AI lessons from the boardroom
Daan Debie on how NS brings AI from pilot to practice
By Rutger Huizenga
AI is on virtually every agenda. Not as a gadget, but as a strategic question: what do you use it for, how do you scale it up, and how do you prevent experiments from getting stuck in prototypes? In the interview series AI lessons from the boardroom, we speak with leaders and executives who do not merely discuss AI, but have to apply it every day. Today: Daan Debie, Director of Data, Innovation and AI at NS, and new board member of KickstartAI on behalf of NS.
There comes a moment when AI stops being a topic of discussion and becomes part of how an organization works. For Daan Debie, Director of Data, Innovation, and AI at NS and recently appointed board member of KickstartAI, that point has already passed. He doesn’t talk about AI as something that is still on its way. He talks about it as that is already part of daily operations: useful when it works, instructive when it doesn’t.
At NS, many AI applications are already part of the passenger experience. The NS app, for example, predicts how busy a train is expected to be. The OV-fiets app estimates whether bikes will be available at a given pick-up point and time. Both are predictive models running in production, not experiments waiting to prove their value.
"People quickly think of language models, of generating or summarising text," says Debie. "But predictive models, image recognition, logistical optimisation, that is all AI too. It is already running. It’s just that nobody notices it."
That invisibility isn’t a problem, but rather a sign of maturity. And Debie knows better than most what it feels like to build that layer of invisibility. He started his career as a hands-on data engineer, built data systems at KLM, and brought machine learning models into production. He later moved into the startup world, joining Source.ag as CTO, a company applying AI in greenhouse horticulture. "That gives you a unique perspective. At a large corporate, you face different challenges in rolling out AI than at a startup where the motto is: move fast, break things. I learned to work at two different speeds."
That experience shapes how he now leads a team of three hundred people at NS, building data and AI solutions for virtually every part of the organisation.
Making adoption deliberate
The public debate about AI still revolves largely around tools, but Debie prefers to think in terms of organisations. Who decides? Who is accountable? And how do we ensure that new technology doesn’t reach only the people who were already ahead?
"That is the real risk," he says. "Not whether AI works but whether it becomes unevenly distributed. That two years from now, you may have one group of people who use it fully, and another group who have barely encountered it. That kind of gap is hard to close."
This is why the NS board made a clear decision around Microsoft Copilot. "We explicitly said in the boardroom: our office staff must learn to work with this. This is not: you are welcome to try it. Nor is it: there is a pilot for those who are interested. No, this is how we are going to work, and we are going to actively train people to do so,” says Debie. At the same time, we are looking across all parts of the business for concrete areas where it can add immediate value."
NS has also put its entire leadership group through what Debie calls a kind of car wash. They are working through a series of training courses on data and AI over a period of eighteen months. And the training goes well beyond learning to use Copilot. "It’s about understanding what kinds of AI exist. About basic statistics. How do I read a graph? How do I spot errors in data? What questions should I be asking? Executives don’t need to become data scientists. But they do need to understand what they are steering."
Choosing where to focus
Once AI adoption becomes a serious priority, ideas start coming from across the organisation. For Debie, that is a good sign, but it also makes prioritisation essential.
"That energy is wonderful," he says. "But you need to know what is already in motion. Capacity and money are finite, so you have to make choices."
Some parts of the business have more appetite for AI than others. But that appetite alone cannot determine where NS invests its money and capacity. That is why NS established an AI Centre of Excellence earlier this year: a central team that helps business units surface use cases, build proof of concepts, and standardise ways of working. The starting point is always the passenger. "If we can use AI to improve customer satisfaction and get more people out of their cars and onto the train, that application goes to the top of the list."
From demo to platform
The Centre of Excellence helps NS choose where to focus and get the right projects started. For Debie the harder question is what happens after a project has been approved. Going from idea to prototype is now relatively straightforward at NS. But going from prototype to production is a different matter.
"That central team cannot maintain everything indefinitely," he says. "At some point, a business unit has to take ownership, free up budget, feel responsible, be accountable. Someone has to be willing to wake up on a Saturday evening when it stops working."
That is easier said than done. Ownership requires more than a signature on a budget. It requires people to let go of familiar ways of working and place their trust in something new. Sometimes, even something they had been doing themselves for years.
A good example is the camera gantries that NS has recently installed along the tracks. These photograph trains passing at high speed, and an image recognition model checks whether there are defects in the wheel sets or pantographs. "Before this technology existed, engineers would literally walk along all the wheel sets at night with a torch," says Debie.
Moving from that kind of manual inspection to algorithmic detection changes more than the process. "Those engineers suddenly have to trust an algorithm over their own visual inspection. That is quite a leap of faith. And if you do not manage that well, your AI product stalls at adoption, and you have quite literally thrown money away."
In order to make that leap of faith smaller, Debie suggests involving users from the early stages of building. “Instead of saying “here is the system, good luck,” consider something like this: This is the problem, this is how we want to solve it, do you want to help build the solution? Only then do you create something that people actually need.”
A circle completed
His love for co-creation also explains Debie's enthusiasm for KickstartAI, the collaborative initiative of NS, KLM, ING, and Ahold Delhaize, a foundation through which large Dutch organisations work together on concrete AI applications, share knowledge, and challenge one another's thinking.
The collaboration has already led to tangible results at NS. One example is the system that automatically detects unattended luggage at stations, built together with KickstartAI. Until now, staff had to monitor screens continuously for abandoned items. Soon, an AI model will help take over that task. "An AI expert from KLM looks at an NS problem with completely different eyes than one of our own data scientists," says Debie. "At a large corporate, you can become a prisoner of your own culture. That fresh perspective is invaluable."
For Debie, KickstartAI is also personal. When he was still at KLM, he helped lay the groundwork for the foundation together with co-founder Sander Stomph. Returning now as a board member feels like a circle completed. But it is about more than sentiment.
The stakes go beyond business. Ultimately, whoever doesn’t develop AI themselves hands the rule-making to others. "If you let it happen from the outside, you lose control. We care deeply about privacy in Europe, far more than they do in the United States. If we want to protect those values, we have to be the one doing the building. And no company can do that alone, so let’s do that together with the help of KickstartAI?”
The 5 AI lessons of Daan Debie
Do not confuse AI with GenAI. AI is far broader than ChatGPT. Look beyond language models towards predictive models, image recognition, optimisation, and operational applications.
Make AI a governance matter, not just a technical one. If you let it emerge bottom-up, adoption will be slow. AI needs a mandate from the boardroom.
Train leadership as well as the organisation Executives don’t need to become data scientists, but they must understand what they are steering. That starts with basic literacy: how do you read data, how do you spot erors, and what questions should you ask?
Bring AI all the way to production. You can build a beautiful model, but at some point it has to run somewhere and be maintained. Real value only emerges after the demo.
5 Involve users from day one AI adoption is much harder when users only see the finished system. Bring them in early, so the solution fits the reality of their work.
