Frictions in AI implementation in specialised Norwegian healthcare: An interview-based study
Research paper, SCIS 2025 (Oslo, Norway)
AI technologies are envisioned as a solution to current societal challenges, including in the healthcare sector. However, their implementation in real-world hospital settings remains limited. In this paper we examine the slow implementation of AI in healthcare as a problem of installed base constraints. AI technologies are fundamentally different from traditional medical information technology. However, existing implementation practices are still shaped by regulations, development and implementation approaches established for conventional medical information technologies, constituting the installed base. We examine these constraints as frictions between novel AI technologies and the installed base through an interview-based study with 14 interviews with 11 experts in the Norwegian healthcare sector. We address the following research question: How do frictions between novel AI technologies and the existing sociotechnical installed base shape the process of AI implementation in public hospitals? Our findings reveal seven key frictions relating to data complexities, hospital information infrastructures, regulatory landscape, resource requirements, accountability, implementation strategies, and incentive structures. We contribute to understanding the shift needed to support the implementation and use of AI technologies in healthcare.
AI, healthcare, hospitals, installed base, frictions
In het kort
Dit artikel onderzoekt waarom AI zo traag zijn weg vindt naar de dagelijkse praktijk in Noorse ziekenhuizen. We bekijken dit als een botsing tussen een nieuwe technologie en de bestaande “installed base”: het geheel van systemen, werkwijzen, regels en organisatiestructuren dat ooit voor conventionele medische IT is ingericht. De spanningen die daarbij ontstaan noemen we fricties. Het materiaal bestaat uit kwalitatieve interviews met experts uit de Noorse gezondheidszorg die betrokken zijn bij AI-projecten in ziekenhuizen.
We vinden zeven fricties, elk veroorzaakt door een kenmerk waarin AI verschilt van eerdere technologie:
- Datacomplexiteit. AI ontleent zijn functionaliteit aan data, terwijl gewone IT zonder data gebouwd kan worden. Data voor zeldzame aandoeningen is schaars, annotatie kost veel tijd, en een Amerikaans model presteerde slecht op Noorse CT-scans omdat Noorse protocollen een lagere stralingsdosis gebruiken. Toegang tot data is in de praktijk de grootste drempel.
- Ziekenhuisinfrastructuur. Systemen zijn als silo’s gebouwd en praten slecht met elkaar. IT-ondersteuning is regionaal gecentraliseerd en staat ver van de kliniek. Monitoring van AI-prestaties na ingebruikname is nauwelijks geregeld, hoewel de MDR en de AI Act dat vereisen.
- Regelgeving. De MDR en CE-markering zijn gemaakt voor statische technologie en grote bedrijven. Definities van AI in de regelgeving zijn zo vaag dat ze op alles van toepassing zijn, en instellingen leggen regels verschillend uit.
- Middelen. Er is geld voor onderzoek, maar nauwelijks voor implementatie. Ziekenhuizen missen software-engineers en datawetenschappers, rekenkracht in de cloud is duur, en technology transfer offices zijn onderbemand.
- Verantwoordelijkheid. Zorgprofessionals dragen een persoonlijke verantwoordelijkheid die moeilijk te rijmen is met het ondoorzichtige karakter van AI. Het is onduidelijk wie aansprakelijk is bij een foute diagnose en op welke data modellen zijn getraind.
- Implementatiestrategie. Ziekenhuizen denken in “zelf bouwen of kopen”, maar ook gekochte modellen moeten lokaal worden bijgesteld, en zelf gebouwde modellen moeten alsnog door de certificering. In de praktijk ontstaat daarom een hybride aanpak.
- Prikkels. Onderzoekers worden beloond voor publicaties, niet voor klinische toepassing. Clinici zien AI soms als extra werk in plaats van minder. De voorwaarden van technology transfer offices maken commercialisering onaantrekkelijk.
In de discussie laten we zien dat fricties twee kanten op werken: de installed base remt AI, maar AI zet ziekenhuizen ook aan om infrastructuur en regels te herzien. Welke frictie het zwaarst weegt, hangt af van de fase waarin een project zit: vroeg vooral data, later vooral regelgeving. Fricties hebben bovendien een dubbel karakter. Strenge regels voor verantwoordelijkheid en gegevensbescherming vertragen nu, maar kunnen later juist bijdragen aan betrouwbare AI. Op dit moment werken de fricties vooral remmend, wat past bij de vroege fase waarin AI in ziekenhuizen zich bevindt. De Noorse uitgangspositie, met vroege invoering van elektronische patiëntendossiers, uitgebreide registers en een publiek zorgstelsel, kan op termijn een voordeel worden.
Voor de praktijk stellen we voor de zeven fricties te gebruiken om de eigen organisatie door te lichten, fricties niet alleen als obstakels te zien, en op nationaal niveau te werken aan een uniforme uitleg van dataregels en flexibelere financiering voor de stap van onderzoek naar implementatie. Vervolgonderzoek bestaat uit casestudies van twee AI-projecten: detectie van hersenbloedingen en van een acuut hartinfarct.
