The pilot almost always works. That's not a compliment to the technology — it's a symptom of how pilots are scoped: small data, motivated users, a sponsor paying close attention. The failure shows up later, in production, when none of those three conditions hold anymore.
We tracked forty AI initiatives across eleven enterprises from pilot to their first anniversary in production. Twenty-six stalled or were quietly shelved. Of those twenty-six, twenty-one failed for the same reason: nobody had planned for what happens when the model is wrong.
A pilot proves a model can be right. Production proves an organization knows what to do when it isn't.
The fourteen initiatives that made it past their first year shared a pattern that had nothing to do with model architecture: each had an explicit, pre-agreed escalation path for low-confidence or contested outputs, owned by someone with the authority to override the system without a meeting.
That's a governance decision, made before launch, not a technical fix applied after a public failure. It's the single clearest predictor we found of whether an AI initiative survives contact with a real customer.