AI in US healthcare: where the money goes

Three-quarters of 2025 generative AI app spend in US healthcare went to documentation and to coding and billing: $1.05B of $1.4B (75%, derived here from the figures above, not reported by the source).

The fastest growth is in patient engagement and in prior authorization, where providers are buying AI to deal with insurers.

The cost effects are plausible. I found no independent causal measurement of them.

That is generative AI app spend only. NIH grants and drug-R&D rounds are different kinds of money, and larger ones: NIH’s AI category was $2.3B and one Isomorphic round was $2.1B.

In research the early science is strong. As of October 2026 I found no approved drug that is plainly AI-discovered. There is no standard definition of the term.

Each figure is labelled. Verified means I read it on the primary page. As reported by means I read it in secondary coverage. Lawsuit claims are allegations, surveys of belief are opinion, and a trade group’s analysis is its claim. The evidence against this has its own section.

Providers bought $1B of the $1.4B of healthcare generative AI apps in 2025 (Menlo Ventures). Ambient documentation was $600M and coding and billing $450M (Menlo Ventures).

Growth is elsewhere. Patient engagement grew 20x year on year and prior authorization 10x, from smaller bases. Documentation grew 2.4x. Prior authorization here is provider-side tooling for dealing with insurers, which Menlo files under front-office revenue cycle management. The insurer-side category, payer operations, grew 5x, and Menlo calls it more nascent than the provider-side categories (Menlo Ventures).

Payers show $50M. That is not what payers spend on AI. Menlo counts bought generative AI apps only, and leaves out the models and rules engines insurers build themselves (Menlo Ventures). It also leaves out research AI. Menlo is a venture firm and this is a survey-based estimate.

In Menlo’s executive survey, 27% of health systems, 18% of outpatient providers and 14% of payers were using domain-specific AI (Menlo Ventures). The year before, Menlo put spend at $500M (Menlo Ventures). That came from a cross-industry survey with a different method, so the two years are not strictly comparable.

McKinsey, a consultancy, surveyed US healthcare leaders: the share who had implemented generative AI went from 25% in Q4 2023 to 47% in Q4 2024 and 50% in Q4 2025. Leaders most often cite administrative efficiency as the domain of greatest potential (McKinsey, a consultancy, in its survey of US healthcare leaders). That is potential, not spend.

Billing automation gained +25 points in a year, ahead of scheduling at +16 points and outpatient follow-up at +9 points (ASTP/ONC).

Physicians point the same way, as opinion. In the AMA’s 2023 survey, 54% were enthusiastic about AI for documentation of billing codes, charts or notes and 48% for prior authorization (an opinion survey of 1,081 physicians surveyed by the AMA in 2023). In its separate 2024 survey, 57% named administrative burden the top opportunity (an opinion survey of physicians in the AMA's 2024 survey, a separate survey from 2023). Enthusiasm is not use.

Generative AI is newer. In 2024, 31.5% of hospitals used generative AI integrated with their EHR, and 24.7% planned to within a year (JAMA Network Open). Major teaching hospitals led at 53.9%, then system-affiliated at 38.5%, then independent at 16.3%. The paper does not say what it is used for. Its authors suspect ambient scribes and message drafting.

Patients use it too. 32% of adults used an AI chatbot for health information in the past year. Among those users, 19% said not being able to afford seeing a health professional was a major reason, and 32% of users with incomes under $40,000 (an opinion survey of 1,343 US adults in the KFF Tracking Poll, fielded Feb 24 to Mar 2, 2026 (margin of error 3 points); an opinion survey of chatbot users in the same KFF poll; an opinion survey of chatbot users with incomes under $40,000 in the same KFF poll). That is a self-reported poll, and it is not spending.

Overall use of predictive AI went from 66% to 71% of hospitals (ASTP/ONC). At least 60%+ of health systems use AI or technology on denials and prior authorization (as reported by Modern Healthcare (reprinted by the Connecticut Hospital Association)).

Time saved is measured. Dollars saved are not.