1 Introduction
AI technologies are expected to address various contemporary societal challenges, with healthcare being a significant area of focus. AI holds significant promise in multiple aspects of healthcare, including automating tasks currently performed by healthcare professionals, supporting clinical decisions, improving clinical work efficiency, and diagnostic outcomes (Kaul et al., 2020). Currently, developments within medical AI are based on Machine Learning (ML) or Deep Learning (DL) technologies, mostly used for image analysis and diagnostics. Such developments position the medical areas of radiology, pathology, dermatology, and ophthalmology at the forefront of ongoing advancements (Rajpurkar et al., 2022, p. 31; Wang et al., 2019, p. 293). However, despite these prospects, the implementation of AI into real-world healthcare settings is still limited, in its early stages, and large-scale implementation is progressing slowly (Aristidou et al., 2022; Davenport & Kalakota, 2019; Sharma et al., 2022).
To understand why AI implementation processes in healthcare are holding back, we take a sociotechnical installed-based perspective (Aanestad et al., 2017). We aim to bring attention to the pre-existing, built environment of practices, conventions, tools, and systems (Aanestad et al., 2017) affecting the implementation of AI technologies. This perspective has been shown to be valuable in the study of IT implementation in healthcare (e.g., Aanestad & Jensen, 2011; Aanestad et al., 2017; Ellingsen et al., 2022), as well as in other domains (e.g., Bygstad, 2010; Hustad et al., 2020). For instance, it has shown how the quality of the installed base influences design decisions (Grisot & Vassilakopoulou, 2015), how the installed base can be cultivated according to different strategies (Aanestad & Jensen, 2011; Grisot et al., 2014; Vassilakopoulou & Marmaras, 2017), and how the installed base can strongly oppose innovation (Klein & Schellhammer, 2017).
Specifically, in this study, we are interested in understanding how the existing sociotechnical installed base both enables and constrains AI implementation in healthcare. More specifically, to understand the constraining and enabling aspects of the sociotechnical installed base, we make use of the analytical concept of friction and focus our analysis on how specific characteristics of novel AI technologies trigger frictions during interaction with the existing sociotechnical installed base (i.e., regulations, IT infrastructure, practices). Frictions are active forces with a tendency to favour existing values (Håkansson & Waluszewski, 2011) and can indicate how existing logics endure (Vassilakopoulou et al., 2019). Based on these premises, we address the following research question:
How do frictions between novel AI technologies and the existing sociotechnical installed base shape the process of AI implementation in public hospitals?
To answer this question, we have researched the introduction of AI technologies in Norwegian hospitals (i.e., the specialist healthcare services). We conducted an exploratory study in the spring and autumn of 2024 to understand the current situation of AI development, implementation, and use in specialised care in Norway. The study consisted of interviews with various stakeholders in different job positions. The interviews focused on the stakeholders’ experiences with AI implementation in hospitals, their views on the current approaches to AI implementation, the existing challenges, and possible improvements.
Our findings reveal how the implementation of AI technologies gives rise to various frictions as novel AI technologies interact with the existing sociotechnical installed base. According to our findings, these frictions emerge in relation to data complexities, hospital information infrastructures, regulatory frameworks, resource requirements, accountability concerns, implementation strategies, and incentive structures. Through our analysis, we show how specific characteristics, unique to AI technologies, act as triggers for these frictions when they meet established healthcare practices and systems.
Our findings contribute to the literature on AI implementation in healthcare by proposing to understand the AI implementation challenges in hospitals as frictions between the novel characteristics of AI technologies and the installed base, and by providing a detailed analysis of how these frictions manifest. An understanding of the AI implementation ‘gap’ as frictions suggests that these frictions can serve as both barriers and catalysts, requiring practitioners to actively negotiate, adapt, and transform existing practices as they integrate novel AI solutions into the existing sociotechnical environment.
The structure of the paper is as follows: First, we position this study in relation to existing literature on AI implementation and present our own conceptual framing. Next, we present the context of the study and the research methodology. Finally, we present our preliminary findings and discuss implications and future research plans.
3 Theory: Sociotechnical installed base and frictions
In this paper, we make use of the concept of sociotechnical installed base (Aanestad et al., 2017) to understand the process of AI implementation in healthcare. This is a core concept within the information infrastructure perspective (e.g., Hanseth & Lyytinen, 2010; Grisot et al., 2014; Grisot & Vassilakopoulou, 2017), which draws attention to how existing work practices, technologies, and regulations both enable and constrain digital infrastructure innovation. The focus on the role of the installed base in innovation processes brings attention to the inertia of such processes in realising change and the opportunities for change (Aanestad et al., 2017).
