Why most AI projects fail before any model is involved
Most AI engagements are framed as model problems. In practice, the failure mode that arrives first is structural. Before a model is selected or a prompt is written, an organization has already decided — implicitly — what is worth producing, who is allowed to decide it, and how the result will be validated. When those decisions remain implicit, no model can compensate for their absence.
The structure that precedes the model
A model executes against a question. The question itself is the output of an organizational system that frames what counts as a question worth answering. When that framing is unstable, model output becomes unstable for reasons that have nothing to do with the model.
This is why two teams using the same tool, on similar data, produce results of different quality. The variance lives upstream — in the part of the operation that decides what to ask.
What collapses first
In engagements where AI is introduced into an operation that was never structured for it, the first thing to collapse is not the model. It is the relationship between teams about what the model is for. Each function holds a partial view of the intended use; coordination cost rises in the spaces between them.
A tool cannot resolve a disagreement that the organization has never made explicit.
Where the work actually begins
The work begins one layer beneath the tool — in the decision infrastructure that determines what is produced, by whom, on which evidence, and how the result is validated. This is the layer that allows AI to be integrated without compounding existing operational drift.
Once that layer exists, model selection becomes a small decision. Without it, model selection becomes the wrong conversation entirely.
AI does not fail because the technology is insufficient. It fails because it is asked to operate inside structures that were never written down. The interesting work is building those structures — quietly, and before the model is chosen.