The easiest way to produce a ranking of medical AI products is to put their largest public numbers in a table and sort downwards. The result will look authoritative. It may also compare physicians with patient visits, a registered membership with a quarter's activity, and a survey response with a server event.
A spreadsheet cannot repair that category error. It can only make it look organised.
Doctors looking for a useful tool deserve a better comparison. Published adoption figures can tell us something about distribution, activity and the reach of a product. They become misleading when those different observations are squeezed into a single claim about who leads the market.
The starting point is to preserve the unit that came with the number, even when that makes the final table less exciting.
What the published figures actually describe
The examples below were reviewed on 19 October 2026. Their dates remain attached because a historical disclosure and a changing homepage counter should not quietly become measurements from the same week.
| Source and date | Published observation | What it can describe |
|---|---|---|
| Johns Hopkins Medicine announcement, 13 October 2026 | OpenEvidence is described as used daily, on average, by a majority of US physicians | A physician adoption claim requiring its activity definition and denominator |
| Heidi homepage, observed 19 October 2026 | A counter displayed 2,954,736 patient visits this week | A company-displayed volume of visits, rather than a count of individual clinicians |
| Doximity disclosure exchange, February-March 2026 | The discussion distinguishes registered membership from quarterly active providers | Different measures of a professional platform's reach and use |
| AMA physician AI survey, March 2026, page 5 | 72% reported incorporating at least one AI use case into practice; this question had 1,342 respondents | Survey-reported use of the category, rather than market share for one vendor |
Each row answers a different question. A visit counter can rise because existing clinicians see more patients. A membership number can remain high while recent engagement changes. A survey can capture the use of several products by the same respondent. None of those properties makes a measure useless. They determine what the measure is useful for.
The AMA report itself illustrates why the underlying categories matter. Its broader 81% figure combines the 72% reporting at least one use case with 9% uncertain which AI use cases their practice offers. Repeating 81% as an unqualified use rate would discard the distinction shown in its own chart.
Nor does a source becoming more prestigious change the counting unit. A partner announcement can carry a company adoption claim without turning it into a comparable, independently sampled market estimate.
Doximity shows why definitions deserve their own space
In its February 2026 letter, SEC staff asked Doximity to explain what qualified as use of its clinical workflow tools and clarify the relationship between registered members and active users. This was a request for clearer disclosure, not a finding that the numbers were false.
Doximity's 23 March response proposed counting a provider once per quarter after using a specified workflow function. Qualifying activities included calls longer than ten seconds, faxes, AI prompts, scheduling and ambient notes. It also explained that registered physicians must have claimed or created profiles and satisfy stated professional criteria. The company committed to adding definitions and comparative workflow data in its 2026 annual report.
That is a useful distinction for anyone building a medical AI comparison. A provider who qualifies through a communications tool is an active workflow user under that definition. The number cannot then be relabelled as the population using the company's AI assistant.
The same principle applies to any expanding platform. Once several functions share an account, a platform-level activity total needs feature-level context before it can answer a product-specific question.
First match the clinical job
Medical AI is a broad shelf label. Preparing a consultation note, retrieving evidence, drafting correspondence and supporting revision are different jobs. Products may cover several of them, and their boundaries can change. A useful comparison identifies the specific function being assessed rather than assuming the company name represents one permanent category.
Suppose, in an illustrative practice, one clinician uses an ambient documentation feature during most appointments, an evidence assistant for occasional difficult questions and a revision tool before an examination. The documentation feature will probably generate more recorded encounters. That does not establish that it provides better answers to clinical questions or teaches more effectively.
Frequency is partly a property of the task. A system used once for a difficult decision may contribute more value to that decision than one used repeatedly for routine drafting. Conversely, reliable performance on a frequent administrative task can create substantial practical value. The point is to compare outputs that serve the same purpose.
This is why iatroX's Ask-iatroX, Tutor and exam resources should be assessed against their respective jobs. Listing those functions does not establish a market position, and this comparison assigns iatroX no numerical rank.
Then match people, periods and actions
A defensible comparison needs three aligned dimensions. The population should use the same professional and geographical boundaries. The period should cover the same dates. The qualifying action should represent a comparable level of engagement.
Consider two hypothetical products. One reports 100,000 registered clinicians accumulated over several years. Another reports 25,000 clinicians who completed an eligible task in the past month. The first number is larger, but we cannot deduce which has more monthly users. We also cannot turn the second number into a registration estimate.
Even two monthly active figures may diverge in meaning. One service might count opening the application; another might require completing a clinical workflow. If both definitions are disclosed, readers can make an informed comparison. If the definitions are absent, precision in the displayed totals is cosmetic.
The missing comparison should remain missing. Estimating an unpublished number by borrowing another company's engagement ratio would introduce an assumption where readers expect evidence.
Overlap is part of the market
Users do not become exclusive property when they register. The hypothetical clinician who uses three products can properly appear in all three user totals. Adding those totals does not reveal the number of distinct doctors reached by the category.
Equally, a product's presence among most surveyed clinicians would not establish that it receives most of their clinical AI activity. Reach, frequency, preferred tool and expenditure are different possible definitions of market share.
For a buyer, overlap can be informative. It may show complementary functions or unresolved needs that a single product does not meet. The practical question is whether another tool adds enough value to justify another account, another workflow and another source of information to assess.
An adoption statistic should help answer that question. It should not pressure clinicians to conform to an imagined unanimous profession.
Publish a comparison that can survive scrutiny
The honest alternative to a league table is a dated evidence table with explicit units, source attribution and blanks where compatible data are unavailable. It can be supplemented by task-specific evaluations: time to an acceptable note, support for a clinical recommendation, correction burden or learning performance.
A vendor wanting a numerical rank can publish the missing definitions and participate in a comparison using common methods. That would make the resulting position more meaningful, even if the headline number became smaller.
Until then, the largest visible number tells us which company has published the largest visible number. Choosing a clinical tool requires a better question.
Frequently asked questions
Which medical AI product has the most users?
The public examples here do not establish a comparable ranking. Their populations, activity definitions and reporting periods differ. A defensible answer would require aligned measurements for the same type of clinician, geography, period and qualifying activity.
Are patient visits a useful adoption measure?
Yes, for understanding the volume of a workflow that processes visits. They do not identify how many individual clinicians generated that activity, how often those clinicians returned, or whether another product performs better on a different clinical task.
How should a doctor choose when comparable user figures are missing?
Begin with the job you need done. Assess the output, source access, correction burden and fit with your setting using representative tasks. Public adoption can provide context, but it should not substitute for observing whether a tool solves your problem.