The installed base comprises ‘all that there is’ (Aanestad et al., 2017, p. 28), including current tools, routines, divisions of labour, regulations, and other sociotechnical elements. Hanseth and Lyytinen (2010) define it as the set of ICT capabilities with their users, operations, and design communities. Lanzara (2014) adds the importance of considering existing institutional and organisational components as part of the installed base. In this perspective, infrastructures ‘evolve’ rather than being designed from scratch. As Star and Ruhleder (1996) point out, “infrastructure does not grow de novo, it wrestles with the inertia of the installed base and inherits strengths and limitations from that base” (p. 113). This evolutionary perspective on technology makes the installed base both enabling and constraining innovation, as well as the existing sociotechnical arrangement being transformed by innovation; managing the installed base means building on it and transforming it at the same time (Aanestad et al., 2017, p. 30). In relation to AI, this means for instance that AI needs to fit into the established arrangements while at the same time reshaping them.
To analyse the relation between the installed base and the novel AI technology, we use the concept of frictions from innovation studies (Håkansson & Waluszewski, 2011; Hoholm & Olsen, 2012). Håkansson and Waluszewski (2011) observe that technological systems are difficult to change because different established arrangements become ‘cemented’ upon each other, making alterations and novelty difficult to establish. They define frictions as, “How an alteration force applied to one resource is transferred to resources it interacts with and how this friction can act both as a stabilizer and a de-stabilizer of existing resource interfaces” (pp. 176-177). As their research is within the field of industrial marketing, their understanding of frictions pertains to the resistance organisations encounter as they try to adapt and change within networks of institutional relationships. However, they also observe how cemented arrangements can transform through friction (Håkansson & Waluszewski, 2011). Friction is thus conceptualised as the manifestation of a problem and as an ‘active force’ causing changes. They point out that it is “a force that relates what happens today with what has happened earlier” (Håkansson & Waluszewski, 2011, p. 179), connecting current innovations with prior arrangements and investments.
Accordingly, friction as a phenomenon has three key characteristics. First, it is relational, as it only appears “when a force is directed towards two interacting surfaces” (Håkansson & Waluszewski, 2011, p. 176). Friction occurs when different resources or entities interact, and forces attempt to create movement or change between them. Second, it is time-dependent, since the same force can have different effects depending on when it manifests (p. 176). The timing of when a force is introduced can change how friction manifests and impacts the interacting surfaces. Third, friction affects the features of the interacting resources, producing a ‘coordinated’ movement and transformation of these interacting resources (p. 178). This means friction resists change and actively reshapes the innovation and the installed base it encounters.
Friction also displays a ‘dual’ nature, which is essential when understanding it in the context of technological change. Nowotny (1993) calls friction a “janus-faced phenomenon” (p. 40), as it can have both a stabilising and destabilising effect. In other words, friction can simultaneously resist change by preserving the installed base, and enable change by transforming existing arrangements into new combinations (Håkansson & Waluszewski, 2011, p. 180). Hoholm and Olsen (2012) further expand this concept of frictions by examining innovation as consisting of two opposing forces: mobilising processes (i.e., the process of mobilising resources, activities, and actors to secure commitments), and explorative processes (i.e., the process of learning that creates new or revised understandings about what the innovation can achieve). Building on Actor-Network Theory (ANT) (e.g., Latour, 2007; Callon, 1984), they point out that innovation networks do not simply appear but emerge through the interactions and negotiations of these two opposing forces. Examining infrastructural shifts in eHealth, Vassilakopoulou et al. (2019) show how frictions result in the perpetuation of elements of the past during change processes.
In this paper, we use the concept of friction to analyse the encounter of novel AI technologies, with their specific characteristics, and the sociotechnical installed base, including the established healthcare practices and systems designed for different technologies.
4 Research methodology
The main purpose of this study is to investigate how specific characteristics of AI technologies trigger frictions in interaction with the existing sociotechnical installed base in Norwegian hospitals. The Norwegian healthcare system is public and organised into four Regional Health Authorities (“Regionalt helseforetak”, RHF). Hospitals in the same local region are grouped together into Health Trusts (“Helseforetak”, HF). As mentioned in the introduction, the number of realised AI deployments in real-world hospital settings is currently limited. According to a report from the Norwegian Directorate of Health (Helsedirektoratet et al., 2022), each of the four Regional Health Authorities is utilising medical AI in their hospitals. However, only two out of the five projects mentioned in the entire report seem to concern use in actual hospital clinical practice and not just as pilot projects. Most of the projects referred to in the report are either research projects, innovation projects, or trial projects testing out specific technologies (Helsedirektoratet et al., 2022, pp. 11–14). In August of 2023, a large Health Trust in Norway made headlines in several national newspapers as the first Health Trust to take AI into clinical use in the image diagnostic department of their four hospitals for detecting bone fractures (Sundby, 2023).
