Data and AIArticle · 1 min read
Why AI projects fail before the model
The conversation always starts with the algorithm. But what sinks most projects happens weeks earlier, in a place nobody wants to look: the data the company already has.
In almost every first meeting about artificial intelligence, the question that comes up is which model to use. It is the wrong question, and it is expensive: the team spends weeks comparing approaches only to find out, at the first real training run, that half the records are missing the field the model needs.
This is not a tooling problem. Operational data was created to keep the operation running, not to answer questions. The customer record exists to issue the invoice, not to segment customers.
The model is the easy part
Across the projects we delivered over the last two years, training and evaluating the model rarely took more than a sixth of the effort. The rest was understanding the business and fixing the data.
With no owner for the data, every report turns into an argument about the report.
The most reliable sign that the problem is not technical
If you are evaluating an AI project right now, do the opposite of what is usually proposed: ask for a short, paid data assessment before any fixed scope. If the vendor will not look at the data before selling you a model, that is the signal.