From Readiness to Regulation: Practice-Oriented Measures to Increase the Adoption of Generative AI in the Swiss Public Administration

2.1 Insights from Existing Research

For years, researchers have been engaged in AI research (Zuiderwijk et al., 2021; Yigitcanlar et al., 2024; Wirtz & Müller, 2019). Since AI’s impact on daily life has grown increasingly significant, it has become a popular field of study (Wirtz, Weyerer & Geyer, 2019). The focus of AI research has traditionally been within the field of computer science, emphasizing the technology itself and its functionality (Simon, 1995; Xu et al., 2021). In contrast, AI research with a focus on social sciences, particularly within the context of the public sector, has been largely neglected (Aoki, 2020; Sun & Medaglia, 2019; Wirtz, Weyerer & Geyer, 2019; De Sousa et al., 2019; Janssen & Kuk, 2016). Historically, research emphasis has been placed on the private sector (Desouza, Dawson & Chenok, 2020; Neumann, Guirguis & Steiner, 2024). Zuiderwijk et al. (2021) attribute this to the significantly more advanced adoption of AI in the private sector compared to the public sector.

Despite the limited attention, Yigitcanlar et al. (2024) demonstrated in their scientometric analysis that AI adoption in the public sector spans five decades, dating back to the late 1980s. They identified four key phases of AI literature in the public sector over the last 50 years: The foundational era (late 1980s), the conceptual era (mid-1990s), the exploratory era (mid-2000s), and the booming era (mid-2010s to late 2023). In recent years, during this “booming era”, AI adoption research in the public sector has gained significant momentum (Zuiderwijk et al., 2021; Wirtz, Weyerer & Geyer, 2019; Mikhaylov, Esteve & Campion, 2018). For example, to understand the sector-specific challenges and drivers of AI adoption in the public sector, Neumann, Guirguis and Steiner (2024) conducted an empirical analysis of the AI adoption process in eight Swiss public organizations.

This new surge of AI adoption research in the public sector is partly driven by rapid technological advancements (Aoki, 2020) and the corresponding increase in attention from policymakers (Mergel et al., 2024), closing the gap with the private sector (Desouza, Dawson & Chenok, 2020). This trend has been further accelerated by the COVID-19 pandemic. As Mergel et al. (2024) note, the pandemic amplified calls for more citizen-friendly service delivery, which led to significant government investments in AI technologies. Together, these developments have set the stage for a growing focus on Gen AI within public sector organizations.

Building on this interest, current literature primarily addresses the economic and societal impacts of Gen AI in the public sector (Bright et al., 2024). Economically, McKinsey & Company (2023) estimates that Gen AI could contribute a productivity benefit of approximately $480 billion within the public sector and related fields. Boston Consulting Group similarly projects substantial financial gains, estimating global annual productivity improvements from Gen AI in the public sector could be worth up to $1.75 trillion by 2033 (BCG, 2023). Beyond purely economic aspects, Mellouli, Janssen and Ojo (2024) emphasize that Gen AI offers significant benefits to the public sector. These include gains in productivity (e.g. Margetts, Dorobantu & Bright 2024; Beltran, Mondragon & Hun, 2024; Wang et al., 2025), improved communication (Chun & Noveck, 2025), better internal services and automation of tasks (e.g. Yigitcanlar et al., 2024; Wirtz, Langer & Fenner, 2021), enhanced citizen engagement (e.g. Mellouli, Janssen & Ojo, 2024), service co-design and delivery (e.g. Wang et al., 2025), personalized public services (e.g. Chun & Noveck, 2025), and improved decision-making and forecasting (e.g. Maragno et al., 2023). This potential translates into concrete productivity gains by reducing bureaucratic workloads, allowing civil servants to focus on higher-value tasks (Margetts, Dorobantu & Bright, 2024), and shifting up to one-third of their time from routine to more impactful work (Berryhill et al., 2019), ultimately contributing to increased public value (Wang, Teo & Janssen, 2021; Criado & Gil-Garcia, 2019).