4.1 Data collection and analysis
The study is designed as an interview-based study following an interpretive research approach (Walsham, 2006). We conducted 14 qualitative semi-structured interviews with 11 experts in the Norwegian healthcare sector. The informants were selected for being established actors in AI and healthcare and having experience from working with AI implementation in Norwegian hospitals, either in technology development, project management, or policy development. Two informants (#7 and #9) were interviewed multiple times as they held management positions overseeing numerous AI projects at major hospitals, providing broader insights than could be captured in a single session. The roles of the informants are described in Table 1.
| Informant ID | Role in the Norwegian healthcare sector | Number of interviews |
|---|---|---|
| 1 | Relevant management function at a Regional Health Authority | 1 |
| 2 | Part of the deployment team at a procurement project | 1 |
| 3 | Involved with research and development of several projects | 1 |
| 4 | Working in certification and classification of medical AI | 1 |
| 5 | Involved with research, development and procurement projects | 1 |
| 6 | Part of the deployment team at a procurement project | 1 |
| 7 | Manager of AI projects at a major hospital | 2 |
| 8 | Working for the Norwegian government on medical AI | 1 |
| 9 | Manager of AI projects at a major hospital | 3 |
| 10 | Developer in several medical AI projects at a major hospital | 1 |
| 11 | Developer in several medical AI projects at a major hospital | 1 |
The interviews were semi-structured and followed an interview guide. The informants were asked about their background, their role in relevant past and current projects, their experiences with AI implementation in hospitals, their views on the current approaches to AI implementation, as well as their views on how the situation could be improved. Interviews lasted around one hour each and were conducted both in person and online, according to the interviewee’s preference. The study adhered to national guidelines for research ethics and received approval from SIKT1. All informants were informed about the study’s scope and objectives and provided their written consent prior to participating. All interviews were audio recorded with the informants’ consent and then fully transcribed. To protect confidentiality and ensure privacy, all data was anonymised, with personal identifiers removed or altered without compromising the integrity of the data.
The first author conducted the data analysis of the interview transcripts, aiming to understand the informants’ views on the challenges of AI implementation based on their experiences. The analysis followed a thematic analysis approach in which recurring topics were identified and coded throughout the transcripts, and followed Braun and Clarke’s (2006, pp. 86–93) six-step method: initial familiarising by re-reading the transcripts, noting down initial lower-level codes, searching for themes in these codes, reviewing, defining themes, and producing the analysis. This process enabled the identification of recurring themes related to how the informants viewed the challenges of implementing AI. The analysis process resulted into the following overall themes: data complexities, hospital information infrastructure, regulatory landscape, resource requirements, accountability, implementation strategies, and incentive structures. As a next step, we used the theoretical lens of friction towards the categories and identified specific AI characteristics as friction triggers.
5 Findings: Frictions in AI implementation
In this section, we present our findings structured according to the key themes that emerged from the thematic analysis process: data complexities, hospital information infrastructure, regulatory landscape, resource requirements, accountability, implementation strategies, and incentive structures. Our findings show that frictions are visible when novel AI technologies, with their specific characteristics, interact with established healthcare practices and systems designed for different technologies. These frictions are not just obstacles, but active forces that shape how AI is implemented, potentially leading to changes in the technology and the existing healthcare arrangements. By analysing AI implementation through the lens of frictions with the installed base, we gain a more nuanced understanding of how the specific characteristics of AI technologies interact with existing healthcare systems, leading to unique implementation challenges.
5.1 Data complexities
AI’s reliance on data distinguishes it from traditional medical technologies. While other medical technologies can be designed and implemented with fixed functionality largely independent of data, AI systems derive their very functionality from data. This dependence requires a novel interplay between the data-dependent nature of AI, the existing data landscape in Norwegian hospitals, and established hospital practices related to data acquisition and management. Our informants highlighted the challenges arising from this friction. As one informant explained: “There is something about Artificial Intelligence, where we create functionality based on data. And it’s actually quite unusual, because in normal IT systems, you can create the system and functionality without data” (#8).
This fundamental difference requires a shift in thinking about the role of data, as well as a shift in data practices, in how technology is developed and deployed in healthcare.
In addition, unlike other medical technologies with predetermined rules, AI algorithms learn from data, requiring extensive datasets to identify patterns and make accurate predictions. This poses a challenge for medical AI, where data for specific conditions can be scarce. The issue is further complicated by the need for high-quality, annotated data. One developer explained the difficulty of collecting enough data for rare conditions like glioma tumours: “Every year you may only have less than 100 cases […] so you have to take time to collect enough cases” (#10).
Such constraints make it challenging to guarantee that training data is both large enough and consistent enough to support robust AI models. The type of data also plays an important role in AI performance. Different imaging protocols, data formats, and variability within patient populations can all impact how well an AI model performs. As one informant recounted: “We had this commercial tool […] for [stroke] testing, part of a big research project, and we found that it didn’t really work well on our data. And nobody understood why until we realised that the protocol in Norway used low-dose CT. So we used a much lower dose than what these models were trained on. […] Then you have different texture in the images, and it just performed really, really badly. And that was like a shock to this American company because they’ve never seen data like this before.” (#7).
This incident highlights the importance of not only data quantity but also its compatibility. Such mismatches emphasise the need for careful consideration of data characteristics and the need for fine-tuning models with local data. Informants also emphasised that gaining access to data is, in practice, the main barrier to AI development in Norwegian specialised healthcare, as the processes for data access (and privacy oversight) can vary significantly between hospitals. Also, several informants noted that using data from different sources (e.g., medical images, patient records, lab results, genomic data) for multimodal AI poses additional technical and regulatory difficulties. Hospitals often have large registries and rich data sets, but bridging these silos requires not only technical expertise but also consistent interpretation of data protection rules across institutions. As one informant pointed out: “Different data protection officers interpret the rules in different manners” (#5), highlighting that establishing standardised, national-level guidelines would really accelerate the development and implementation of medical AI.
