been the canonical operational-AI sell for three years. In May 2026 it is the most visibly stale piece of vendor copy on the procurement desk. CMS-0057-F, the Interoperability and Prior Authorization final rule from CMS (Centers for Medicare and Medicaid Services), makes electronic prior auth mandatory for Medicare Advantage, Medicaid, CHIP, and qualified-health-plan payers from 1 January 2026, with full FHIR-API (Fast Healthcare Interoperability Resources) implementation by 1 January 2027. The voluntary AHIP-HHS June 2025 pledge has already cut prior auths by 11% across the participating insurers by April 2026, per Healthcare Dive (AHIP is America's Health Insurance Plans; HHS is the US Department of Health and Human Services). That is 6.5 million fewer prior auths, and a 15% reduction in Medicare Advantage specifically. The American Medical Association reports at least 18 US states took legislative action on prior auth in 2025 alone. Gold-card programmes in Texas, Arkansas, and West Virginia expanded into group practices.

Stanford's HealthAdminBench, published 15 April 2026, gives a procurement-stage benchmark. Stanford researchers report frontier AI models hitting 82.8% on individual subtasks but only 36.3% on end-to-end administrative completion, with prior auth and appeals as the worst failure modes. The benchmark is recent and one team's work, so use it as directional rather than settled. An AI sold to your hospital on prior-auth automation in May 2026 is being sold a contracting volume base, a regulator-shortened decision window, and a Stanford-benchmarked 36% real-world completion ceiling.

This does not retire the operational-AI case. It moves the durable levers elsewhere: scheduling, capacity planning, RCM (revenue cycle management) denials management, and ambient documentation where the value model is retention rather than coding revenue.

2. Vendor concentration risk under Epic's bundling

At Epic's 2025 user group meeting, Epic introduced Penny, a generative-AI copilot for revenue cycle covering coding and denial appeals. Summit Health reported 42% faster medication prior auth submission and 92% of AI-generated responses accepted unedited. Those Penny numbers are real and they sit awkwardly next to Friction 1: they describe a faster lap inside a track that the rule changes are shrinking. Epic announced a no-code Agent Factory for building and orchestrating AI agents, to be showcased at HIMSS 2026 (the Healthcare Information and Management Systems Society conference). Epic is the dominant US EHR by a wide margin. If your hospital runs it, every point-vendor proposal on your desk now competes with a feature that may ship inside your existing licence in the next 12-18 months.

Bob Berbeco at Mahaska Health in Oskaloosa, Iowa, put the procurement language in Becker's in December 2025: "We stepped away from running isolated, bolt-on AI pilots that were not proven by being embedded in real clinical workflows or were likely to be duplicated by Epic's AI roadmap." That sentence is becoming standard procurement vocabulary across the systems running EHR-centred AI evaluations this year.

Olive AI's collapse is the warning, not the template. Olive scaled a broad-platform operational AI play on the assumption that hospitals would pay for an aggregating layer above the EHR. They built that layer; it duplicated workflows that were going to live inside Epic anyway; the customer base did not show up at the volume needed. The procurement question for any point-vendor proposal in 2026 is plain. What does this do that Epic's roadmap will not absorb within the licence in 18 months? If the answer is "nothing that Epic will ship faster," the vendor pricing has to account for an 18-month payback ceiling.

3. The execution-capacity gap

The single most important number for a mid-market CIO reading the headlines is the ambient-scribe ROI math. UCSF Health published in JAMA Network Open on 9 January 2026 a quasi-experimental study (Holmgren et al) across 1.2 million encounters and 1,565 physicians. Adopters generated 1.81 additional RVUs (relative value units) per week, 0.80 additional encounters per week, and $3,044 per physician per year in additional Medicare reimbursement. The accompanying invited commentary from Shah and Garcia, same issue, priced ambient-scribe subscriptions at $200-$600 per clinician per month, observing that revenue alone often does not cover the cost. The Peterson Health Technology Institute (PHTI), the most rigorous independent evaluator in the field, concluded in May 2025 that "AI scribes decrease burnout and cognitive load for clinicians but have not proven a financial return on investment" and that "the economic value proposition has yet to be proven at scale."

Scribes still work. They work as a retention lever. Shah and Garcia, in the same JAMA Network Open commentary, estimate per-physician replacement costs of $500,000 to $1 million. If a hospital can move physician turnover by a single percentage point with $300-$600 a month of ambient AI, the ROI shape is retention, not coding revenue. That is a different procurement case than "this will pay for itself in revenue uplift," which is the case most vendor decks make.

The pattern repeats across the operational stack. The mega-IDN numbers reflect what is achievable when an organisation has its own data engineering team, its own AI governance committee, and the executive cover to absorb 18 months of integration debt. Mid-market hospitals do not have those structural inputs. Daniel Barchi, the senior executive vice president and CIO at CommonSpirit, told Becker's in April 2026 that "healthcare technology is really 80% people. It's about 15% process, and it's really only about 5% technology." That ratio is harder to honour at mid-market scale, because the people-and-process budget is smaller in absolute terms.

What mid-market operators should actually buy in 2026

The narrow-scope decision framework is three questions.

First, what line item will this AI move? A category answer ("admin AI") is not a procurement case. The line items that work in 2026 are concrete: AR (accounts receivable) days, no-show rate, coding denial rate at a specific payer, OR (operating room) utilisation, physician turnover. If the vendor cannot point to a line item on your existing P&L, the procurement case has not been built yet. The same line-item discipline is what carries our earlier boring middle framing into the hospital setting.

Second, who has published evidence at your scale? KLAS (the healthcare-IT analyst firm) reported in December 2025 that 1 in 3,000 surveyed organisations was using agentic AI in production. The agentic-AI procurement case at mid-market scale in 2026 is still at the pilot stage. Evidence from CommonSpirit or Mayo is useful context, but ask the vendor to introduce you to a system in the 200-800 bed band with the same outcome before signing.

Third, what is the value model if revenue does not show up? Ambient scribes break even via retention. RCM AI breaks even on a specific payer's denial-rate movement. Capacity-planning AI breaks even on bed-day yield. A generic "AI for RCM" pitch identifies a category; a line on this quarter's denial report identifies a value model. Naming the value model before procurement is the single highest-yield discipline a mid-market buyer can adopt this year, and the one piece of work that does not change when the underlying model does.

Healthcare AI in 2026 is a scope-discipline buy. The hospitals with a real P&L story this year are the ones who picked the right line item, sized the deployment to their own data engineering capacity, and held the vendor to evidence at their scale. Pull last quarter's denial-rate report by payer before you read another vendor deck. Pick the line item first. Everything else is downstream.