Why AI struggles to reach the hospital bedside: European experts map the barriers

Artificial intelligence has produced a steady stream of algorithms that promise to improve diagnosis in radiology, pathology, oncology and intensive care, yet only a small proportion of European hospitals actually use such systems in routine care. A qualitative study from the European CAIDX network, published in Frontiers in Digital Health, set out to explain that gap. Researchers conducted systematic expert interviews with 55 stakeholders from four European countries working across the healthcare sector, exploring what clinicians, hospital managers, IT staff, vendors and legal experts perceive as the added value of AI, the barriers to its implementation, the usability requirements it must meet, and its potential health-economic implications.

The interviews were conducted face to face between November 2023 and April 2024 at five sites within the Clinical AI-based Diagnostics project, which is funded by the Interreg Baltic Sea Region programme of the European Union and brings together thirteen partners, including five hospitals, from Denmark, Poland, Germany, Estonia, Finland and Sweden. Twenty participants were interviewed in Aalborg, six in Aarhus, nine in Lund, sixteen in Rostock and four in Wrocław. Participants were recruited purposively based on their expertise, availability and professional connections, using word-of-mouth sampling, and included AI developers, tool providers, clinical staff and members of hospital management, legal and IT departments. A shared semi-structured interview guide was used by trained moderators, and the recordings were transcribed, anonymised and analysed with thematic content analysis.

The analysis identified five overarching topics: stakeholders and teamwork, the added value of AI, lessons learned, needs for optimisation, and features specific to AI. Within the first topic, the researchers found several recurrent sources of tension. Clinicians were often treated as scarce resources whose time should be protected, which sometimes excluded them from early development and procurement decisions, yet insufficient clinical involvement was itself cited as a reason why some products failed to meet everyday needs or lacked maturity. One managing director interviewed for the study noted that when physicians are invited to discuss AI, it is often the head physician or someone in research who shows up, rather than the people who would use the tool every day, such as staff in the emergency room.

A second tension arose because different professional groups operate with different expectations. Clinical users generally favoured stable, reliable tools that fit established routines, while IT staff and vendors worked in environments characterised by ongoing updates, legal adjustments and technical iteration. Responsibilities were also frequently fragmented: legal, IT security and data protection units had substantial influence over implementation, but accountability for the overall process was not always clearly assigned. Because many AI systems require integration with hospital infrastructure or the transfer of patient data, uncertainty about privacy, cybersecurity and system integrity often slowed implementation. A fourth tension was the fear of investing time and effort in projects that might never reach routine use, since promising applications could fail due to high costs, poor workflow integration, unmet data requirements, technical incompatibilities or regulatory concerns, and unsuccessful attempts consumed resources and reduced motivation for future initiatives.

The interviews also revealed how the sheer structure of hospitals complicates adoption. Implementing AI involves a broad range of internal stakeholders, including physicians, IT departments, procurement, legal services, data protection officers, finance departments and management, along with external stakeholders such as commercial vendors and research partners. Organisational structures varied greatly, even within a single country, and procurement pathways were often described as opaque, inconsistent and difficult to reconstruct, even by those directly involved, so different actors within the same hospital sometimes had different understandings of how implementation actually worked. Respondents reported that implementation commonly took many months and could extend beyond a year, similar to other medical technologies, with the expectation of increasing complexity under evolving AI regulations.

When it came to perceived value, respondents associated AI adoption with both broad innovation goals and specific expected benefits. Hospitals sometimes pursued AI because it signals institutional modernity, but acquiring it was usually accompanied by expectations of concrete added value for patients, clinical users and institutions. The most common aspiration was improved outcomes combined with reduced workload. Clinicians emphasised AI’s potential to improve the quality and consistency of care while freeing time for direct patient contact, particularly in contexts of high workload, fatigue, staff shortages or time-critical decision-making such as emergency care and night shifts. Some believed AI had already approached or exceeded human performance in narrowly defined tasks, and others highlighted its value as a second reader or safety layer for diagnostic support. Economic value was discussed at the level of the individual clinician, the hospital and the wider healthcare system, with themes including time savings, workload reduction and improved throughput, and one urologist observed that the cost of not using a useful tool can also be shown to payers as a way of saving money.