5.2 Hospital Information Infrastructures
The implementation of AI in Norwegian hospitals also reveals friction between the demands of this technology and the existing hospital information infrastructures. These infrastructures, built largely around legacy systems designed for specific tasks and built as silos, struggle to support AI’s need for interoperability and scalability. Informants frequently pointed to this infrastructure inadequacy as a major obstacle to successful AI implementation. One informant said: “These systems do not talk to each other in a good way” (#7).
This lack of data integration impacts AI implementation, as many algorithms, especially multimodal models, rely on accessing and processing information from diverse sources. In addition, the current IT support structure for hospitals is centralised at regional level. One informant said: “They sit in a completely other location, they don’t have clinicians […], they don’t know that much about how a hospital is operated and have no domain knowledge from the clinician” (#9).
This disconnection between centralised IT support and the specific needs of hospitals further highlights the infrastructure challenges. Furthermore, one informant emphasised the dynamic and evolving nature of AI itself, stating that: “It’s not a fixed technology […] A lot of things will change when you install that technology component” (#1).
This dynamic nature of AI requires flexible infrastructures as models need to continuously learn and adapt, requiring not just initial large datasets but also ongoing access to new and evolving data sets. Several interviewees also highlighted that monitoring of AI performance over time (e.g., detecting performance drifts and data anomalies) is currently insufficiently addressed. One informant argued that: “No one is talking about monitoring… It’s even in MDR and the EU AI Act that you need to have this. They don’t even know this” (#9), highlighting how hospital IT infrastructures, currently, often lack the necessary provisions to effectively monitor AI models after deployment.
5.3 Regulatory landscape
The implementation of AI in Norwegian hospitals presents a challenge for existing regulatory frameworks. These frameworks, designed for traditional medical technologies, are now confronted with the unique characteristics of AI, leading to friction in several areas. One frequently mentioned concern revolves around the complexity and perceived burden of the Medical Device Regulation (MDR)2 and CE marking processes3. One informant described the MDR documentation requirements as “a mess,” pointing out: “It is not built for small start-up companies. It’s built for big, big companies” (#5).
This suggests that the current regulatory process, built with conventional medical technology in mind, is not well-suited for the agile, iterative development common with AI, potentially hindering smaller companies or hospital-based development initiatives. In addition, interpreting and applying existing regulations to AI poses significant challenges. AI’s data-driven nature and continuous learning capabilities introduce complexities not fully addressed by traditional frameworks. One informant said: “The biggest problem with the AI regulation is that no one even knows what AI is… The definitions they used in the regulations, […] you could apply that definition to everything” (#9). This lack of clarity makes it difficult for practitioners to navigate the regulatory landscape and to determine the best ways to development and implementation.
Another set of concerns come from the centralised nature of Norwegian healthcare IT services, which were set up before AI became as prominent as it is now. Navigating the layers of approvals, and consistent interpretations among different healthcare organisations, data protection officers, and ethical committees, can stall or impede AI projects significantly. Several interviewees highlighted the need for clearer national guidelines that standardise interpretations of data access, privacy, and security requirements.
5.4 Resource requirements
Resource limitations within the installed base create friction when implementing AI in Norwegian hospitals. The demands of AI in terms of funding, personnel, and computational power, differ from other medical technologies. Funding gaps are particularly evident in the stage of translating research into clinical implementation. This indicates the installed base’s historical emphasis on procuring finished technologies, not developing research-based solutions. One informant noted: “To get funding for implementation, there’s almost no money available at all” (#6), highlighting that academic research initiatives may secure financial support for early proof-of-concept stages, but are unable to secure the resources to move promising prototypes into actual clinical use.
In addition, the need for specialised personnel like software engineers and data scientists presents another challenge, as these roles are not typically part of existing hospital staff. One informant highlighted: “You might actually also need to have more like software engineers actually in the hospital or something like that to do this, both to like implement the algorithms but also to run them.” (#9). This emphasises the need for new skillsets not traditionally found within hospitals.
Another interviewee mentioned the high costs associated with cloud-based solutions and compute resources needed for robust AI training, noting that expenses can become large quickly, especially for institutions not accustomed to covering such continuous computational expenses.
Moreover, technology transfer processes — often managed by under-resourced Technology Transfer Offices (TTOs)4 — amplify these resource constraints. Informants pointed out that turning a promising, in-house-developed AI tool into a clinically implemented product often requires navigating a complex path, with insufficient institutional support. This discourages in-house innovation and places an excessive burden on smaller teams attempting to move from research to practice.