Recently, research has also begun to explore the risks and unintended adverse effects of Gen AI in the public sector. Several studies have emphasized the negative consequences of AI adoption, addressing legal, regulatory, ethical, political, technical, social and organizational challenges (Agarwal, 2018; Bannister & Connolly, 2020; Janssen & Kuk, 2016; Wirtz, Weyerer & Geyer, 2019; Wirtz, Langer & Fenner, 2021; Selten & Klievink, 2024; Sun & Medaglia, 2019; Hopster & Maas, 2024; Bright et al., 2024; Margetts, Dorobantu & Bright, 2024; Taeihagh, 2025). For instance, Cantens (2025) emphasizes the inevitability of Gen AI adoption while also underscoring its associated challenges, including limited explainability of outputs, confidentiality constraints, and persistent mistrust among civil servants toward this emerging technology. Neumann, Guirguis and Steiner (2024) even refer to “justified scepticism and fear of governments using AI” as it may reinforce existing inequalities (Margetts, Dorobantu & Bright, 2024), compromise citizens’ privacy (Hopster & Maas, 2024), and ultimately threaten democracy (Allen & Weyl, 2024). Collectively, these studies underline that Gen AI adoption is best understood as a “Double-Edged Sword” (Salah, Abdelfattah & Al Halbusi, 2024; Neumann, Guirguis & Steiner, 2024), promising transformative opportunities but simultaneously carrying profound social, political, and organizational risks for the public sector (Kuziemski & Misuraca, 2020).

2.2 Gaps in the existing Literature

Despite the progress made in recent years, research on Gen AI in the public sector remains scarce (Sun & Medaglia, 2019; Wirtz, Langer & Fenner, 2021; Yigitcanlar et al., 2024). Mergel et al. (2024) observe, “Even though we were able to bring together a range of new papers on AI implementation in the public sector, there is still an immense research gap to be filled (…).” Similarly, De Sousa et al. (2019) note that, despite the recent increase in AI initiatives across government sectors worldwide, there remains a lack of systematic understanding regarding the motivations, processes, outcomes, and impacts of AI implementation in the public sector. Taeihagh (2025) attributes this to the literature’s inability to keep pace with the rapid technological advancements, leaving the study of AI management in public administration still underdeveloped relative to the field’s evolution.

Similar to broader AI research in the public sector, scholarly investigation into the adoption of Gen AI, specifically in public administration, remains in its infancy. So far, much of the current knowledge on Gen AI adoption in public administration stems from theoretical discussions, often focused on in-depth case studies of individual AI tools (Neumann, Guirguis & Steiner, 2024; Selten & Klievink, 2024). Beltran, Mondragon and Han (2024) highlight a “notable gap in the academic literature concerning a systematic understanding of governments’ roles and functions in this AI-infused era.” Bright et al. (2024) state that only a few studies explore the implications of Gen AI for public sector workers, emphasizing that “Answers to these questions will be vital in helping us to safely and robustly actualize the potential of Gen AI technologies in the public sector.”

This “widening empirical evidence gap” (Mergel et al., 2024), particularly on everyday use by public sector employees, creates a disconnect between academic discourse and practical application (Newman, Cherney & Head, 2016). In response, recent studies have begun to examine specific applications of Gen AI across different sectors (e.g. Zhang & Boulos, 2023; Qadir, 2023; Madan & Ashok, 2023) and national contexts (e.g. Guenduez & Mettler, 2023; Wang et al., 2025; Alshahrani et al., 2022), providing early insights into how the technology is being integrated and the factors shaping its adoption.

While these studies highlight important sectoral and national variations, little attention has been paid to the adoption of Gen AI within the distinctive political, administrative, and regulatory framework of Switzerland.

3.1 The Swiss Context

Switzerland offers a well-suited empirical context for analysis due to its highly decentralized federal system, consensus-based governance, linguistic and cultural diversity, and strong subsidiarity principles (Fleiner, 2009). In comparative analyses, Switzerland consistently ranks at the top for both fiscal and policy decentralization (Füglister & Wasserfallen, 2014). Given this strong federal structure, all levels of government are endowed with significant autonomy. The Swiss constitution enshrines the principle of subsidiarity, stipulating that all responsibilities not explicitly allocated to the federal level fall under the jurisdiction of the cantons (Linder & Vatter, 2001; Füglister & Wasserfallen, 2014).

