AI projects rarely fail because the model was wrong. They fail because the ground underneath was never built to hold one.
AI projects rarely fail because the model was wrong. They fail because the ground underneath was never built to hold one.
Every few months, a leadership team decides this is the year of AI. A budget is approved, a use case is chosen, a model is built. Then, quietly, it never reaches production. The post-mortem blames the model, the vendor or the team. The real cause sits one layer below.
AI runs on data. If that data is fragmented across systems, defined differently by every department, owned by no one and trusted by few, no model can rescue it. A prediction is only as good as the numbers feeding it, and in most organisations those numbers cannot survive a second question.
This is why the most valuable work is often the least glamorous. Before the model comes the foundation: a single source of truth, clear ownership, traceable lineage, quality you can prove. It is unfashionable work. It is also the work that decides whether the AI on top of it is an asset or an expensive demo.
Our advice is consistent, if unpopular: do not start with the model. Start with whether your data could survive being trusted. Get that right and AI becomes almost anticlimactic to add. Get it wrong and no amount of model sophistication will save the initiative.
That is the work behind the decision. It is where we start.