In a randomized trial, one scribe cut time in the note by 9.5%. The other’s change was not significant (as reported by Nabla, a scribe vendor, in its own press release; as reported by UCLA Health). Kaiser Permanente reported 15,800 saved over 2,500,000 encounters, observational and self-reported by the health system (as reported by Becker's Hospital Review). That is about 23 seconds each (derived here from the figures above, not reported by the source). On return on investment, the evidence review found limited evidence of a return, and a risk of raising overall costs, in the Peterson Health Technology Institute's evidence review (as reported by Fierce Healthcare).

At six systems that announced ambient scribes, visits coded at the high levels rose at every one. New patients rose 13.5 to 19.5 points, established patients 7.4 to 11.8 points (computed from the values below) (Trilliant Health).

This is correlation. Trilliant chose the systems because they announced scribes. It gives no adoption dates. The window overlaps the 2021 coding guideline change. Trilliant says the causes are not clearly understood.

A payer trade group says hospital coding cost Blue plans $942M extra over two years on one set of procedures. It says the complex share of inpatient cases went from 37% to 40% (claimed by the Blue Cross Blue Shield Association, a payer trade group, as reported by Fierce Healthcare). It does not identify which claims involved AI. The hospital association replied that older patients and more complex cases explain the shift, payers' own automated downcoding is part of it, and the test is whether "the patient's medical record supports the diagnosis" (claimed by the American Hospital Association, a hospital trade group, as reported by Becker's Payer). Both are interested parties.

A rigorous Medicare study of scribes and spending is funded for 2026 with results to follow, so there is no causal estimate yet (as reported by the University of Minnesota School of Public Health).

• KFF, ACA Marketplace plans in 2023: insurers denied 20% of in-network claims, and below 1% of denials were appealed (KFF). In Medicare Advantage in 2024, 7.7% of 52.8 million prior authorization determinations were denied in full or in part. 11.5% of denials were appealed, and 80.7% of appeals were partly or fully overturned (KFF). KFF does not attribute any of this to AI. It shows the scale of denials and how few are appealed.
• A lawsuit, Estate of Lokken v. UnitedHealth, alleges 90% of appealed denials were reversed under UnitedHealth’s nH Predict tool (an allegation by the plaintiffs in Estate of Lokken v. UnitedHealth, unproven). A suit against Humana alleges the same (an allegation by the plaintiffs in Barrows v. Humana, unproven, as reported by Fierce Healthcare). These are allegations, not findings. UnitedHealth says the tool is not used to make coverage decisions.
• A Senate majority-staff report: UnitedHealthcare’s post-acute prior-authorization denial rate went from 10.9% in 2020 to 22.7% in 2022 (as reported by Healthcare Dive). Accounts of the report differ on the baseline, so the size of the rise is uncertain. The minority and the insurers are not in the report. It does not show AI caused the rise.
• 61% of surveyed physicians believe payer AI is raising denials (an opinion survey of 1,000 physicians surveyed by the AMA, as reported by the American Medical Association (press release)). That is belief, not a measured rate.
• The federal WISeR model tests AI-assisted prior authorization in traditional Medicare: in six states from the start of 2026, with vendors paid a share of the spending they avert; the ACP says this "could create a financial incentive to deny services" (as reported by ACP Advocate).
• Payers have their own fears. In Menlo’s survey they worry that AI prior-authorization, status and appeal tools will bring "surging call and claims volume", and that coding tools built to maximize reimbursement will raise costs (an opinion survey of payers).

NIH funding in its AI category went from $600M in FY2019 to $2.3B in FY2023, in nominal dollars, about 3.8 times (derived here from the figures above, not reported by the source). That is 4.7% of the NIH budget (Journal of Medical Internet Research). The category includes clinical and administrative AI research, not only discovery.

Isomorphic Labs raised $2.1B in May 2026 (Isomorphic Labs). Xaira launched with $1B+ (as reported by BioSpace).