This high level of decentralization also presents challenges for digital transformation in the public sector. According to Schedler and Summermatter (2005), it partly explains Switzerland’s consistently moderate rankings in e-government performance. With approximately 3,000 political authorities deciding independently on their e-government initiatives, there is limited national coordination. Furthermore, Switzerland has an advanced regulatory framework, particularly in data protection, which adds further complexity. These challenges are reflected in international benchmarks. For example, the 2024 eGovernment Benchmark Report by the European Commission, shows that Switzerland lags significantly behind the European average in the availability and maturity of digital public services, ranking 29th out of 35 (Capgemini et al., 2024).

To address this fragmentation, the organization Digital Administration Switzerland (DVS) was established to coordinate and promote digitalization efforts across all levels of government. Within the strategy Digital Administration Switzerland 2024–2027, DVS develops annual implementation plans to advance its objectives. The most recent plan for 2026 also includes the Coordination Body for Data Science & AI as an active working group composed of experts from public administration that can be engaged as needed by the executive steering committee. The establishment of this coordination body signals a growing institutionalization of AI-related issues within the Swiss public administration. At the same time, it reflects a broader political acknowledgement of AI’s strategic relevance.

In its policy agenda for the 2023–2027 legislative period, the Swiss Federal Council highlighted both the challenges and risks that AI poses for Switzerland as a business hub, as well as the need to explore appropriate regulatory responses. In this context, the Federal Council mandated the Federal Chancellery to develop a dedicated strategy for the use of AI systems within the federal administration. Published in 2025, this initial strategy is structured around three guiding pillars: strengthening institutional competencies, ensuring trustworthy application, and enhancing efficiency through the use of AI systems (Bundesrat, 2025a). Together, these pillars establish the foundational framework for AI adoption in the Swiss administration.

In light of this growing institutional awareness, discussions have begun to emerge around the development of a dedicated national Gen AI system tailored to the specific needs and values of the Swiss public administration.

3.2 Development of a Swiss Governmental GPT

In March 2025, the Swiss newspaper Tagesanzeiger revealed that the Swiss federal government has been developing and operating a proprietary, isolated AI system called Gov-GPT since November 2024. This system runs in a specially secured IT environment within the federal administration. Developed by the Federal Office of Information Technology, Systems and Telecommunication (FOITT), Gov-GPT is based on the open-source language model LLAMA by Meta AI (Knellwolf, 2025).

Gov-GPT has similar functionalities to ChatGPT and other large language models, but its protected environment allows it to process data that would not be permissible under current data protection and information security regulations (Anz, 2025). According to the official guidelines on the use of Gen AI tools in the Swiss federal administration dated January 18, 2024, inputting many types of information into Gen AI tools is prohibited. Specifically, information classified as “internal,” “confidential,” or “secret” may not be processed. Even non-classified documents containing sensitive content, including any form of personal data, are excluded from use in public Gen AI systems (Bundesrat, 2024). This stands in sharp contrast to Gov-GPT, where only “secret” information remains prohibited which implies that Gov-GPT can process data that was previously off-limits to conventional AI systems (Landis, 2025).

The emergence of Gov-GPT coincides with a growing political debate on digital sovereignty in Switzerland, which the Federal Council defines as the state’s ability to retain control and autonomy in the digital domain (Bundesrat, 2025b). This debate has intensified following the recent introduction of Microsoft 365 within the Swiss federal administration. While aimed at increasing efficiency, this initiative has also raised concerns related to data control. In response, the Federal Chancellery launched a feasibility study in 2024 on the use of open-source software (OSS) for office automation, assessing its suitability as a fallback solution. This line of inquiry is further reflected in the Swiss Federal Council’s 2025 report “Digital Sovereignty of Switzerland” (Bundesrat, 2025c), which emphasizes the promotion of OSS as a key governmental measure to reduce technological dependencies and identifies AI as a particularly sensitive domain to ensuring Switzerland’s digital sovereignty.

Against this backdrop, Gov-GPT can be interpreted as an early institutional response to the challenge of developing more sovereign Gen AI tools. It thus serves as a practical example of how Switzerland seeks to advance the broader implementation of Gen AI within its federal administration while embedding the abstract concept of digital sovereignty into concrete administrative practice.

4.1 Nationwide Survey

Main Findings

The survey indicates a limited level of Gen AI adoption in the Swiss public administration: Over one-third of respondents (35.9%) have never used Gen AI at work, while almost the same proportion (36%) use it only rarely. Overall, more than 70% of public sector employees use Gen AI either never or only occasionally (see Figure 2). In contrast, 12.4% of respondents reported using the technology approximately once a week, 13.2% several times per week, and only 2.4% reported daily use in their work.

