7 practical lessons from over 150 AI projects

“We already saw it in 2018,” says Benoît Hespel, Head of AI at Proximus ADA.

“Everyone wanted to use neural networks, even though they were not the right choice for certain projects.” Technology for technology’s sake is never the best starting point. But that’s just what is happening today with LLMs and AI agents. “Companies don’t want to fall behind. They are eager to ‘do something with AI’, but often don’t depart from a real problem,” says Dirk Luyckx, CTO at Codit. “Then we have to go back to the essence together: what exactly do you want to solve?”

Ambition is important, of course, as long as you split it into workable steps. “Saying that you want to ‘improve the efficiency of the network’ is too vague,” according to Benoît. “What specifically does that mean? Which KPIs do you want to improve and to what degree?” Without a clear measurement framework, endless iterations take place without decisions. “In addition you must be very strict with yourself. If the expectations are not clearly defined, the project will fail before it even starts.”

“The value of AI doesn’t just come out of thin air,” says Dirk. “You need data, and that must be usable, available and correct.”

All too often, at the start of a project, it turns out that the necessary labels are missing, or that what teams thought they had is unusable in practice. “You have to objectively assess the maturity of your data,” says Benoît. “Without governance, insight into data ownership and privacy, you inevitably get stuck.”

Technical integration also often turns out to be a stumbling block. “Data live in silos,” says Dirk. “You have to break down those silos and let the data work together. Integration requires more than pipelines; it demands consistency, semantic coordination and often also organizational agreements.”

AI requires a multidisciplinary approach. “The biggest mistake that organizations make is thinking that it will work with just a data scientist and a data engineer,” according to Benoît. “You also need stakeholders from the business, professional experts, IT architects and analytic translators.” That last group is essential for making the translation between business expectations and technical models.

Dirk picks up on this: “The team must think about the adoption of the technology from day one. Besides building the model, you also have to explain, motivate and guide it. Otherwise you run up against resistance.” AI causes change – and change can only succeed if there is widespread support for it.

“Developing a proof of concept is one thing,” says Dirk. “Scaling up the solution is much more difficult right away.” Often a model relies on a limited dataset without considering processing in real time, integration or maintenance. “Then a model works well during the test phase, but doesn’t manage to hold its own when it goes live on a larger scale.”

According to Benoît, preparation is the key. “Think from the beginning about your intended final situation. If you want real-time applications, the infrastructure and dataflows must be prepared for it. Otherwise you start again from zero.” Developing a POC with the attitude ‘we’ll see afterwards’ is not the best idea, say the experts. “Work with the larger goal in mind.”

Ethics is not an optional step when AI is involved. “AI is fundamentally different from standard IT applications,” says Dirk.

“The outcome of a question to a LLM changes over time, even if the prompt remains unchanged.” Monitoring of the associated risks is therefore a must. “We do that with tools for content safety, bias detection and model drift,” explains Dirk.

The intent behind the use of an AI model plays a role too. Benoît: “You can use the same technology in an ethically acceptable or unacceptable way. Sentiment analysis, for example, is fine for generic trend reports, but you can also use it in a malicious way, for example to assess individual employees. So monitoring is also needed to stay compliant.”

AI systems become outdated. Data change, behavior shifts and performance falls. “And yet a budget for follow-up is seldom provided in an AI program,” says Benoît. “Without monitoring you have no overview of degradation, and without updates the solution doesn’t stay relevant.” Adoption too comes to a standstill if support of the application disappears after the go-live.

The solution? “See support as an extended development phase,” says Dirk. “That way the team remains actively involved, with room for feedback and further optimization. That too is an important lesson. An AI project doesn’t end with the release; that’s really when it begins.” Bear that in mind throughout the process, including the budget.

The greatest value of AI is in the application, but often that’s just where it goes wrong. Technology alone isn’t enough; companies must also include governance, scalability, ethics and adoption in the process. And in particular, anyone who starts with AI must keep learning. “Experience makes the difference,” Benoît concludes. “Not only to prevent mistakes, but also to become better. Every use case fine-tunes your approach.”