Hospitals have AI products. Why so little of it is in daily use, and what the usual numbers do and do not show

Radiology has had commercial AI products for longer than any other clinical specialty: tools that flag lung nodules on CT scans, prioritise urgent studies in a worklist, measure structures automatically. Availability has not translated into routine use. In a 2020 survey of its members by the American College of Radiology's Data Science Institute, published in 2021, about 30% of responding radiologists said they used AI in their clinical practice. The authors described the penetration as modest.

Jan Beger, Global Head of AI Advocacy at GE HealthCare, made the same observation in an interview for Tech Anatomy, recorded at the Adopt AI summit at the Grand Palais in Paris in November 2025. "When you see what is deployed and in day-to-day use there, it's still very low usage, even in a leading domain such as radiology," he said. Beger previously founded GE HealthCare's Edison Accelerator, its programme for AI start-ups, and his current role centres on AI literacy among healthcare professionals.

His explanation for the gap is mostly not about the technology.

Point solutions and the inventory problem

Beger lists the familiar obstacles first: reimbursement, interoperability, data quality. He then describes two that receive less attention.

The first is strategic. Many health systems, he says, have deployed AI one department at a time, a nodule detector here, a triage tool there, without a plan for the institution as a whole. Leaders often do not know what is already in use. His suggested first step is modest: a spreadsheet, filled in team by team, listing which AI tools are running and how they are used. He expects many hospitals that believe they "don't use it at all" would find otherwise.

The second follows from the first. A narrow tool is hard to evaluate financially. When an algorithm covers one task in one workflow, the return on investment, if there is one, is spread thin and difficult to isolate. Beger expects this to change as multimodal models (which combine images, text and other data types) and agentic AI (systems that carry out several steps of a task under human oversight) cover broader use cases. That is a forecast.

Literacy as three skills

The non-technical point he stresses most is AI literacy, which he breaks into three abilities: evaluating an AI output responsibly, working alongside a tool efficiently, and explaining what a tool produced to patients and to the rest of the care team. His preferred method for teaching it is the one medicine already uses: share concrete use cases, measure their effect and look at outcomes.

The example he chose was ambient AI scribes, software that listens to a consultation and drafts the clinical note. He presents them as a way to give back some of the attention that screens and keyboards have taken from the consultation. The evidence here is recent and still limited. A randomised trial at a single US academic centre, published in NEJM AI in late 2025, assigned 238 outpatient physicians to one of two commercial scribes or to usual care. One product reduced time spent writing notes by about 9.5% compared with the control group; the other showed no significant change. Survey measures of burnout improved in both scribe groups, which the authors treat as secondary findings to be confirmed in larger, multicentre trials.

Beger also referred to research suggesting that patients prefer chatbot answers to those of physicians, including on empathy. The best-known study, published in JAMA Internal Medicine in 2023, compared 195 physician replies on a public Reddit forum with ChatGPT's answers to the same questions. Evaluators preferred the chatbot in 78.6% of assessments and rated it more empathetic far more often. The evaluators were healthcare professionals rather than patients, the chatbot's answers were several times longer, and the authors note that accuracy was not separately assessed. Beger's conclusion from it was not that chatbots should replace the conversation, but that empathy is the part of the job clinicians will need to cultivate.

The figures behind the case for AI

In his keynote at the summit, Beger built the argument for AI on three pressures: the volume of data and literature, the quality of care, and workforce shortages in ageing societies. Some of the figures he cited circulate widely in the sector, and each can be traced to an origin.

"Every 72 days, all medical data doubles" traces back to a 2011 paper by Peter Densen on medical education. Densen estimated that the doubling time of medical knowledge was about 3.5 years in 2010 and projected 0.2 years, roughly 73 days, by 2020. It was a projection from a curve, not a measurement, and it concerned knowledge rather than data.

The claim that a general practitioner would need to read for 21 hours a day to keep up comes from a 2004 study by Brian Alper and colleagues. They timed physicians appraising a sample of articles and estimated that evaluating the monthly output of journals relevant to primary care would take 627.5 hours, which works out at about 21 hours a day over a 30-day month. The estimate is more than twenty years old.

The statement that 30% of the world's data is healthcare-related is usually attributed to RBC Capital Markets, which published it without a cited method. At least one data company has argued the real share is closer to 11 to 14%. We did not find a published source for a third figure Beger mentioned, that a dermatologist would need to read 3,900 papers a year, and have left it aside.

Two readings

One reading takes these numbers as broadly right in direction, whatever their precision. Clinicians cannot read everything, documentation consumes time, staff are short, and tools that summarise, draft or triage address real constraints. On this view the slow uptake reflects missing skills and strategy, which can be fixed.

A second reading treats slow adoption as partly rational. Beger himself argues that validation and regulatory approval are slow and expensive "but it's worth it", and that responsibility, transparency and explainability cannot be afterthoughts. A radiology department that declines a tool with uncertain financial return and little outcome evidence may be exercising the very literacy he calls for.

What the interview does not establish

The conversation was short and recorded at a conference. Beger's account of how health systems approach AI comes from his work with customers and partners, not from a published survey. His view that 2025 was a tipping point, because of multimodal and agentic AI, is a forecast. He also speaks for a company that sells imaging equipment and AI software, which gives him both a broad view of hospitals and an interest in their adoption of AI.

What to watch

Two things would test his diagnosis. The first is whether hospitals that run an AI inventory and set an institution-wide strategy show higher sustained use than those that buy tools one by one. The second is whether broader tools such as ambient scribes produce the measurable returns that narrow ones struggled to show, in multicentre trials rather than single sites. The 30% figure from radiology dates from 2020; a repeat of that survey would show whether the gap he describes is closing.

Listen to the episode

Why Doctors Are Choosing AI Chatbots with Jan Beger, GE Healthcare

Listen · 14 min