AI and Healthcare Payment: The $942 Million Nobody Checked

Contents10 sections

AI follows the money

AI in healthcare does what the payment system rewards. Most American care is paid fee-for-service: each visit, test and procedure on its own. Under that system the first tools to spread are the ones that document, code and bill. Clinical AI that could change how patients do spreads slowly, because nobody is paid for the complication that never happened.

Three pieces of evidence from the last six weeks show the pattern from different sides. A research institute found administrative AI pushing billing up. An insurers' association put a price on rising hospital coding, $942 million, and the hospitals' association answered that nobody had checked a single medical record. And Medicare has started paying some providers for results instead of activity. That's the one change that would point the same tools at patients.

What PHTI found

The Peterson Health Technology Institute (PHTI), which evaluates health technology independently of the companies that sell it, updated its report Driving Adoption of High-Value AI in Healthcare on 15 October 2026. It defines high-value AI as AI that "measurably improves health outcomes, expands access to care, and lowers spending." Adoption hasn't followed it. Tools for prior authorization (an insurer's approval before care is given), coding and billing have spread fast. Under current incentives, says the report, "administrative AI entrenches inefficient workflows and drives inflationary billing." Clinical AI is held back by how care is paid for, purchased and regulated.

Anyone who has run a clinic will recognise this.

Under fee-for-service, a visit coded at a higher level earns more. A hospital stay with more recorded secondary diagnoses, the conditions listed beside the main reason for admission, can move into a higher-paying group. A tool that helps a clinician write a fuller note will, almost by construction, produce notes that support higher codes. Whether those codes are more accurate or more generous is the whole argument, and it's an argument both sides have every financial reason to keep having for as long as a fuller note and a bigger bill arrive in the same envelope.

One example in the report gives the size. It comes from a multihospital system and its AI scribe, software that listens to the visit and drafts the note. After the scribe arrived, Level 5 visits for established patients rose 5% and Level 4 visits rose 7%. Level 5 is the top evaluation-and-management billing level for an office visit, and Level 4 the one below it. Revenue increased as well. Each provider using the scribe brought in $1,004 more a month on average. PHTI cites a KLAS Research review of the scribe vendor Suki from January 2026, so the number comes from one vendor's customers.

Insurers haven't shown savings either. PHTI reports that "insurers and vendors have not yet reported evidence that AI reduces the average cost per claim after factoring in the cost of the AI solution." Both sides are spending.

The $942 million dispute

On 24 September 2026 the Blue Cross Blue Shield Association (BCBSA), which speaks for the Blue insurance plans, put a number on hospital coding intensity. Coding intensity is how often, and how far up, claims are coded toward higher-paying categories. Across inpatient claims, cases coded as medically complex went from 37% at the start of 2023 to 40% at the end of 2025. The price came to $942 million over two years. Of that, $653 million, about $11,000 per extra complex case, came from secondary diagnoses billed more often. Its sharpest example was major bowel procedures, where claims at the highest complexity level rose from 10.2% to 22.7%, more than doubling. BCBSA tied the rise to hospitals adopting AI-assisted documentation and billing tools, though its analysis didn't say which claims had passed through one.

BCBSA's PDF refused automated access. The figures here come from Fierce Healthcare's report of the release, and the hospitals' reply confirms the $942 million.

That reply came on 5 October 2026, in a blog post by Molly Smith and Jen Holloman of the American Hospital Association (AHA), and its four points deserve to be taken one at a time, since any of them could hold without the others:

• No records were reviewed. BCBSA didn't look at medical records, so it can't say whether the added diagnoses were clinically supported.
• Inpatients are sicker. Lower-acuity care has moved to outpatient settings, so the patients left in hospital are older and more complex.
• Fuller notes help care. Better documentation can capture conditions that were under-recorded before, which helps care planning.
• The effect is small. By BCBSA's own earlier numbers, it's about 0.3 percentage points of total commercial spending growth.

The AHA's best line is a test both sides could adopt: "the appropriate test for whether a patient has been correctly coded is not whether reimbursement increased." It also notes that insurers do the same in Medicare Advantage. There, federal audits have found unsupported diagnoses from some Blue plans.

Neither side has published the evidence that would settle it — independent coders reviewing a sample of records, with AI and without. Nobody has done it. The AHA didn't show the added diagnoses were supported either. Coding intensity rose; AI adoption rose at the same time; the causal link is unproven.

Both sides are buying more AI

The likely next step is more AI on each side. Providers buy tools to capture every diagnosis the record supports. Payers buy tools to find codes the record doesn't support, and to deny or down-code claims. Each side's AI is a reasonable answer to the other's. The contest (software, staff, appeals) comes out of the same premiums and taxes. None of it is aimed at whether a patient's blood pressure came down.

