Every leadership team we meet has spent money on AI by now. Licences, a pilot, a chatbot sitting on top of the data, sometimes an agency. Ask where the profit went up and the room goes quiet.

The technology works. Getting from a working model to a change in the numbers takes three things that rarely arrive together.

01 / 03

The data has to be clean

AI does not fix bad data. It repeats it faster and with more confidence.

What we find is usually ordinary. The same product has three names across three systems. Freight sits in one cost account for one market and a different one elsewhere. Returns are recorded when the parcel arrives, and sometimes when the refund clears. None of it is dramatic. All of it makes the answer wrong.

So the first job is unglamorous and unavoidable. Connect the sources, agree the definitions once, write them into the pipeline. This is much faster than it used to be. It still has to happen before anything else is worth doing.

02 / 03

Someone has to translate the problem

The gap generally lies somewhere between the problem and the tool.

A planner says we bought too much of the wrong thing. That is not a prompt. It is a question about forecast error by product group, about lead times, about how the buy was signed off and what that person could see at the time. Someone has to sit with the planner, find where the decision actually breaks, and then decide what to build.

This is usually where AI projects stop. The models are available to everyone. Knowing which one to point at which problem, and what the output has to look like before anyone will act on it, tends to be the scarce part.

03 / 03

It has to survive a normal Tuesday

Talking to your data demos well. It rarely changes a decision.

The tools that work arrive inside the routine that already exists. The buy sheet that opens with the recommendation already in it. The alert that fires when a product will run out before the next inbound lands. The margin view that is ready before the Monday meeting instead of being built during it.

That takes operational experience. You have to know how a buying cycle runs, when the numbers get locked, who signs what, and where people already spend their hours. Without that you build something technically correct that nobody opens twice.

We are two people holding different halves of this problem.

One half is scientific. A PhD in AI, published research, and years spent on machine learning that has to run in the real world rather than in a paper.

The other half is operational. Scaling a consumer brand through more than 500% growth, running technology and operations through an LVMH investment, and years of planning, logistics, and margin work.

Neither half gets there alone. The models need someone who knows where the money leaks. The operation needs someone who can build.

So, where is the profit? In the inventory you did not have to discount, the freight you stopped overpaying for, and the hours your team stopped spending on work a machine does better. It was always there. The work is building the thing that reaches it.