Your AI problem is an operating-model problem
Two years and a lot of spend later, the pilots that stall and the ones that scale differ on workflow and ownership, not model quality. Budget accordingly.
If your AI initiative is stuck, the model is almost certainly not the reason. The reason is that you attached a new capability to an operating model that was never redesigned to use it.
The evidence is hard to wave away
MIT's research found that roughly 95% of enterprise generative-AI pilots delivered no measurable impact on the P&L — and located the cause not in model quality but in the "learning gap" between the tool and the organization around it. The same research found the money pointed the wrong way: more than half of GenAI budgets went to sales and marketing, while the durable returns sat in the back office, where the friction — and therefore the savings — is largest.
The part executives resist
This is the part executives resist, because it implicates them rather than the vendor. A model is something you buy. An operating model is something you have to change: who approves what, where the handoffs are, which steps exist only because a human used to be slow. Drop a capable model into an unreconsidered process and you get an expensive pilot that impresses in the demo and moves nothing on the ledger. The 5% that worked did the unglamorous thing — they redesigned the process first and bought the model second.
None of this argues for less ambition. It argues for aiming the ambition at plumbing instead of theater. The back office does not make a good keynote. It makes a good quarter.
So before approving the next AI line item, ask which process it redesigns — not which model it licenses. If the answer is only the latter, you are funding a pilot, not a return.