Phase 2: The buyer · Prelude
The newest medical AI doesn't need the FDA. It still needs a doctor to check its work.
Most AI on the FDA's list is imaging and heart AI. The newer AI, such as note-writing and answer tools, works across specialties and hasn't gone through FDA review. By who uses it, about half of US medical-AI money is an organization paying for its own people, and most of the rest buys a result. Phase 2 follows the money into the workflow.
The question: Who pays for medical AI, and who uses it?
By Chi-way Li · Nov 4th 2026
On Nov 1 I wrote about OpenEvidence, an AI that answers doctors' questions. Doctors use it for free, drug-company ads pay for it, and it isn't on the FDA's list of AI devices. One comment stuck with me: the person who uses an AI tool is often not the one who buys it. Doctors picked up OpenEvidence on their own. This year Sutter Health, Mount Sinai and NYU Langone announced plans to build it into their Epic records, which makes it a hospital decision. Another doctor argued it would have spread even if it cost money, and free may have mattered less than I thought.
Phase 2 of this study is about that buyer. Before it starts, here is what Phase 1 already tells a buyer, and what it doesn't.
Two waves of medical AI
Phase 1 started from the 1,614 AI-enabled devices the FDA had authorized by June 2026. Most are imaging or heart devices: the FDA's radiology panel reviewed 1,230 and its cardiovascular panel 154, together 86%. I think that's partly because medical images went digital early, and partly because of how the law is written. Software that analyzes a medical image or a monitor's signal is a medical device. Software that drafts notes, books visits or answers a doctor's question usually isn't.
Nearly all of these devices, 1,576 of the 1,614 (97.6%), are sold to help a clinician, who still has to check the AI's work. Only seven authorizations, for three products that read retina photos, diagnose on their own. The FDA's main route tends to keep it that way. A 510(k) clears a device by showing it is as safe and effective as one already sold for the same use, so the first device of each type sets the claim and later ones tend to copy it. Phase 1 called the cost of the clinician's check the verification tax.
The AI now spreading across hospitals looks different. Note-writing AI works for any clinician who writes notes, and answer tools for any doctor with a question. None of them has gone through FDA review, since the law exempts most such software. A hospital tends to buy note-writing AI centrally, through IT or its chief medical information officer, while imaging AI, which has spread widely too, is usually bought one department at a time. OpenEvidence needed no purchase at all, and its maker says a majority of US doctors use it daily, on average. Hospitals and practices paid about $0.6B for note-writing AI in 2025, by a venture firm's estimate that Phase 1 used, without a single FDA device.
I think that's much of why this AI has spread so quickly. It skipped FDA review, one purchase or none covers every specialty, and checking a draft replaces work clinicians used to do. Next week tests part of this: who checks the AI's work, and what they get back.
Who pays, and who uses it
Phase 1 estimated about $2.9B of US spending on medical AI in 2025. Only about $1.3B of it bought FDA-authorized products. The chart sets the same money beside who uses the AI.
Imaging and heart teams get AI from four payers, while note-writing AI comes from one
US spending on medical AI in 2025, about $2.9B, grouped by who uses it. Color shows who pays.
A first, rough map. Spending is Phase 1's US estimate for 2025 without the $0.57B drug makers paid for data; sizes are rough. "Who uses it" is the main user of each kind of AI, from public cases; Phase 2 checks it. "Own organization" means the organization where the people who use the AI work. Hatched: AI assistants and agents, the newer wave: note-writing AI, patient-call and care-team agents, EHR and reference-tool assistants and OpenEvidence's ads, about $0.92B; mostly generative AI (the patient-call agents include about $0.05B of older conversational AI) and none of it on the FDA's list. The imaging and heart row includes the autonomous eye exam, used in primary care, and patient-finding AI on pathology slides. Not shown: about $2.8B of insurer-paid tests and monitoring and $1.1–1.4B of employer-paid programs where AI is one part of what is sold, and billing and coding AI, which Phase 1 did not count.
