A three-year study of 23 large companies in Switzerland has identified a recurring obstacle facing organisations that adopt generative artificial intelligence: while businesses have spent heavily on giving staff access to general-purpose AI tools, most have struggled to translate that investment into significant organisational value. The research, conducted between 2022 and 2025 by academics at the University of St. Gallen and HEC Montréal, points to a specific organisational structure, which the researchers term the "AI spine", as a factor distinguishing companies that have managed to scale generative AI use across business functions.
The study drew on data from more than ten workshops and 87 in-depth interviews with front-line employees, business leaders, and technology and data executives across sectors including retail and investment banking, health and general insurance, medical coding, energy, law, manufacturing, postal services and technology consulting. The participating organisations were members of a research consortium focused on generative AI adoption.
Generative AI refers to systems, including large language models, that can produce text, code or other content in response to prompts, rather than simply analysing or classifying existing data. Since 2022, many companies have deployed such tools to help employees draft documents, summarise information or write code more quickly. According to the researchers, this has generally improved individual productivity but has not, in most cases, led to the kind of process-level change that would deliver a measurable return on investment or a competitive advantage.
The researchers identified three practices shared by leaders who succeeded in scaling generative AI beyond individual use. First, they broadened the range of use cases across multiple processes rather than concentrating on a single task. Second, they treated each use case as an ongoing project to be refined over time rather than a finished deployment. Third, they moved quickly to discontinue use cases that failed to demonstrate measurable value, rather than allowing unsuccessful projects to continue indefinitely.
According to the study, most traditional companies are not structured in a way that supports these practices. Many large organisations operate through multiple business divisions, each responsible for its own profit and loss, with duplicated functions and limited flow of information between units. The researchers note that this structure, combined with internal competition for resources between divisions, makes it difficult to spread generative AI applications across an organisation once they have proven useful in one part of the business.
Many companies have attempted to address this through a "hub-and-spoke" model, in which a centralised team of AI specialists provides technical support to individual business units on request. The researchers found that this conventional approach has limits, since it typically depends on units separately identifying and requesting support for their own use cases, rather than enabling ideas and expertise to move more freely across the organisation.
The organisations that the researchers found most effective at scaling generative AI had instead developed what the study calls an "AI spine": a cross-functional internal structure intended to connect ideas, expertise and technical resources across business units on an ongoing basis, rather than treating each unit's needs separately. According to the researchers, this structure allows employees and teams from different parts of an organisation to share ideas about where generative AI might improve a process, drawing on lessons learned elsewhere in the company rather than starting from scratch.
The study also found that companies using this approach applied more disciplined project governance, with clearer criteria for assessing whether a given use case was delivering value and a stated organisational tolerance for abandoning projects that were not. The researchers argue that this combination, broader idea-sharing paired with a mechanism for concentrating resources on the most promising applications, helped participating companies avoid a common pattern in which generative AI initiatives multiply without any becoming embedded in core business processes.
The findings add to a wider body of academic and industry commentary examining why many organisations have found it difficult to move from experimental use of generative AI to systematic organisational adoption. Other researchers cited in the study have separately argued that scaling generative AI in the workplace requires structural change within organisations, rather than simply expanding access to existing tools, and that some conventional organisational models may need to be reconsidered as generative AI and large language models become more embedded in day-to-day operations.
The research does not suggest that adopting an AI spine structure guarantees successful scaling of generative AI, nor does it quantify the financial return achieved by the companies studied. Rather, the study identifies a set of organisational practices and structures that were present in companies the researchers judged to have scaled generative AI use cases more effectively than others in the consortium. The authors note that multidivisional organisations, a structure common among large companies across many industries, face particular difficulties in coordinating generative AI initiatives because of existing barriers to cross-unit collaboration and resource-sharing that predate the technology itself.
The study's authors, Kevin Schmitt of the University of St. Gallen's Institute of Information Systems and Digital Business, Gregory Vial of HEC Montréal, and Ivo Blohm, also of the University of St. Gallen, conducted the research as part of a broader consortium examining generative AI adoption among Swiss companies. Their work adds empirical detail to ongoing discussions among business leaders and researchers about how large organisations can move beyond individual productivity gains from generative AI tools towards changes that affect entire business processes.
As companies across sectors continue to weigh further investment in generative AI, the study suggests that organisational structure, rather than the sophistication of the underlying technology alone, may play a significant role in determining whether such investment translates into measurable business value. The researchers indicate that further work is needed to establish how widely the AI spine model, or similar cross-functional structures, might apply beyond the sample of companies examined in this study.