Why Top-Down AI strategies fail – and how em­ploy­ees de­liv­er res­ults

Once a month, researchers at Copenhagen Business School provide Børsen readers with a current and research-based perspective on the challenges facing leaders.

This time, Associate Professor Christian Hendriksen from the Department of Operations Management highlights the steps leaders can take to successfully unlock the value of AI.

Across the world, CEOs are trying to make sense of how to use generative AI and large language models. Some may have read reports on the potential, or hired consultants to identify possible use cases. 

But when leadership presses the big button – buying Microsoft 365 Copilot for everyone and expecting clear results – they are often disappointed. The explanation is simple: it is the employees, not top management, who know how the technology delivers practical value. That is why top-down AI projects often miss the mark. 

As a researcher and educator, I frequently speak with employees in companies, municipalities, and government agencies who tell me they are using ChatGPT on a personal account for work-related tasks. 

They glance around, lower their voices, and tell me the same thing: they are already using it – because it works. And they are doing so to complete real tasks faster and better, not to impress anyone with a demo. 

Yet many hesitate to tell their managers – either out of fear of being told off or because there is no space for open knowledge sharing. At the same time, managers are waiting for employees to "come up with use cases." But without openness, clear frameworks, and the freedom to explore the best models, everything comes to a halt. 

Users are the experts

Language models are general-purpose technologies: they can be used for many things, but the benefits only emerge when professional expertise gets hands-on with the tools. 

A manager prompting for half an hour doesn’t have the same insights as the sales rep using the best model for several hours a day. Without real experimentation – and recognition that the users are the experts – large AI initiatives will fall short. 

Close the gap between AI and everyday work

This is a challenge both for those eager to move fast and those choosing to wait and see. The solution is to bring learning closer to the actual work. That means shifting AI out of the project plan and into everyday processes, where small improvements can be measured and embedded. 

Three practical recommendations 

1. Give employees space and tools to experiment 
Provide your employees with trust, resources, and access to the best models – in their real work contexts, not in a closed test sandbox or last year’s free version of Copilot. If Jytte from marketing is to understand the tech, she needs access to the best tools. It is affordable: top models cost around $25 a month. 

Also, set clear guidelines for data use: what can be shared in which tools, and what is off-limits? For example, create a whitelist of documents that are always safe to share with a chatbot, so employees know they are on solid ground. If real data is off-limits entirely, no one will learn anything – and Jytte won’t get anywhere. 

2. Create an open learning space 
Hold short demo sessions where employees show how they’ve used AI to complete tasks better, faster, or more safely. Turn the strongest examples into shared templates and small process tweaks so that experience becomes operational – and let leadership listen. 

3. Use the tools yourself 
Your calendar is full, but when leaders use the same models, the quality of questions, priorities, and expectations improves. You will better understand employees’ concerns and ideas, and you will be in a stronger position to assess which ideas are ready to scale – and where data or systems are falling short. 

This last point is key: if you are thinking “I don’t have time to experiment,” your employees are probably thinking the same. 

Ask yourself: do we have time and space to try new things and share knowledge? Is there room for someone to take the lead and test whether the newest ChatGPT model can handle a complex task? If not, then your first task is to create that time and space. 

Build AI with patience

General-purpose technologies take time to take hold. It took nearly 35 years from Edison’s first power line in New York before electric motors replaced steam in half of all US factories. Something similar – though faster – is now happening with generative AI. 

It requires patience, but also speed in learning. The same principle applies: major productivity gains only come when work is organised around the technology, and processes, systems, and routines are adjusted accordingly. That won’t happen with a six-month "Copilot project." 

Start with something tangible: give a focused team access to the best models, clear data rules, and a regular rhythm of knowledge sharing. Let employees experiment with the technology based on their professional expertise and understanding of their own tasks. And let their insights form the foundation of your company’s AI strategy.