5.5 Accountability
The introduction of AI into healthcare settings also creates a friction point concerning accountability and trust. Unlike traditional medical technologies where the lines of responsibility are generally clear, the “black box” nature of AI algorithms, coupled with their potential to influence clinical decisions in a high-stakes environment, raises difficult questions about liability and user trust. One key concern revolves around the question of responsibility in cases of AI-driven misdiagnosis. With traditional technologies, accountability typically rests with the clinician or the manufacturer. However, AI’s autonomous nature blurs these lines. As one informant noted, the personal responsibility inherent in healthcare professions creates a sensitivity to these issues: “No other professions have the same kind of personal responsibility as healthcare professionals” (#4).
This personal responsibility clashes with the nature of accountability in AI systems, where it can be difficult to pinpoint the source of an error. This concern may be amplified by the lack of transparency in AI solutions, where, as one informant pointed out, it’s often unclear: “What kind of data they have trained on and how they have developed these models” (#11). This could further damage trust and make it challenging for practitioners to confidently incorporate AI into their workflows. At the same time, some informants also raised concerns around how fully autonomous AI might ultimately reduce clinical skills in case of overreliance, introducing additional ethical questions surrounding both patient safety as well as professional accountability.
5.6 Implementation strategies
The implementation of AI in Norwegian hospitals reveals a tension between practitioners’ existing decision-making frameworks and the characteristics of AI. Informants often approached AI implementation through the familiar lens of build-or-buy, where building refers to in-house development strategies and buying refers to the procurement of commercially available, CE-marked applications. This dichotomy, however, seems to fall short when applied to medical AI, revealing a friction between established strategic frameworks and the emerging needs of this novel technology. For instance, even when opting for commercial solutions, hospitals often see the need for customisation and fine-tuning with local data, as highlighted by one informant: “If you buy commercial AI models, I think you still need to fine-tune the model with your own hospital data, otherwise […] you have a risk to miss some detections or segmentations” (#10).
This requires a degree of in-house technical expertise and effort, effectively creating a hybrid approach even when the initial intention was to solely procure a ready-made solution. Conversely, in-house development efforts are limited by the requirement of regulatory approval (CE marking, MDR), perhaps requiring collaboration with commercial partners. This suggests that even in-house development must work within the existing regulatory framework designed primarily for commercial products, again blurring the lines between the build and buy strategies. One informant argued for a different approach: “A hybrid approach where you have a mix of both, […] in-house competence and development at hospitals and then external stuff” (#9).
Several informants also emphasised that hospitals are forced to take on “producer responsibility” if they develop AI in-house, which might entail legal liability and the need to comply with all MDR requirements. This can be overwhelming for institutions not well-equipped with large regulatory teams. Yet, purely relying on external vendors can lead to poor local performance if the vendor’s models do not match Norwegian patient data and hospital workflows. As a result, a middle ground has emerged in practice, aiming to strike a balance between external products and local adaptation.
5.7 Incentive structures
In addition to the prior themes, many of our informants pointed to another important friction: the misalignment of incentives across different stakeholders involved in AI development and implementation. Researchers, for instance, are incentivised by and rewarded for producing publications rather than turning their work into clinical practice. One interviewee pointed out: “So if you ask a researcher to spend one, maybe two, maybe three years of their career developing an application that can be used in a clinic, he or she won’t have any publications for those years. […] That’s very damaging for a research career” (#3).
Clinicians, on the other hand, may be reluctant to adopt AI tools that appear to add complexity to their already demanding workflows. As one informant noted, an AI system that provides more images to review (instead of shortening the diagnostic process) can become an extra burden: “If you have a process you’re doing at work and you add a tool, you add maybe more images given to you by AI, you actually have more images, more information to look through. It takes more time, actually” (#6).
Furthermore, the structure and resourcing of Technology Transfer Offices (TTOs) in Norway can limit commercialisation processes. One informant highlighted that the TTO agreements sometimes fail to provide attractive terms for researchers, which could discourage them from pursuing commercialisation: “Who’s ever going to do anything with that [5% ownership]? You can just forget it” (#9).
These incentive issues are connected to the previously mentioned resource requirements and implementation challenges. Well-developed AI tools might struggle to reach clinical implementation if the reward structures and work pressures push researchers, practitioners, and hospital managers away from focusing on actual clinical implementation. While incentive structures may be less visible than, for example, data or infrastructure challenges, they are an important part of the frictions that arise when attempting to integrate AI into the installed base of healthcare.
5.8 Summary
Table 2 presents a summary of the findings, provides a short description of frictions and the specific characteristics unique to AI as a technology that act as ‘friction triggers.’