That's the sense in which incentives decide adoption more than capability does. The model that writes a billing-ready note could summarise a record for a nurse. It could check a medication list against kidney function, or chase an abnormal lab result nobody followed up. Some of those uses are in pilots, and they spread slowly.

Paying for outcomes

PHTI's roadmap has four parts: learn what works from real-world evidence; set rules for coverage, licensure and liability; buy differently; and redesign care around clinical AI. Buying differently means paying on outcomes and judging vendors on outcome measures. PHTI names that as the place to start.

Medicare has one running. The Centers for Medicare & Medicaid Services (CMS) launched the ACCESS model (Advancing Chronic Care with Effective, Scalable Solutions) on 5 July 2026. It's voluntary. It runs for ten years in Original Medicare and tests technology-supported care for chronic disease, starting with tracks for early cardio-kidney-metabolic conditions, cardio-kidney-metabolic conditions, musculoskeletal pain, and behavioural health (depression and anxiety). Heart failure, COPD (chronic obstructive pulmonary disease), substance use disorder and tobacco cessation join from 1 April 2027. ACCESS pays recurring Outcome-Aligned Payments. Results decide the full amount. CMS lists more than 160 participants.

A payment like that changes the question a buyer asks. "How many more Level 5 visits?" becomes "did more of our hypertension patients reach target?" A tool that documents more diagnoses earns nothing unless the patient improves. A tool that catches a missed refill does.

Three cautions apply.

ACCESS is one model, and most of Medicare is still paid per service. Outcome measures can be gamed too, by picking healthier patients or measuring what's easy. And nobody knows yet whether ACCESS participants do better; the model is less than four months old.

Two rules for MARA

MARA, the Medical Agentic Resident Assistant this project is designing, gathers a patient's record, checks it and recommends, and a person makes every decision. Two rules in her design documents follow from the evidence above.

• MARA doesn't up-code. She doesn't suggest billing codes or prompt for documentation to raise a visit level, and a diagnosis she surfaces is there because the record supports it for care.
• MARA is graded on outcomes. Her own learning is credited only when an outcome confirms it, never on activity counts such as notes drafted or minutes saved, and a miss lowers her authority.

Both rules cost something. Under fee-for-service, a tool that refuses to raise revenue is harder to sell. Outcome-linked payment is the funding model under which an assistant built to help patients can pay for itself.

What nobody knows yet

• Cause is unproven. No published record-level review shows whether AI tools caused the rise in coding intensity, contributed to it, or arrived at the same time.
• Missed illness is untested. The AHA argues the added diagnoses reflect real illness that was missed before, and BCBSA's analysis doesn't test it.
• The scribe figure is narrow. The 5% and 7% rises, and the $1,004 a month, come from one health system in one vendor's review.
• ACCESS has no results. Whether it improves outcomes or lowers spending will take years to show.

Why it matters

A hospital and an insurer can each buy a better model every year and still be fighting over the same claim. The money moves between them, and the patient's blood pressure stays where it was. Changing the model changes the speed of that contest. Changing what gets paid changes its direction. That's the lever that decides whether the next clinical AI tool is built to raise a code or to lower an HbA1c (the blood test that tracks diabetes control over months).

Sources

• Peterson Health Technology Institute. Driving Adoption of High-Value AI in Healthcare. Last updated 15 October 2026. Web page and full report, p. 15 (cost per claim) and p. 16 (the scribe figure, citing KLAS Research, "Suki ROI Validations 2026," January 2026). https://www.phti.org/research-insight/driving-adoption-of-high-value-ai-in-healthcare/ and https://www.phti.org/wp-content/uploads/2026/10/PHTI-Driving-Adoption-of-High-Value-AI-in-Healthcare.pdf
• Smith M, Holloman J. "Better Information, Better Care: How Coding Supports Today's Complex Patients." AHA blog, 5 October 2026. https://www.aha.org/news/blog/2026-10-05-better-information-better-care-how-coding-supports-todays-complex-patients
• Fierce Healthcare. "Hospitals' use of AI coding tools cost BCBSA plans $942M more for similar care: analysis." 24 September 2026. https://fiercehealthcare.com/finance/hospitals-use-ai-coding-tools-cost-bcbsa-plans-942m-more-similar-care-analysis
• Blue Cross Blue Shield Association. Hospital Coding Intensity Analysis: Major Bowel Procedures. September 2026. https://www.bcbs.com/media/pdf/BCBSA-AI-Coding-Intensity-Whitepaper.pdf (refused automated access; figures taken from the two reports above).
• Centers for Medicare & Medicaid Services. ACCESS (Advancing Chronic Care with Effective, Scalable Solutions) Model. https://www.cms.gov/priorities/innovation/innovation-models/access

Glossary