On the technical side, hospital IT systems were described as heterogeneous, difficult to navigate and often incompatible with external solutions. AI applications typically require access to data from multiple internal systems, yet this data is often neither well structured nor easily retrievable. Even when a vendor has developed and tested a product extensively, its performance in a local hospital environment may differ due to data quality, system architecture or workflow variation. Digital maturity emerged as a key issue: in less digitised systems, weak data infrastructures, insufficiently structured clinical information, missing digital signatures and limited capacity for remote work obstructed implementation, while even more digitally advanced settings struggled with data integration and standardised formats. The authors conclude that digitalisation alone is insufficient, and that data quality, interoperability and process redesign are equally important.

The study also identified features that distinguish AI from other medical technologies. AI literacy among clinicians was uneven, with many reporting only a basic understanding acquired informally rather than through structured training, and understanding levels varied markedly across individuals and countries. Many clinicians did not seek in-depth methodological knowledge and expected certified tools to be reliable, yet still wanted some transparency and control, creating a demand for layered systems that allow simple use in routine cases and deeper inspection when needed. Perceptions of AI were generally balanced rather than uniformly enthusiastic or sceptical: many were optimistic about its potential in standardised, high-volume areas and diagnostic assistance, but concerns were raised about overreliance, opacity, misuse, cybersecurity, data protection and threats to professional roles. Bias was a particularly important concern, with respondents emphasising that AI systems may reproduce the limitations of their training data and perform unevenly across populations, and that accuracy alone is insufficient to reveal such problems.

Trustworthiness, in the respondents’ view, depended on transparency, explainability, certification and oversight, yet many were unfamiliar with how certification works in practice, and some questioned whether current procedures adequately capture the unique properties of AI. Unlike conventional software that supports isolated work steps, some AI systems produce outputs that resemble autonomous diagnostic judgments, raising new questions when human and machine assessments diverge. In the discussion, the authors note that most physicians see themselves as users and do not want to verify an application’s reliability and validity, while current medical device regulations and certification rules are not well suited to AI applications, creating what they call a dangerous gap in responsibility. They point to model cards, structured documents that accompany trained machine learning models and provide benchmarked evaluations, as an increasingly promoted solution, though these are not yet widely used.

Based on the findings, the consortium issued practical recommendations for hospitals, vendors and regulators. Hospitals are advised to involve clinical users early and continuously, define responsibilities without gaps or overlaps, streamline acquisition processes, establish central innovation offices, and structure their data with quality management if they wish to train AI applications. Vendors are encouraged to provide clear information on the purpose of a tool, its added value, performance metrics, training data characteristics, validation methods, explainability, limitations and potential misuse, alongside robust interfaces and support. The EU AI Act, which came into force in 2024, is described as an important step toward ensuring patient safety and ethical standards, although the interviews predated its practical effects. The authors acknowledge limitations: physicians were surveyed more than therapists or nurses, recruitment relied on snowball sampling within the partners’ networks, which may have overrepresented AI enthusiasts, southern European countries were not represented, and the interviews did not cover the AI Act’s consequences. They conclude that successful implementation requires more than technical performance, demanding organisational readiness, clear responsibility structures, meaningful clinician involvement and trust among stakeholders.

Subject of Research: Barriers and requirements for implementing AI-based diagnostic applications in European hospitals, based on qualitative stakeholder interviews

Article Title: Requirements and key factors for the successful implementation of AI in clinical practice: a qualitative study from the European CAIDX network

Article References: Goerss, D., Østervig Byskov, C., Jonasson, S. B., Lyngholm, I., Tupikowski, K., Nalliah, S., Jovinge, S., Pikani, J., Teipel, S., & Dyrba, M. (2026). Requirements and key factors for the successful implementation of AI in clinical practice: a qualitative study from the European CAIDX network. Frontiers in Digital Health, 8, 1954586. https://doi.org/10.3389/fdgth.2026.1954586

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1954586

Keywords: artificial intelligence, clinical implementation, hospitals, qualitative research, CAIDX, healthcare stakeholders, AI literacy, interoperability, EU AI Act, digital health, procurement, trust