| Unique AI characteristics as friction triggers | Short description | Friction theme |
|---|---|---|
| AI requires vast, high-quality data for training, development, and ongoing access for evaluation, improvement, and model adaptation. Compatibility issues, such as different imaging protocols, can impact performance. | Data are scarce, have limited quality, require time to be accumulated and prepared for AI, vary in quality, and may be incompatible with AI models trained on different datasets. Data access is also inconsistent due to varying interpretations of data protection rules. | Data complexities |
| AI requires interoperable, flexible, and adaptable infrastructures, along with ongoing monitoring to detect performance drifts and data anomalies. | Hospital IT infrastructures are siloed, lack integration, and centralised IT support is disconnected from clinical needs. Also, AI performance monitoring appears to be currently insufficient. | Hospital information infrastructures |
| AI requires regulatory frameworks that address its continuous learning, black box nature, model transparency, and standardised interpretations of compliance requirements. | Existing regulations (e.g., MDR and CE marking) are too complex for local development initiatives and were designed for static technologies. Additionally, regulatory definitions are vague, leading to inconsistent interpretations across organisations. | Regulatory landscape |
| AI requires novel expertise of technologists and health personnel, both in technical and healthcare domains, along with proper funding that goes beyond the early research phases. | Hospitals lack specialised AI personnel, including software engineers and data scientists, as well as general knowledge about AI models and functionality. They also face funding challenges for transitioning from research to clinical implementation. Additionally, Technology Transfer Offices (TTOs) are under-resourced for supporting AI implementation. | Resource requirements |
| AI’s black-box nature makes its processes opaque, disconnecting responsibility from accountability, which in turn makes it difficult to ensure clinician trust and legal clarity. | Accountability is traditionally strongly linked to professional responsibility. In healthcare, personal responsibility is deeply ingrained, but AI’s opaque decision-making makes it difficult to assign liability. The lack of transparency in AI models creates trust issues among practitioners, further complicating accountability in medical decision-making. | Accountability |
| AI implementation requires novel and flexible procurement models, along with regulatory strategies that support the entire implementation lifecycle. This approach should balance in-house expertise with external procurement to ensure successful deployment. | Current implementation strategies are based on build-or-buy models, but these existing models are inadequate. AI requires local adaptation, even when commercially sourced. Hospitals face regulatory challenges when developing in-house AI, leading to a push towards hybrid approaches. | Implementation strategies |
| AI adoption requires alignment of academic, clinical, and commercial incentives to facilitate innovation, development, and real-world implementation. | Incentives in research, clinical practice, and TTOs, do not support practical AI implementation. Researchers are incentivised to publish rather than focus on translating AI into clinical practice. Clinicians may resist AI if it increases their workload instead of reducing it. Additionally, TTO agreements often discourage commercialisation efforts, further hindering AI implementation. | Incentive structures |
6 Discussion
Our study aimed to understand how specific characteristics of AI technologies trigger frictions in interaction with the existing sociotechnical installed base, and to provide an understanding of the challenges of implementing AI in Norwegian hospitals as frictions between AI’s novel characteristics and the existing sociotechnical installed base. In this chapter, we discuss the findings based on the theoretical lens of frictions.
6.1 Friction characteristics
The frictions we identified emerge from the relation between AI technologies and the installed base of Norwegian healthcare. For example, data complexities arise because AI’s requirement for extensive, high-quality, annotated training data interfaces with hospital data practices not designed for AI. Similarly, the accountability friction emerges from the interfacing of AI’s “black box” nature and professional responsibility. Our findings also show that the frictions between AI and the installed base vary in importance across different stages of the implementation process. During early development and training stages, data complexities are perceived as the primary frictions, while later, regulatory hurdles (e.g., MDR certification and CE marking) are perceived as more important. This temporal aspect suggests that our informants reported frictions that were most relevant to their current stage in the AI development and implementation process.
Also, our findings indicate that frictions affect both of sides, i.e., the introduction of AI both shapes and is shaped by the installed base. For example, hospital information infrastructures may limit AI development and implementation, but at the same time it has prompted hospitals to revisit and reconsider their infrastructures. Similarly, the regulatory landscape is actively being adapted to accommodate AI’s unique characteristics while developers are adapting to work within existing regulatory frameworks. This mutual shaping process resembles the process of friction as it is known in physics, where it produces heat on both of the interacting surfaces. AI technologies are being adapted to fit within the Norwegian healthcare context, while the installed base evolves in response to the introduction to AI.
6.2 The dual nature of frictions
Our findings are in line with Nowotny’s (1993) concept of friction as a ‘janus-faced phenomenon’ with both stabilising and destabilising effects. For example, the Norwegian healthcare system’s robust accountability practices and regulatory requirements currently slow down AI implementation — i.e., a stabilising effect. However, these same practices may later support trustworthy AI deployment by ensuring systems meet similarly high standards — i.e., transformative potential. Similarly, the EU’s highly regulated data protection practices are perceived as constraints that cause friction in data access. But these same rigorous practices could enable Norway to develop trustworthy and high-quality AI systems due to better data governance.
Building on Nowotny’s concept, our findings also align with Hoholm and Olsen’s (2012) view of innovation as consisting of two opposing forces: mobilising processes and explorative processes. For instance, implementation strategies, like build-or-buy, reveal traditional pathways, mobilising people and resources in predictable, structured ways. However, AI implementation seems to require more explorative processes based on what works in practice, i.e., hybrid approaches that do not precisely fit into either existing strategy. Similarly, the ‘incentive structures’ friction also reveals tension between mobilising processes (focus on academic publications and clinical targets) and explorative processes (developing novel AI solutions where those incentives don’t always align).