Show the numbers
| What is bought | Who pays | Who uses it | Where from | Newer wave | $B, 2025 |
|---|---|---|---|---|---|
| AI-scored lab tests, such as Decipher | Insurers | Doctors who get an AI score (the lab runs the AI) | Outside | 0.72 | |
| Heart-scan analysis and other AI billed per use, including the autonomous eye exam and add-on payments | Insurers | Imaging and heart teams | Outside | 0.23 | |
| Imaging and heart AI subscriptions | Hospitals and practices | Imaging and heart teams | Own organization | 0.5 | |
| AI options and upgrades on scanners | Hospitals and practices | Imaging and heart teams | Own organization | 0.215 | |
| Drug makers paying for patient-finding AI on heart tests and slides (Viz.ai, Anumana, Ultromics, PathAI) | Drug makers | Imaging and heart teams | Outside | 0.07 | |
| Patients paying $40 for an AI mammogram read | Patients | Imaging and heart teams | Outside | 0.03 | |
| Note-writing AI | Hospitals and practices | Clinicians writing notes | Own organization | Yes | 0.6 |
| EHR vendors’ priced AI beyond notes, such as Epic Art | Hospitals and practices | Clinicians writing notes | Own organization | Yes | 0.01 |
| Drug-company ads in clinician answer tools, mainly OpenEvidence | Drug makers | Doctors looking up answers | Outside | Yes | 0.095 |
| Hospital-paid AI in reference tools, such as UpToDate Expert AI | Hospitals and practices | Doctors looking up answers | Own organization | Yes | 0.01 |
| Patient-call, scheduling and intake agents | Hospitals and practices | Patients, calling or being called | Own patients or members | Yes | 0.16 |
| Care-team agents, such as Hippocratic AI | Hospitals and practices | Patients, calling or being called | Own patients or members | Yes | 0.035 |
| AI care calls to health-plan members | Insurers | Patients, calling or being called | Own patients or members | Yes | 0.01 |
| Procedure and dental AI, such as polyp detection | Hospitals and practices | Proceduralists and dentists | Own organization | 0.13 | |
| VA and other public contracts | Governments | Clinicians in public health systems | Own organization | 0.07 | |
| Total | 2.885 |
Source: Cleared to Paid US spending estimate (article 1.3); users from public casesHealthcare AI Insights
About half, $1.54B, is an organization paying for AI its own people use, such as a hospital buying note-writing AI for its clinicians. About $1.15B comes from outside the organization that uses the AI. Most of that, about $0.95B, buys a finished result, mostly paid by insurers: an AI-scored lab test, or a heart-scan analysis billed per use. The last $0.21B is hospitals and health plans paying for AI their own patients talk to. The pattern I see is that the further the payer is from the user, the more it pays for a result. Most help is bought by the user's own hospital or practice. Who carries the check that comes with it is next week's question.
The same teams also draw on several pockets. Imaging and heart teams get AI from four payers. Doctors looking up answers get OpenEvidence, paid by drug-company ads, and UpToDate's AI, paid by their hospital.
This is a first, rough map. Phase 1 counted who pays. Phase 2 looks at who uses the AI, who checks its work and whose budget gains, by specialty, by role and by workflow.
What Phase 2 asks
Phase 2 follows the money into the workflow, one role at a time:
- The people who use the AI: who checks its work, and what do they get back? (next week)
- Buyers outside the hospital: drug makers, insurers buying for their own work, employers, governments and patients. Does each one buy again? (Nov 17)
- Budgets inside the hospital: workflow by workflow, which one keeps paying once the pilot money runs out? (Dec 1)
If you use AI at work, did anyone ask you before it was bought?
Join the discussion on LinkedIn: the post
Sources
FDA AI-enabled device list, downloaded 2026-10-11 (1,614 devices, decisions to 2026-06-29); FDA 510(k) and De Novo summaries; Federal Food, Drug, and Cosmetic Act section 520(o), added by the 21st Century Cures Act (2016); Cleared to Paid device coding (data/classification/final_axes.csv); the Cleared to Paid estimate of US medical-AI spending in 2025 (article 1.3); Phase 2 buyer and workflow map (public cases, to be published in Phase 2); Viz.ai and Bristol Myers Squibb agreement on Viz HCM (Business Wire, 2023-03-03); articles 1.1, 1.2 and 1.3; Sutter Health (Business Wire, Feb 11, 2026), Mount Sinai (Apr 2026) and NYU Langone (Oct 29, 2026) announcements on OpenEvidence in Epic; NYU Langone Health news release, Oct 29, 2026 (OpenEvidence's reach); Menlo Ventures, 2025: The State of AI in Healthcare (note-writing spending, as used in article 1.3).
How our totals fit. Phase 1 estimated $3.5B of US spending on medical AI in 2025, including $0.57B that drug makers paid for data, which is not AI itself; without it, $2.9B (article 2 of this phase). Phase 1's rule left out billing, coding, prior-authorization and operations AI; adding our estimate for those, about $0.7B, gives $3.6B (article 3). About $1.3B of the $3.6B buys FDA-authorized devices; the other two-thirds does not (the prelude). Article 3's spending chart shows the $3.3B that pays for care work; the rest is drug makers' AI deals, government contracts and insurers' own AI.