Rock Health: AI-enabled startups took 54% of $14.2B in US digital health venture funding in 2025 (Rock Health), about $7.7B (derived here from the figures above, not reported by the source). For 2024, the earlier report gives 37% of the $10.1B first published, across 191 deals (Rock Health). The 2025 report repeats 37% beside a restated $10.5B total and does not recompute it (Rock Health). The 2024 and 2025 reports define “AI-enabled” differently: 2024 counts any solution powered by AI, not just companies selling AI as their core product, and 2025 counts startups using AI “as a core part of their product or offering”. The 2025 report does not say the definition changed, so part of the rise from 37% to 54% may reflect the wording, not growth alone. Rock Health excludes biotech, so this is not research money.

CB Insights counts global funding and includes AI drug developers. It puts AI-focused companies at 42% of digital health funding in 2024 (a later report gives 41%), then 60% in Q1 2025 and 69% in Q2 2025, single quarters (CB Insights). Global digital health funding was $22.3B in 2025 (CB Insights). CB Insights publishes no public split between administrative, clinical and research AI.

I found no source that compares administrative AI dollars with research AI dollars on one definition. Menlo’s buyer split is the closest, and it covers bought apps only.

As of October 2026 I found 0 full FDA approvals of AI-discovered drugs (as reported by IntuitionLabs). The count depends on the definition. Zasocitinib is under FDA Priority Review, with a decision expected in the first quarter of 2027; it was designed with physics-based computation, and whether it counts as "AI-discovered" is contested (as reported by Takeda).

AlphaFold won a Nobel Prize (as reported by The Next Web). Rentosertib is a published Phase 2a result in Nature Medicine with an AI-derived target and molecule, small and short (as reported by BioSpace (carrying Insilico's own release)). 80 to 90% of AI-discovered molecules passed Phase I, in a sample of about two dozen. Phase II was about 40%, near the historical average (as reported by HT World). Setbacks: two Phase II programs discontinued and the pipeline cut (as reported by GEN (Genetic Engineering & Biotechnology News)).

A count of approvals this early settles little either way. The next data point is zasocitinib’s FDA decision.

• Saved time is real. The trial and Kaiser both show less documentation time (as reported by Nabla, a scribe vendor, in its own press release; as reported by Becker's Hospital Review).
• Clinical use is large. In McKinsey’s survey, 54% of care-organization respondents had implemented generative AI for clinical productivity, the most widely adopted domain. McKinsey says clinical-care organizations are moving beyond using it mostly for administrative tasks (McKinsey, a consultancy, in its survey of US healthcare leaders).
• Better coding is not necessarily upcoding. The hospital association says capturing real diagnoses is legitimate. The trade group cannot say which claims involved AI (claimed by the American Hospital Association, a hospital trade group, as reported by Becker's Payer; claimed by the Blue Cross Blue Shield Association, a payer trade group, as reported by Fierce Healthcare).
• Some of the rise may predate AI. The 2021 guideline change overlaps the Trilliant window, and Trilliant says the causes are not clearly understood (Trilliant Health).
• The denial claims about nH Predict are unproven. They are allegations in two suits, and UnitedHealth says the tool is not used to make coverage decisions (an allegation by the plaintiffs in Estate of Lokken v. UnitedHealth, unproven).
• Automation may save money. CAQH estimates the savings opportunity if routine administrative transactions move from manual to electronic workflows at $20B a year. That is an opportunity, not a realised saving, and it is not about AI (the CAQH Index). An NBER working paper by McKinsey and Harvard authors estimates AI could save 5% to 10% of US health spending, much of it administrative (NBER). That is an estimate, not an outcome.
• The biggest single bets are in drug R&D. CB Insights says AI drug discovery dominated digital health deal activity, with the largest deal an AI drug developer in Q1 2026 (CB Insights) and funding was centred on AI drug discovery in Q2 2026 (CB Insights), including Isomorphic’s $2.1B round. Its Q1 2025 report also lists AI small molecules at $204M and Isomorphic’s earlier $600M round (CB Insights). On the other side, clinical documentation raised $372M in Q1 2025 and $639M in Q2 2025, but those are two quarters of top-market lists, not a full split (CB Insights).
• The research has milestones. A Nobel, a published Phase 2a result in Nature Medicine with an AI-derived target and molecule, small and short and a drug under Priority Review. “No approved drug yet, strong early science” is what the evidence supports.
• Research money is large. NIH’s $2.3B category exceeds Menlo’s $1B of provider app spend, though the two measure different things.

