Putting Artificial Intelligence to Work in Healthcare

October 14, 2026

Nature Medicine

Overview

AI tools are entering healthcare rapidly, but evidence that implementation improves outcomes for patients, clinicians, or health systems remains sparse. The article argues that clinical value depends on workflow integration, adaptation to local settings, safety evaluation, ongoing monitoring, and engagement with clinicians and patients—not model performance alone.

Background

Machine-learning and generative AI tools are being applied to activities such as medical-image interpretation, clinical documentation, and decision support. The article describes rapid deployment alongside a gap between enthusiasm and perceived usefulness: a cited survey found that more than 90% of US health systems had begun deploying AI tools, while no more than 53% considered them highly useful. Retrospective metrics and reader studies do not by themselves establish value in clinical practice. Implementation also occurs across healthcare systems with differing resources, data infrastructure, care models, and workforce capacity.

Data Highlights

Measure or example	Reported finding	Source context
US health-system AI deployment	More than 90% had initiated deployment for activities including radiology and clinical documentation.	Survey cited in the article
Perceived usefulness	No more than 53% considered the tools highly useful.	Same survey cited in the article
Deployment in China	More than 800 healthcare institutions had deployed DeepSeek since its release.	Article context
Clinician engagement	Engagement with an AI-based clinical decision support tool diminished over a 4-week pilot, alongside declining adoption.	Pilot assessment described in the article
Cross-country implementation	The same diabetic retinopathy AI system required markedly different implementation strategies in India, Thailand, and Australia.	Tiwari et al., as summarized in the article

Key Findings

AI performance in retrospective evaluations or reader studies does not establish clinical value unless the tool is incorporated into clinical workflows.
Yan et al.'s experience with an AI eye-care system in China indicated feasibility only when the technology was integrated throughout the care pathway.
Kabukye et al. highlighted the need to re-engineer AI as locally responsive, resource-conscious solutions in low- and middle-income countries, where computational constraints and specialist shortages are relevant.
Tiwari et al. reported that implementation of an identical diabetic retinopathy AI system in India, Thailand, and Australia required different strategies reflecting local priorities and workflows.
Research and real-world evaluations have identified inaccurate or misleading medical advice from large language models and safety concerns in applications including mental health; Lång argues that safety should be established before potential clinical benefits are assessed, particularly in screening.
Shah et al. reported that model iterations and changes in electronic-record documentation could contribute to performance drift after deployment; the article also describes declining clinician engagement and adoption during a 4-week pilot.

Clinical Implications

The article describes implementation as requiring prospective, real-world evaluation of effects on clinical outcomes, care delivery, costs, workflows, and staff experience, alongside attention to patient perspectives. It also identifies continued post-deployment oversight and governance that assigns responsibility for monitoring performance, addressing emerging risks, and implementing corrective measures as necessary considerations.

Conclusion

The article presents clinical AI implementation as a continuing process of assessment before, during, and after deployment. Its account emphasizes that workflow fit, local adaptation, safety, monitoring, and trust from clinicians and patients are central considerations alongside technical performance.