In Part 1, “Cleared, Not Paid,” I looked at who pays for medical AI. Analyst Chi-way Li counted 1,614 FDA-authorized AI devices and could tie only 79 to a billing code of their own. Most of the rest are bought from hospital budgets. This part asks what that buyer gets for the money, and what it leaves unpaid. It stands alone if you are arriving here first.
I have spent decades in GI practice and now work as a hospitalist. I am also founder and CEO of AiMOS Systems Corporation, so weigh my argument accordingly. Every figure below is attributed, and where a number is secondhand, I say so beside it.
The verification tax
Li offers a working explanation for why AI so rarely earns payment of its own. Until AI is reliable enough on hard cases that nobody needs to check it, a clinician reviews each output. The time the AI saves, or the miss it catches, is then hard to see and harder to bank. Li calls this a verification tax and labels it a working explanation, not a finding. I treat it the same way: an idea worth testing.
Scope matters here. The evidence below covers two kinds of tools, ambient scribes that draft visit notes and AI drafts of replies to patient messages. These are the best-measured cases. I found no comparable evidence for imaging AI, and I make no claim about it.
Li draws one implication. If the idea is right, AI whose mistakes are cheap to catch, such as scribes drafting notes that doctors read anyway, should spread faster than AI that reads images. That is why I looked at scribes and inbox drafts. My inference is that the check ought to be easiest to measure there.
Li also names a limit on his own idea. The autonomous eye exam needs no clinician’s check and has had a billing code since 2021, yet Medicare paid doctors only $138,000 for it in 2024. The verification tax cannot be the whole story.
What scribes show
Start with the largest analysis I found. A multicenter study in JAMA of 1,809 scribe users, as reported by MDLinx in June 2026, found about 16 fewer minutes of documentation and 13.4 fewer minutes of total EHR time per eight-hour shift, with no significant drop in after-hours work. I have not read the primary paper, which another report attributes to Rotenstein and colleagues, so treat these figures as reported. The report I read did not explain why the total-EHR saving is smaller than the documentation saving.
A Penn Medicine pilot points the other way on after-hours work. In a news release, Penn described a 46-clinician pilot with 20 percent less EHR time, 30 percent less after-hours time and about two more minutes of face-to-face time. This was an early-phase pilot, the release does not name the vendor, and a release is not the paper. ENT Today reported secondhand that note time fell from 10.3 to 8.2 minutes and after-hours time from 50.6 to 35.4 minutes.
The third source is the one I find most useful, because it measured both time and perception. In a 2026 study in the American Journal of Managed Care, Adler-Milstein and colleagues studied 252 physicians. Measured documentation time fell from 22.2 to 17.3 minutes per scheduled hour. The savings were concentrated among physicians with high baseline documentation time.
Taken together, scribe studies report real but modest savings. After-hours effects are mixed: Penn reported a fall, and the large JAMA analysis, as reported, did not.
I would not rank these sources against each other too eagerly. The Penn pilot was small and early, and it comes from an institution’s own release. The JAMA analysis is much larger but reaches me secondhand. The AJMC study is the only one of the three I read in full. Each is a piece of the picture. None settles whether a given hospital will bank the savings.
What inbox drafts show
Drafting replies to patient messages has a thinner record. A University of California San Diego randomized study found that AI-drafted replies did not reduce response time. I am citing only that headline result. The release did not carry the detailed percentages, and I have not read the full paper.
A Stanford study in JAMA Network Open of 162 clinicians ran five weeks with a single group and no comparison arm. It found no change in the time to take action, write or read. Utilization was about 20 percent, meaning clinicians used the drafts in a minority of cases, so this is a limited test, not a definitive one.
Happier, not faster
The most consistent finding across scribe studies is not about time. It is about how clinicians feel.
Chowdhury and colleagues ran an open-label randomized crossover trial of two scribes in JAMIA. Of 160 participants enrolled, 136 were analyzed. Both tools reduced burnout scores, and the two tools did not differ. There was no arm without a scribe, so the trial shows burnout scores were lower after use, and cannot show the scribes caused it.
UW Health announced in December 2025 a trial of 66 physicians and advanced practice providers. It described a clinically meaningful reduction in burnout correlated with about 22 fewer minutes of documentation per day. That is a press release, it gives no effect sizes, and I would want to read the full paper before leaning on it.
At Kaiser Permanente, researchers reported nearly 16,000 hours saved from October 2023 to December 2024, and 88 percent of surveyed physicians reported a positive effect on their visit interactions. The design was observational and self-reported, with about 102 survey responses.
Now the perception gap. In the AJMC study, 86.5 percent of the 252 physicians said the scribe reduced their documentation time. The authors found no significant association between perceived and measured change (odds ratio 0.975, 95 percent confidence interval 0.94 to 1.01, P=.144). The confidence interval spans 1.0. In plain terms, how much time a physician felt they saved did not reliably track how much time the record showed.
A word of caution about a more dramatic number. A flash poll of 103 physicians reported by Medical Economics in May 2026 found that 42 percent saved zero minutes and 20 percent saved more than an hour. The poll was sponsored by Heidi, a vendor, and its method was not disclosed. I would not rest an argument on it, in either direction.
Where the time goes
If the savings are modest and the clinic day does not shrink, where do the minutes go? Medical Republic, reporting in April 2026 on a JAMA analysis, said that savings may be reallocated to other work. It also warned that health systems may convert saved time into more visits. I have not read the JAMA paper behind that account, and I do not mix its figures with those above, because it describes a different subgroup and comparison.
