Almost every stalled procurement AI pilot gets diagnosed as a model problem. Almost none of them actually are.
The pattern is consistent enough to name: a team runs a promising proof of concept on a clean, curated sample of data, gets a confident demo, and greenlights a wider rollout. Then the model meets the organization’s actual material master, where the same item has five descriptions, three suppliers have no performance history on file, and half the historical pricing has no source attached to it. The model does not get worse. The ground it is standing on was never solid, and the demo simply had not exposed that yet.
Why this keeps happening
Two organizational habits drive it. AI initiatives tend to get funded and scoped around what the model can demonstrably do, not around whether the underlying data can support it at scale — data readiness is not exciting to put in a pitch deck. And data quality in procurement specifically has historically been nobody’s dedicated job. It is everyone’s problem in the abstract and no one’s responsibility in practice, which means it degrades quietly until an AI project finally makes it visible.
The part that makes it worse
Unlike a broken report or a stalled dashboard, an AI system trained or run on inconsistent data does not fail loudly. It produces an answer. The answer looks exactly as confident as a correct one would, which means the failure mode is not "the system does not work," it is "the system produces plausible-looking recommendations nobody can fully trust," and that erodes adoption far more slowly and far more damagingly than an outright crash would.
The fix is sequencing, not more sophisticated models: audit and structure the data first, then apply AI to what is now a reliable foundation.