Three lessons from our AI projects

Over two years, we supported dozens of AI projects, from initial exploration to a model that runs every night. Some went exactly as planned; others stalled because of something no one had anticipated. Looking back at what made the difference, we keep coming back to the same three lessons.

Lesson 1: start with the decision, not the model

The question that makes a project succeed is not which model performs best, but which decision will be made differently as a result. Once that decision is clear, we also know who the user is, how often a prediction is needed and how accurate it needs to be. Projects that start by choosing a model often produce a polished demo that then fits into no one’s working day.

Lesson 2: data quality is an organisational issue

Almost every project that was delayed ran into problems with data, not technology. Columns whose meaning changed halfway through the year, manual corrections in an export file, or two departments defining the same concept differently. You cannot fix that with a data-cleaning script. Someone in the organisation needs to own the definition and have the authority to establish it.

Lesson 3: plan the handover from day one

A model that no one can explain will not be used. That is why we agree at the outset who will manage the model, how often it will be retrained and which signals should prompt action. Those agreements take an hour in the first week and save months of work after delivery.

None of these lessons is about algorithms. They are about clarifying the question, taking the data seriously and arranging management before it becomes necessary. Get those three things right, and you will usually find that the technology is the easiest part of the project. If you would like to talk about this further, feel free to get in touch; we are happy to share what we encountered along the way.