• A one-definition comparison of administrative and research AI dollars.
• A public figure for insurers’ in-house AI spend, or for pharma’s in-house AI research spend.
• A peer-reviewed causal estimate of AI coding tools on billing.
• Primary pages for several figures. Those are labelled as reported.

• Menlo Ventures · Menlo Ventures, "2025: The State of AI in Healthcare" (21 Oct 2025): survey-based estimate of generative AI app spend
• ASTP/ONC Data Brief No. 80 · ASTP/ONC, "Hospital Trends in the Use, Evaluation, and Governance of Predictive AI, 2023-2024", Data Brief No. 80 (Sep 2025), from the AHA IT Supplement: non-federal acute care hospitals, n=2,080 in 2024
• Everson, Nong and Richwine, JAMA Netw Open 2025 · Everson, Nong, Richwine, "Uptake of Generative AI Integrated With Electronic Health Records in US Hospitals", JAMA Network Open 2025;8(12):e2549463 (12 Dec 2025): 2,174 nonfederal acute care hospitals, 2024 AHA IT Supplement
• Modern Healthcare via CHA · Modern Healthcare on a McKinsey survey of revenue-cycle leaders (30 Jan 2026)
• Nabla release on the UCLA trial · Nabla press release on the NEJM AI randomized trial of ambient scribes
• UCLA Health on the scribe trial · UCLA Health news release on Lukac et al., "Ambient AI Scribes in Clinical Practice: A Randomized Trial", NEJM AI (Dec 2025)
• Becker's on Kaiser scribes · Becker's Hospital Review on the Kaiser Permanente (TPMG) ambient scribe rollout, citing the KP release and NEJM Catalyst
• Fierce on the PHTI report · Fierce Healthcare on the Peterson Health Technology Institute ambient scribe report (25 Mar 2025)
• Healthcare Dive on payer AI · Healthcare Dive on the Senate PSI majority staff report on Medicare Advantage prior authorization (17 Oct 2024)
• Lokken v. UnitedHealth order · Estate of Lokken v. UnitedHealth, D. Minn.: memorandum opinion and order (13 Feb 2025), via the Georgetown litigation tracker
• Fierce on the Humana suit · Fierce Healthcare on the class action against Humana, Barrows v. Humana, W.D. Ky. (filed 12 Dec 2023)
• AMA prior authorization survey · AMA 2024 Prior Authorization Physician Survey, 1,000 physicians (released Feb 2025)
• ACP Advocate on WISeR · American College of Physicians, ACP Advocate (12 Sep 2025), on the CMS WISeR model
• Fierce on the BCBSA analysis · Fierce Healthcare on the Blue Cross Blue Shield Association analysis of hospital AI coding tools (Sep 2026)
• Becker's on the AHA reply · Becker's Payer on the American Hospital Association's reply to the BCBSA analysis (5 Oct 2026)
• Trilliant Health · Trilliant Health, "Increased outpatient coding intensity following hospital adoption of AI-enabled scribing": six systems that announced ambient scribes, all-payer claims
• UMN on the scribe study · UMN School of Public Health: a rigorous nationwide study of ambient scribes and spending (NIHCM-funded, 2026)
• CAQH Index 2024 · 2024 CAQH Index: cost of routine administrative transactions and savings from electronic workflows
• NBER w30857 · Sahni, Stein, Zemmel, Cutler, "The Potential Impact of Artificial Intelligence on Healthcare Spending", NBER Working Paper 30857 (2023): not peer-reviewed, an expert estimate
• Le et al., JMIR 2026 · Le et al., J Med Internet Res 2026;28:e84861: NIH funding in the "Machine Learning and Artificial Intelligence" category, FY2019 to FY2023
• NLM on Bridge2AI · NLM news release: NIH launches the Bridge2AI program (Sep 2022)
• Isomorphic Labs · Isomorphic Labs press release on its Series B (12 May 2026)
• BioSpace on Xaira · BioSpace (Business Wire): Xaira Therapeutics launches (23 Apr 2024)