Although frictions can both enable and constrain innovation (Håkansson & Waluszewski, 2011; Hoholm & Olsen, 2012), our findings show that they currently act mostly as a constraining force on AI projects in Norwegian hospitals. This might reflect the current early stage of AI implementation, where the installed base is not yet developed to support the specific demands of AI. However, as hospital AI matures, the same installed base could become an asset rather than a constraint. For example, Norway was among the first countries to adopt an Electronic Health Record (EHR) system and has extensive medical data registries (Folkehelseinstituttet, 2024; helsedata.no, n.d.), which could significantly support the development of medical AI. Also, the public nature of the Norwegian healthcare system, shaped by welfare state ideology, oriented towards the public good, and supported by substantial financial resources, offers a potentially strong foundation for AI development and implementation in the future.
6.3 Contributions
Our research provides practical contributions for stakeholders in the Norwegian specialised healthcare sector. First, we suggest that stakeholders use the identified friction areas to assess the implementation challenges in their organisation. Frictions should not be considered obstacles, as the two-way interaction between the two surfaces could provide unique, positive developments (e.g., strict regulatory landscapes might seem burdensome at first, but may contribute to better AI in the long run). Our analysis of the relational and time-dependent nature of frictions also suggests developing holistic strategies that consider the sociotechnical installed base and the AI implementation lifecycle, rather than addressing individual frictions in isolation. Lastly, our findings also reveal certain areas where national-level efforts and coordination could help with constraints coming from certain frictions, such as standardised interpretations of data regulations and more flexible funding mechanisms that could bridge the gap between research and implementation.
Our research also makes theoretical contributions. Our study builds on sociotechnical installed base theory by demonstrating how specific characteristics of AI as a novel technology create specific frictions with specific characteristics of the installed base of Norwegian hospitals. While previous research examined implementation challenges in healthcare infrastructure generally, our focus on AI’s unique characteristics reveals how different technologies interact differently with the same installed base (specifically, how AI interacts differently with this installed base than prior conventional medical technologies). Our use of the concept of friction also contributes to a more nuanced understanding of the ‘gap’ between AI technology and clinical practice (Sharma et al., 2022; Aristidou et al., 2022). We do this by assessing the gap in terms of frictions, going beyond identifying barriers and analysing the dynamic forces at play. Lastly, our research suggests a promising direction for further theory development. As previous studies have shown, innovation in digital infrastructures takes place at different ‘speeds’ simultaneously (Bygstad & Øvrelid, 2021). Our findings regarding the time-dependent nature of frictions could be developed further to account for differences in innovation speeds in relation to specific AI characteristics and their frictions.
7 Conclusion and future research
Our study has examined how frictions emerge between novel AI technologies and the existing sociotechnical installed base when implementing AI in Norwegian hospitals. By analysing interviews with key stakeholders involved in AI implementation projects, we have identified seven major friction points: data complexities, hospital information infrastructures, regulatory landscape, resource requirements, accountability concerns, implementation strategies, and incentive structures. These frictions highlight the challenges that practitioners face when introducing AI into established healthcare settings and emphasise the need for adaptive approaches that can accommodate both innovative technologies and existing practices. We show that successful AI implementation requires more than overcoming technical challenges or changing organisational practices. It requires understanding the relational, time-dependent, and dual nature of the frictions that emerge from the interfacing between novel technology and the installed base. It requires developing better interfaces that allow both the technology and the installed base to move together in a coordinated way, transforming both in the process.
The future research plan is to expand the current data with interviews on the experiences of AI implementation projects in hospitals to provide a more in-depth understanding of the frictions as experienced from the perspective of practitioners. In addition, we plan to conduct a second phase of the research with in-depth case studies of two specific projects implementing AI in hospitals in 2025: one involving cerebral haemorrhage detection algorithms, and one involving acute myocardial infarction detection algorithms. As we have identified the frictions, we are interested in researching how these frictions are addressed and dealt with in real-life projects in our future research. Since our findings have shown the experimental and innovative nature of AI implementation, case studies are a suitable research methodology to further study how approaches and strategies are enacted in context.
References
Footnotes
The Norwegian Agency for Shared Services in Education and Research (SIKT) is the main governmental body responsible for research ethics in Norway.↩︎
The Medical Device Regulation (MDR) is an EU regulation that governs the production and distribution of medical devices in the European market. AI applications designed for human healthcare, such as for diagnosis, treatment, or monitoring, typically fall under the Class II category of the MDR, requiring more rigorous compliance processes (European Union, n.d.).↩︎
Products with a CE mark indicate that they conform to EU standards regarding safety, health, and environmental protection. For medical AI applications, obtaining a CE mark often requires ‘notified body approval’, unlike other products where a manufacturer’s self-declaration might be enough (Malvehy et al., 2022, p. 361). Not all AI tools subject to MDR necessarily pursue CE marking, particularly those used internally within healthcare institutions without commercial distribution.↩︎
TTOs are typically owned by Regional Health Authorities, universities, university hospitals, and research institutes.↩︎