Inside Organisations: KIN Research Cases

Challenge: With advertising executives viewing GenAI investment as a top priority, the share of AI-generated content in advertising is accelerating — about 86% of advertisers already deploying or planning to deploy GenAI for ad creation. Against this backdrop, RTL Nederlands — a major media company with approximately 600 employees asked itself: how does GenAI change the creative work process, and what does it mean for organizing work with clients?

Our investigation: Researchers focused on the advertising functions of the company, that include campaigns for streaming platforms, broadcast television, and promotional materials. The research team conducted a 13-month ethnographic study following a team of 24 creatives responsible for establishing and realizing a creative vision for campaigns. Researchers observed the full workflow — briefing, concepting, concept approval, production, and post-production — tracking how Midjourney used by creatives changed the process at each stage.

One discovery: The researchers describe GenAI as a "spirited technology": powerful and responsive, but also unpredictable and in need of careful direction. For creatives working closely with prompts, it dramatically expands what can be generated and makes it easier to pitch and sell ideas to clients. At the same time, it introduces new challenges. Early AI-generated visuals can quickly become fixed reference points, limiting further exploration. Both creatives and clients may feel they already "see the idea", leading to premature decisions before concepts are fully developed. As the process continues, the focus shifts. Instead of refining ideas, teams spend more time trying to align highly polished, AI-generated images — often with a "Hollywood-level" finish — with real-world production constraints. The result is what the researchers call a "process collapse": activities that were once sequential — brainstorming, visualization, and production — become compressed into a single, fast-moving iterative loop.

Key takeaway: Introducing GenAI into the workplace delivers efficiency — but also unintended organizational side effects (Retkowsky et. al 2024). Wherever work is structured as a sequence of roles and handoffs, process collapse is a managerial challenge that is easy to underestimate until it is already reshaping how teams operate.

Helping managers recognize how workflows, roles, and organizational structures are being reconfigured in the age of AI is a central focus of the Managing AI programme.

Challenge: For the Leiden University Medical Center (LUMC), a regional care provider and national center for highly specialized medicine, integrating AI is foundational to the future of healthcare delivery. However, developing effective AI tools requires close involvement from clinicians, whose expertise is essential for training, testing, and refining algorithms. Yet the realities of clinical work — high workloads, time pressure, and zero tolerance for errors — leave little room to experiment with tools that are not yet fully reliable. How can organizations involve doctors and technicians in shaping AI without disrupting patient care?

Our investigation: Researchers had been collaborating with the LUMC Radiology department for nearly 6 years, studying their work, organizational structures, and how they organize around new Researchers had been collaborating with the LUMC Radiology department for nearly 6 years, studying their work, organizational structures, and how they organize around new technologies. They conducted multiple ethnographic studies (Kim et al. 2025) and observed how a diverse group of medical professionals participated in in-house development, co-creation, and adoption of commercially developed AI tools. (Grootjans et al. 2025). They have turned some of these insights into a teaching case that guides executives on how to lead such technological transformations (Rezazade Mehrizi et al. 2023).

One discovery: The researchers found that doctors and technicians are highly motivated to engage with AI — but only under conditions that allow them to experiment and learn without putting patient care at risk. At the same time, AI developers are eager to learn from real clinical use and refine their tools based on feedback. To bridge this gap, the research team and Radiology department created a "learning lab": a controlled environment where doctors and technicians could test AI tools in semi-real settings. Within this space, teams experimented with different AI models, but also with key design choices — such as sensitivity thresholds, explainability features, and interface design — as well as different workflow configurations. These experiments revealed how clinicians respond to incorrect AI suggestions, how their attention can become overly focused on certain outputs, and how interaction patterns shape outcomes. Over time, the lab became a shared learning environment where clinicians, developers, and managers could jointly explore what works — and what doesn’t — in practice. It generated insights that improved both the design of AI tools and how they are integrated into clinical workflows.

Key takeaway: When the challenges of adopting new technologies are turned into learning opportunities, different stakeholders can reflect on and critically evaluate their assumptions, practices, and mindsets. For experts in particular, this works best when grounded in real-life feedback on how they can perform their work better.

That same principle — learning by doing, with honest feedback built in — sits at the heart of the Managing AI programme, where participants develop their own AI case and receive tailored feedback and guidance on their AI roadmap project. 

Challenge: When a European plant breeding company, a global market leader in the vegetable seed sector, decided to introduce machine learning (ML) into its cucumber breeding process, the question was how ML could be meaningfully embedded in the complex, experience-driven world of plant breeding.

Our investigation: The company piloted ML solutions designed to support breeders in developing and evaluating varieties with desirable traits — shape, colour, taste, disease resistance, and heat resistance. Researchers followed this process through a 25-month ethnographic study, observing how data scientists and cucumber breeders collaborated in developing and implementing these systems. The research traced how ML outputs were interpreted and acted upon across occupational groups — from the R&D department to seed sorters on the production floor.

One discovery: What the research revealed was that to develop an actionable and useful ML solution, it was not enough for developers to consult breeders or ask them about what tools will support their work. Data scientists tried various ways of collaborating, including workshops, joint labelling and work groups — most of those still presented them with complexities they could not resolve. It was only when the data scientists went out of their offices and into the "greenhouse" itself, "getting their hands dirty" — that they ultimately were able to make a breakthrough in their tools (Mayer et al., 2023). Shadowing, working closely and regularly alongside breeders, helped them understand in a holistic way how expertise was actually exercised: through experience, sensory input, and context.

Key takeaway: Developing actionable and valuable ML tools is possible when data scientists and domain experts collaborate closely. This means not only discussing options or consulting on use cases, but shadowing the actual work on the ground. Counterintuitively, not just consultation, but immersion may result in saving valuable time and resources for developing tools that are not likely to be used or provide little value for the users on the ground.

How to design development pipelines and co-creation processes that bridge that gap from the start is one of the core questions the Managing AI programme tackles.

Challenge: Faced with shrinking capacity and growing pressure to act smarter, the National Police turned to AI-based predictive policing. The central question was not whether algorithms could predict crime — but how such systems could be responsibly embedded in everyday police work.

Our investigation: The police piloted an in-house developed AI system, the Criminal Anticipation System (CAS), designed to predict where and when certain types of crime were most likely to occur. Researchers followed this process through a three-year study combining piloting CAS across four cities with 36 months of ethnographic fieldwork—shadowing police officers at the station, in the cars and on the streets.

One discovery: The researchers found that doctors and technicians are highly motivated to engage with The research traced how AI outputs were interpreted, translated, and acted upon by different professionals — from data scientists and intelligence officers to managers and officers on the street. AI does not work on its own. The real impact emerged through new roles, workflows, and decisions about authority and responsibility. Intelligence officers became vital "algorithmic brokers", translating abstract predictions into actionable guidance (Waardenburg et. al 2021). Managers had to decide when to trust AI over experience, while officers adjusted to reduced discretion in patrol decisions. The success of CAS depended less on the algorithm itself and more on how well it was integrated into organizational routines and professional judgment.

Key takeaway: Managing AI is a management and organizational challenge first — and a technical one second. The risks are real, the governance choices matter, and the unintended consequences are often where the most important lessons hide.

Understanding how to identify those risks early — and build governance frameworks that hold up under the pressures of real organizational life — is precisely what the Managing AI programme is designed for.