Figure 2

The survey shows that, while the overall adoption of Gen AI is comparably low across municipalities, cantons, and the federal level, some differences emerge between administrative tiers. At the municipal level, 42.7% of respondents reported never having used Gen AI at work, whereas the proportion was lower at the cantonal (24.4%) and federal levels (21.1%), indicating that a higher share of employees at these levels had at least experimented with the technology. Cantonal employees also appeared to use the technology more frequently, with 19.7% reporting multiple times per week and 3.8% daily usage, compared to 9.7% and 1.8% at the municipal level. For a detailed breakdown of response distributions, including the number of responses per category, see Appendix Table A1.

After controlling for gender, professional experience, and education level, only minor variations in Gen AI usage were found. Men reported slightly higher usage than women, with 33.5% of men and 41% of women never having used the technology. Usage was most common among employees with 1–5 years (around 17%) and 5–15 years (around 13%) of experience, while beginners (<1 year, 43%) and very experienced employees (>15 years, 41%) most often reported no use. Higher education levels were generally linked to slightly more frequent usage.

In summary, the adoption of Gen AI among Swiss public sector employees remains very low, indicating a need to strengthen its adoption across public administration. Without broad uptake, the potential benefits of Gen AI are unlikely to be fully realized. This highlights the importance of targeted, evidence-based measures that reflect the specific needs and capacities of Swiss public sector employees. Such measures are essential to support the effective use of Gen AI and have direct implications for the further development of Gov-GPT as a concrete institutional initiative.

4.2 Delphi Study

4.3 Consolidated Framework for Implementation Research

Main Findings

Figure 3 provides a synthesized representation of the Delphi-informed CFIR results and aggregates these findings into a layered visualization that highlights where adoption-relevant issues cluster in shaping Gen AI use within the Swiss public administration. The figure not only illustrates the structural relationships between domains but also makes visible the interconnections among the Delphi results and their empirical prioritization, thereby revealing the key areas that are most critical for the effective implementation of Gen AI in Switzerland.

Figure 3

As illustrated in Figure 3, the measures identified through the Delphi process are primarily assigned to the “Individual setting” and “Inner setting” domains of the CFIR Framework. These two domains stand out not only quantitatively, due to the number of associated issues, but also qualitatively, as they encompass the two issues rated as most important by the Delphi study participants (1 Ensure Access; 2 Training).

The issues clustered around the “Individual Setting” underscore the central role of personal attitudes and competencies in influencing adoption. Constructs such as Training (MR 2.24) and Reduce Stigmatization (MR 6.45) reflect key individual-level prerequisites for engagement, suggesting that successful implementation hinges on lowering both cognitive and psychological barriers among users.

Simultaneously, the measures mapped to the “Inner Setting”, such as Ensure Access (MR 2.18), Communication (MR 4.83) and Coercion (MR 6.97), underscore the importance of a supportive organizational environment that provides access to Gen AI, emphasizing the role of internal communication and the availability of appropriate tools within the Swiss public administration. These findings resonate with established implementation science literature, which emphasizes the mediating function of the inner setting between broader contextual pressures and individual adoption behaviors (Liang et al., 2007; Coelho, August & Lages, 2011).

Each of the remaining three CFIR dimensions is represented by only one measure, each of which received a medium priority rating from the Delphi experts. For the dimension “Intervention Characteristics”, the relevant factor was Demonstrate Benefits (MR 2.39). The “Outer Setting” was most strongly associated with Clear Guidelines (MR 3.18). Finally, the inclusion of Set an Example (MR 4.06) under the CFIR Dimension “Implementation Process” highlights the strategic importance of role modeling and leadership engagement in fostering legitimacy and trust in AI tools.

Taken together, embedding the Delphi results within the CFIR framework highlights a clear concentration of adoption-relevant measures within the “Individual” and “Inner Setting” domains. This suggests that Gen AI adoption in the Swiss public sector is primarily driven by human and organizational conditions rather than by characteristics of the technology itself or external pressures. By comparison, the “Intervention Characteristics,” “Outer Setting,” and “Implementation Process” domains play a more supportive role, indicating that while these elements contribute to implementation of Gen AI in the Swiss public administration, they are less decisive than individual and organizational determinants.