A saved minute that goes back into the schedule has not been returned to the physician. It has been sold to the next patient. That may be good for access. It is not the saving the purchase order promised, and it does not pay for the review.
The strongest objection
The strongest objection comes in two parts. First, review is already paid. A physician who reads and signs an AI-drafted note is doing what physicians have always done, and that work is compensated inside the evaluation and management visit fee, relative value units or a salary. Second, burnout may be the right metric. If clinicians are less exhausted and stay in practice, minutes may be a poor yardstick.
I concede the second point outright. Lower burnout is a real benefit, and the trials above point toward it, within the limits I listed. A scribe that makes the day more bearable does not need to be faster to be worth buying.
My answer to the first point is an argument, not a finding. The payments that cover review today predate AI review, and they do not scale with the number of AI outputs a physician must check. A visit fee is set per visit. If AI drafts more items for a physician to read, the fee does not rise with the number of items. Review work can grow while its payment stays flat. That is the mismatch I mean by “unpriced.”
A 2026 review in JMIR by Gong, Bang and Lee of 21 studies and 2,885 clinicians concluded that it is unclear whether scribes reduce cognitive burden or shift it to verification. It covered several types of AI, not scribes alone, and many of the clinicians were early adopters, so it is a caution, not a verdict on scribes.
What the gap means for a buyer
A hospital that buys a scribe or an inbox tool is buying two things. One is the output, a draft note or reply. The other is the physician’s time to check it. The vendor prices the first. The second sits inside salaries and visit fees that were set before the tool existed, and I found no line item that tracks it.
This is why Part 1 matters here. Li’s classification, which he says rests partly on inference, is that most cleared AI appears to be paid for from the buyer’s own budget, not by an insurer billing separately for it. When the buyer is a hospital and the payer pays nothing extra, the hospital carries the cost of the check and must find the return in staffing or throughput. Li’s question to hospital leaders is whether the time an AI tool promised to save showed up there. The AJMC finding suggests that physician perception alone is a poor way to answer it.
The gap in the literature
Here is what I looked for and did not find. I found no study that totals review time across the AI tools a physician now uses, or that puts a dollar value on it. I found none measuring review time for imaging AI. A May 2026 commentary in HealthManagement by Kamara reached a similar conclusion, saying direct measurement of the verification burden has not been published. That is commentary, not data, and my searches were not exhaustive. A study may exist that I missed.
What I can say is narrower. The studies measure documentation time, EHR time, burnout and perception. They do not measure the minutes spent reading, correcting and signing AI output, and none that I found tries to total those minutes across tools or attach a price to them.
I also want to mark what is inference. The finding is that perceived and measured savings do not reliably track each other, and that a review of the literature cannot say whether burden is reduced or shifted. The inference is mine: that reading and correcting AI output takes real physician time, and that this time is a cost someone bears. That it is large, or that it cancels the savings, is not established by anything I found. The honest summary is that the cost is probably real, unmeasured and unpriced, and that its size is unknown.
For a hospital leader, that is a practical problem. Li’s advice to hospital leaders is to ask vendors how long the clinician’s check takes and what is left of the promised saving after it. I did not find a published answer, which is what you would expect if few studies have asked. Until someone measures the check, the purchase decision rests on perceived savings, and the AJMC result shows how far perception can drift from the record.
What comes next
If the check is real and unpaid, the next question is what happens when it fails. A physician who reviews an AI draft is expected to catch its errors, whether or not anyone budgeted time for the review. In Part 3, “Who Owns the Failure,” I look at how the AMA’s new governance framework allocates accountability, what malpractice carriers have said on the record, and one preprint on whether expert review catches an AI that is steering the reader.
Brian K. Hudes, MD, is a practicing gastroenterologist and writes The Renaissance Physician. Disclosure: he is the founder and CEO of AiMOS Systems Corporation and has a commercial interest in health care technology and clinical artificial intelligence.
Selected sources
Chi-way Li, Oct 14, 2026, updated Oct 16 (analysis of who pays for cleared AI): Most cleared AI is bought, not billed
Adler-Milstein et al., American Journal of Managed Care, 2026 (physician perception versus measured scribe effects): AJMC 2026;32(1):34-40
MDLinx, Jun 8, 2026 (secondhand report of a multicenter JAMA analysis of 1,809 clinicians): AI scribes and modest savings
Penn Medicine, news release, Feb 2025 (46-clinician AI scribe pilot): Penn Medicine release
Collins, ENT Today, Jul 8, 2025 (secondhand summary of scribe studies): ENT Today article
University of California San Diego, Apr 15, 2024 (randomized study of AI-drafted replies): UC San Diego release
Garcia et al., JAMA Network Open, 2024 (Stanford inbox-draft study of 162 clinicians): Stanford publication page
Chowdhury et al., JAMIA, 2026 (crossover trial of two scribes): JAMIA 2026;33(5):990-999
UW Health, Dec 12, 2025 (news release on a trial of ambient AI): UW Health release
Greene, Kaiser Permanente Division of Research, Apr 1, 2025 (observational, self-reported physician survey): Kaiser Permanente Division of Research
Littrell, Medical Economics, May 5, 2026 (Heidi-sponsored flash poll of 103 physicians): Medical Economics article
Attard, Medical Republic, Apr 22, 2026 (secondhand report on a JAMA analysis): Medical Republic article
Gong, Bang and Lee, JMIR, 2026 (review of 21 studies): JMIR 2026;28:e93618
Kamara, HealthManagement, May 15, 2026 (commentary on the cost of verification): HealthManagement article