Artificial intelligence in the ward: Two paths to adoption in clinical practice

Artificial intelligence (AI) is poised to revolutionize medicine with algorithms that can diagnose, predict, and personalize treatments. And yet its adoption in clinical settings is hindered by technological, organizational, and cultural barriers. In recent years, many studies have examined the factors slowing down AI implementation in healthcare, but large-scale empirical research providing a concrete snapshot of actual adoption, particularly at the regional level, remains scarce.

This is the gap addressed by the study, conducted with the support of Regione Lombardia and Assolombarda. Two questions guided the research:

• To what extent, and how, is artificial intelligence already being used in Lombardy’s healthcare facilities?
• What organizational models are emerging?

From December 2023 to June 2024, researchers collected data from 46 Lombardy-based healthcare organizations (31 public and 15 private) covering over 75% of the region’s facilities. The survey focused on technologies based on supervised machine learning, with no reported cases of generative AI adoption at the time.

The questionnaire was divided into three sections:

1. Mapping AI applications in use: including clinical areas, data types, and supported functions (diagnosis, prognosis, treatment).
2. Organizational governance of AI: dedicated structures, internal training, allocated resources, and partnerships.
3. Perceived barriers to adoption: from regulatory and financial challenges to corporate culture.

The sample identified 56 AI applications in use across 20 organizations. More than one-third of these were already in routine use, while the rest were in development or pilot phases.

Most of the applications support diagnostic functions. Prognostic tools are less common but still significant, particularly in radiology, diabetology, and oncology. AI never replaces clinical judgment but rather supports decision-making.

Among the adopters (43% of the sample), two distinct organizational groups emerge:

1. Developers (13%): facilities (mostly Scientific Institutes for Research, Hospitalization and Healthcare - IRCCSs)) that design AI solutions internally or with tech partners. They are more active in prognostic applications and have further developed governance structures and in-house expertise.
2. Buyers (30%): hospitals that purchase CE-marked commercial technologies, often integrated into medical devices and primarily used in diagnostic imaging.

The perceived barriers vary significantly across groups. Developers are mostly concerned about interoperability, data management, and privacy. Buyers cite corporate culture and lack of clinical evidence as key obstacles. Non-adopters, meanwhile, point to a lack of internal expertise.

The developer group is small but carries significant weight: five IRCCSs and one Local Health and Social Care Agency (ASST). These organizations make long-term investments and retain critical skills in-house. Over 65% have created dedicated AI units, and 83% maintain ongoing collaborations with universities or external providers. Their 27 applications are mostly still in development, often focusing on prognosis, especially in neurology and rehabilitation. As of June 2024, none of the internally developed tools were in routine use, underscoring the long timelines involved. Still, this is a strategic investment that can generate value by tailoring solutions to local needs and eventually exploring commercialization.

The buyer group is larger but less structured. It includes ASSTs, IRCCSs, outpatient clinics, and nursing homes. None of these organizations has established dedicated AI units, though many have created informal teams. The 29 AI applications in use are mostly integrated into CE-marked devices, primarily for assisted diagnosis in radiology. About 59% were already in use or in advanced setup stages. This approach enables faster implementation but demands strong internal capacity for technology assessment, selection, and integration with existing workflows.

Ultimately, adoption strategies appear aligned with the nature of the institutions involved: IRCCSs, with their research orientation, act as early movers in partnership with universities and private firms. Medium-sized, non-research-focused facilities recognize the importance of AI but prefer to engage with the market.

In a rapidly evolving landscape, the next frontiers are AI applications for treatment (early examples are already emerging in Lombardy) and the transition to generative AI.

Ardito V, Cappellaro G, Compagni A, Petracca F, Preti LM. “Adoption of artificial intelligence applications in clinical practice: Insights from a Survey of Healthcare Organizations in Lombardy, Italy.” Digital Health. 2025;11. DOI: https://doi.org/10.1177/20552076251355680.