• Rock Health 2025 · Rock Health 2025 year-end digital health funding overview (excludes biotech and drug discovery)
• IntuitionLabs on approvals · IntuitionLabs, AI drug discovery and FDA approvals (an aggregator; depends on what counts as "AI-discovered")
• Takeda on zasocitinib · Takeda newsroom (14 Sep 2026): FDA Priority Review for zasocitinib in psoriasis
• BioSpace on rentosertib · Insilico Medicine release via BioSpace: Nature Medicine publication of the Phase 2a rentosertib results (3 Jun 2025)
• HT World on AI-drug success rates · HT World on Jayatunga et al. (BCG/Wellcome), Drug Discovery Today (May 2024): phase success rates of AI-discovered molecules
• GEN on Recursion · GEN: Recursion halts four pipeline programs (May 2025)
• The Next Web on the Nobel · The Next Web: Google DeepMind duo win the Nobel Prize in Chemistry (9 Oct 2024)
• Menlo Ventures 2024 · Menlo Ventures, "2024: The State of Generative AI in the Enterprise" (20 Nov 2024): a survey of 600 US IT decision-makers across industries
• AMA 2023 physician survey · AMA (12 Jan 2024): a survey of 1,081 physicians on health care AI, fielded in 2023
• AMA 2024 physician survey · AMA press release (12 Feb 2025): physician enthusiasm for health care AI grows, 2024 survey
• Resourcera, AI Healthcare Statistics 2026 (aggregator) · Further reading: an aggregator (updated 16 Apr 2026) that mostly does not tie its figures to sources. No figure on this page is taken from it.
• AI Healthcare 360, AI in Healthcare Statistics 2026 (aggregator) · Further reading: an aggregator. No figure on this page is taken from it.
• KFF on ACA claims denials · KFF, "Claims Denials and Appeals in ACA Marketplace Plans in 2023" (published 27 Jan 2025, updated 24 Mar 2026): CMS Transparency in Coverage data, HealthCare.gov plans. Does not attribute denials to AI.
• KFF on Medicare Advantage prior authorization · KFF, Medicare Advantage insurers made nearly 53 million prior authorization determinations in 2024 (28 Jan 2026): CMS Part C and D reporting data. Does not attribute the figures to AI.
• CB Insights, State of Digital Health 2024 · CB Insights, State of Digital Health 2024 (16 Jan 2025): global digital health funding; AI-focused companies include AI drug developers
• CB Insights, State of Digital Health Q1'25 · CB Insights, State of Digital Health Q1'25 (17 Apr 2025): global equity funding
• CB Insights, State of Digital Health Q2'25 · CB Insights, State of Digital Health Q2'25 (24 Jul 2025): equity funding
• CB Insights, State of Digital Health 2025 · CB Insights, State of Digital Health 2025 (22 Jan 2026): global digital health funding
• CB Insights, State of Digital Health Q1'26 · CB Insights, State of Digital Health Q1'26 (16 Apr 2026)
• CB Insights, State of Digital Health Q2'26 · CB Insights, State of Digital Health Q2'26 (30 Jul 2026)
• McKinsey, Generative AI in healthcare (Apr 2026) · McKinsey, "Generative AI in healthcare: Adoption matures as agentic AI emerges" (16 Apr 2026, Lamb, Dahab, Jones): a consultancy survey of US healthcare leaders, Q4 2025
• KFF poll on AI chatbots and health information · KFF, "Poll: 1 in 3 Adults Are Turning to AI Chatbots for Health Information, Equaling the Share Who Use Social Media for Health" (25 Mar 2026)
• Rock Health 2024 · Rock Health, "2024 year-end market overview: Davids and Goliaths": AI-enabled means any solution powered by AI, not just companies selling AI as their core product
• Azumo, 60 AI in Healthcare Statistics 2026 (aggregator; Azumo sells AI services) · Further reading: an aggregator whose publisher sells AI services. No figure on this page is taken from it.