Why I never talk about AI at the start of a project.

When an artificial intelligence project launches, the energy often goes toward the most visible part: the algorithm. The idea seems good, the team gets excited, everyone wants to see a demo. I understand this instinct very well. I also see many projects stall at exactly this point.

The appealing demo that never leaves the meeting room

Many AI projects begin with a demonstration of algorithmic feasibility. The same scenario comes up often.

A problem or pain point is identified.

An ROI is quickly estimated, in the best-case scenario.

A proof of concept, a POC, begins with a data sample. Sometimes the data is real. Often, it is simulated.

The result, in 90 % of cases in my experience, looks like a demo presented at a team meeting or slipped into a PowerPoint. Then the project slows down considerably. Moving from this demo to a production-ready solution requires far more effort than expected.

When it comes time to scale up, problems emerge quickly.

• Data that is poorly connected or difficult to use.
• An information system that is still too fragile to support AI.
• Costs that double, or more, with integration, testing and then maintenance of the solution.

A project that seemed feasible in two months with a budget of 60 000 euros can then turn into a year-long undertaking costing 200 000 euros. The original sponsor loses confidence. The team scatters. The idea gets shelved.

Data as the project's first test

When someone tells me about an AI project, my first question is about the data, never the algorithm. In 90 % of cases, it is still difficult to use directly, even when the client thinks otherwise.

I therefore prefer to begin with a POC that validates the data flow. A POC limited to the algorithm on a clean sample gives an overly reassuring picture. This data-focused step attracts less attention. It feels less like a quick win. Yet it prevents us from building an impressive demo on an unusable foundation.

Working on the data first makes it possible to identify technical or organizational obstacles quickly. It also lets us check whether the data holds up at scale. The algorithmic phase then begins under much better conditions, with fewer surprises and fewer makeshift fixes.

The statue analogy

Think of a statue. No one starts with a perfectly sculpted nose in the middle of an enormous block of granite. The work begins with the overall shape. The details come later, once the material has already been roughed out.

The same applies to data. Jumping straight into the AI component before preparing the data is like working on a fine detail in a block that is still raw. The project then consumes time and money before revealing, too late, that its foundation is too fragile.

A stronger project when constraints emerge early

There is another image I like. Dinner rarely begins with dessert. The most satisfying part comes after a minimum of preparation. In an AI project, the algorithm often plays that role. It is appealing. It impresses. It makes the subject tangible.

When the project begins with the data, the team understands sooner what it is getting into. Stakeholders see the real constraints. The sponsor better understands the effort required. And when the AI phase begins, it rests on a much more credible foundation.