Health AI goes clinical.

Clinical AI is no longer a lab experiment. It sits in the consultation room, in the radiology reading list and in the electronic record. In 2026, 81% of US physicians used AI in their professional work, against 38% in 2023, and the average physician now uses it for more than two tasks (AMA, March 2026). The shift is broad, fast and mostly invisible to patients. It is also about to meet a fixed rulebook.

81% of US physicians used AI in their professional work in 2026, up from 38% in 2023 (AMA, Mar 2026)

The first wave is paperwork

The most common uses are not diagnosis. They are summaries of medical research (39% of US physicians), care plans and discharge notes (30%), charting and visit notes (28%) and chart summaries (28%). Assistive diagnosis comes later, at 17% (AMA). Hospitals follow the same order: 71% of US acute-care hospitals ran predictive AI inside their electronic health record in 2024 (ONC via AHA).

The reason is time. US physicians spend 5.8 hours in the electronic record for every 8 hours of scheduled patient time (JGIM, 2024). Employed physicians in Germany spend about three hours a day on administration and documentation (Marburger Bund, 2024). In one large US programme, ambient scribes saved physicians about an hour a day at the keyboard (NEJM Catalyst via AMA).

The workforce math makes this urgent. The World Health Organization projects a shortfall of 11.1 million health workers by 2030. AI that returns time to care is not a luxury purchase; it is capacity planning.

Europe spends fast, Germany adopts slowly

Health is the fastest-growing AI spender in Europe, at 41% a year according to IDC. Adoption on the ward is uneven. In the UK, 29% of doctors used at least one AI system in the past year (GMC, February 2025). Across the WHO European Region, 32 of 50 responding states use AI-assisted diagnostics, yet only 8% have a national health-AI strategy and 86% name legal uncertainty as a barrier (WHO Europe, November 2025).

Germany shows the gap between infrastructure and use. 18% of hospital physicians use AI, up from 9% in 2022, and 15% of practice physicians (Bitkom and Hartmannbund, May 2025). About 73 million electronic patient records exist, but 3.6% are actively used by patients. The country also runs a working market for prescribable apps: 58 digital health applications were listed at the end of 2025. Our German Health brief maps these building blocks in detail.

22 AI offerings, seven of them clinical scribes

The Good AI Prism places 22 AI offerings and three provider records (Isomorphic Labs, Owkin, Tempus) in Health & Life Sciences, each recorded with a dated source from the provider's own pages. Seven capture the conversation between clinician and patient and write the documentation: Abridge, Corti, Dragon Copilot from Microsoft, Heidi, Nabla, Suki and Tandem Health. Six read images and pathology slides, among them Aidoc, Siemens Healthineers' AI-Rad Companion, Brainomix, Gleamer, Lunit and Paige. Six support drug discovery and biology, from Google DeepMind's AlphaFold and Isomorphic Labs to Insilico Medicine, Owkin, Recursion and BioNTech's InstaDeep. Three run clinical trials and precision data, two assess symptoms and one plans hospital capacity.

The split matters because the rules follow the job, not the brand. A scribe that writes a note is treated differently from a model that flags a stroke on a scan. Open the Health & Life Sciences sector brief for the full map by job and the rule that applies to each group.

The rulebook has dates

The US Food and Drug Administration lists 1,614 AI-enabled medical devices, 76% of them in radiology. In the European Union, the AI Act now sets the timetable, as amended by the Digital Omnibus Regulation (EU) 2026/1744, in force since 27 July 2026:

1,614 AI-enabled medical devices on the FDA list, 76% of them in radiology (FDA, Sep 2026)

• Since 2 August 2026, patients must be told when they interact with an AI system (Art. 50).
• From 2 December 2027, EU high-risk duties apply to emergency triage of patients and to eligibility for public health services, both listed in Annex III.
• From 2 August 2028, AI in medical devices and in-vitro diagnostics that need a notified body carries high-risk duties under Annex I, alongside MDR and IVDR.

Other jurisdictions move in parallel. 22 US states enacted health AI laws in the first half of 2026, many of them on disclosure of generative AI in patient communication. Brazil's Federal Council of Medicine requires AI committees and patient information since 10 August 2026. Germany's 2025 Medical Assembly put it in one line: AI supports decisions, final responsibility stays with the physician.

Beyond the clinic: care and drug discovery

The same pressure runs through long-term care. The world will count 2.1 billion people aged 60 and over by 2050 (WHO). Our study AI for Care in Ageing Societies shows how health systems use AI and robotics to give time back to care, with separate US and German editions.

In research, the first drugs designed with generative AI have reached Phase 3 trials (Generate, December 2025; Insilico, September 2026). In breast screening, the MASAI trial reported 44% less screen-reading workload with AI support and non-inferior interval cancer rates (The Lancet, 2026). The Pharma, Biotech and Longevity study follows the evidence and the money.

How to judge a vendor's evidence

Health buyers face a market where every provider claims accuracy and few publish what a procurement team needs. Four questions separate a credible offer from a demo. First, does the provider show clinical evidence for a population like the own patients, not only a benchmark dataset? Second, is there a predetermined plan for model changes, as the FDA now accepts for updates, and how fast can a change be rolled back? Third, where is the data hosted, does the provider train on customer data, and do GDPR, HIPAA or the European Health Data Space rules on secondary use apply? Fourth, who reviews every output before it reaches a patient or the record? Write the answers into the contract, not into meeting notes. The AI vendor claims article shows how to read provider statements against evidence.

What the evidence still lacks

Most published evidence measures accuracy, not time returned or patient outcomes. That is the gap a buyer has to close inside the own organisation. Saved minutes count only when they demonstrably go into patient contact, quality or access. A tool that is accurate in a vendor study may not fit the own patient population, and a model update can change its behaviour after go-live.

In practice

Use case: ambient scribes. In one large US programme, ambient AI scribes saved physicians about an hour a day at the keyboard (NEJM Catalyst via AMA). The gain showed up as time. Whether that time reaches patients is the measure a hospital still has to add to its own pilot.

Best practice: test against the old workflow. The MASAI breast-screening trial randomised AI-supported reading against standard double reading and reported 44% less screen-reading workload with non-inferior interval cancer rates (The Lancet, 2026). It measured workload and patient outcome together, the design a hospital pilot can copy at a smaller scale.

What to do Monday

1. List every AI use, including scribes and chat tools clinicians bring themselves. An AI Estate Audit makes the list complete.
2. Classify by intended purpose. Separate documentation from diagnosis and triage, and map each against device law and the AI Act dates.
3. Name a clinical owner for every AI output that reaches a patient or the record.
4. Pilot documentation first, with a baseline of documentation time, quality and incident reports measured before go-live.
5. Decide on measured outcomes. Scale only on time returned to care, quality and patient outcomes; stop when the threshold is missed.

This article describes market and regulation. It is not medical or legal advice.