5.1 Contributions to Literature

This study makes three key contributions to literature. The first is its empirical contribution, which advances the understanding of Gen AI adoption in the public sector by providing evidence from the Swiss context. Whereas prior studies have been mostly conceptual or normative (Wirtz, Langer & Fenner, 2021; Mergel et al., 2024), this study grounds theory-building in a highly decentralized and compliance-oriented setting and demonstrates the stage of adoption in the Swiss public administration, revealing how public sector employees are beginning to engage with and apply Gen AI in their work.

The second is its methodological contribution, which demonstrates how a sequential mixed-method design can strengthen theory development, by reducing bias (Dzwigol, 2022) and providing a more comprehensive understanding of the phenomenon (Nanthagopan, 2021). By combining a nationwide survey of public servants with a Delphi study of experts and embedding the findings within the CFIR framework, the study illustrates how methodological triangulation can reveal individual and organizational conditions of adoption that might remain hidden in single-method designs.

Damschroder et al., as the developers of the framework, explicitly conceptualized CFIR as a “living, evolving, and improving set of propositions” and already emphasized in their original publication (2009; updated 2022) the need for deeper contextual knowledge. Building on this evolutionary perspective, the third contribution of this study offers technological refinements by extending the CFIR framework to the context of Gen AI adoption in the public sector. It identifies individual attitudes and organizational readiness as particularly critical implementation domains, especially during early-stage adoption within highly decentralized and compliance-oriented administrations.

5.2 Implication for Practice

First, the study offers context-specific guidance for the Swiss public sector by highlighting differences in Gen AI adoption across municipalities, cantons, and the federal administration. Delphi experts consulted at the Swiss Smart Government Day provided tailored priorities to increase Gen AI adoption addressing structural, cultural, and institutional conditions unique to Switzerland. Aligning these measures with observed adoption patterns allows public sector leaders to prioritize interventions that are both practical and locally relevant.

Second, this study provides practical insights for public sector organizations considering the adoption of Gen AI technologies such as Gov-GPT (see “Explanation for Swiss Gov-GPT” in Table 3). For policymakers and managers, secure access to Gen AI tools and targeted employee training emerged as the most critical interventions. The findings differentiate high-impact from low-impact measures, guiding where to focus resources for effective early-stage adoption and avoiding strategies unlikely to influence outcomes, such as coercive approaches or purely technical refinements.

Finally, the study implies a structured approach for implementing the identified priorities within public sector organizations. It emphasizes the combination of individual-level measures with organizational support, ensuring that initiatives like a Swiss Gov-GPT are implemented effectively. This approach allows for a possible sequencing of interventions, highlighting which implementation domains should be addressed first to maximize Gen AI adoption and which can follow once foundational conditions are in place. These findings provide an evidence-based foundation for policy action, suggesting that the Federal Chancellery AI strategy and its implementation plan should incorporate these insights, directly linking individual and organizational measures to broader national initiatives such as Gov-GPT.

5.3 Limitations and Future Work

Several limitations must be acknowledged when interpreting the findings of this study.

First, the generalizability of the results is limited because the study is based solely on the Swiss public administration case. Institutional arrangements, regulatory frameworks, and administrative cultures differ across countries, so the adoption patterns and effective measures identified here may not fully translate to other national contexts. While the insights are relevant for governments with similar decentralized or compliance-driven environments, caution is required when applying them beyond Switzerland.

Second, the selection and composition of the Delphi panel may have influenced the study’s conclusions. Participants were recruited from the Swiss Smart Government Day, but it cannot be conclusively determined that all attendees were experts in public administration. The real-time format of the Delphi process also limited opportunities for reflection between rounds, and the overlap between the initial pool of experts and the final panel may have constrained the diversity of perspectives, potentially affecting the prioritization of measures.

Third, the study did not include a longitudinal analysis. Evolving institutional conditions, such as the gradual reduction of the “confidentiality barrier” (Cantens, 2025) through the deployment of Gov-GPT, may naturally influence perceived risks and adoption behavior independently of targeted interventions (Ajayi et al., 2025). Even without specific measures, rapid adoption of Gen AI is already observable (Arce-Urriza et al., 2025). As a result, the survey data used in this study, collected in 2024, can no longer be considered fully current.