Two different questions wearing one label
"How's the AI pilot going?" is really two separate questions collapsed into one, and most organizations only answer the first. Did the model work is a technical question with a satisfying answer. Did it change what the P&L looks like is a business question that most pilots were never scoped to answer, because nobody assigned it an owner or a number before the work started.
The gap between those two questions is where the bulk of enterprise AI spending currently sits: technically successful, financially unaccounted for, and quietly renewed every quarter because killing it would mean admitting the first question was the wrong one to have asked.
Where the money actually shows up
When AI investment does move the P&L, it shows up in a small number of places: cost lines that shrink because a task no longer needs the headcount it used to, revenue lines that grow because a decision gets made faster or more often than a person could make it alone, or risk lines that shrink because an error gets caught before it compounds. Every pilot worth funding should be able to name which of those three it's aimed at before it starts.
Most can't, not because the teams running them are careless, but because the pilot was approved as a capability demonstration, and capability demonstrations don't need a P&L line to justify their existence. Production does.
A pilot that impresses a steering committee and a pilot that moves the P&L are different projects wearing the same name.
Raef Khan, Chief Executive Officer
The handoff is the failure point, not the model
We've reviewed enough stalled AI programs to know the pattern by now: the model performs well in the pilot, the readout lands well with leadership, and then the work quietly stops moving because nobody owns turning a working model into a line item someone is accountable for. The pilot team disbands or moves to the next initiative, and the operating team that would need to absorb the change was never in the room to begin with.
Fixing this is less about better models and more about a different approval gate: no pilot gets funded without naming the P&L line it's meant to move and the operating owner who will be accountable for moving it, before a single prototype is built.
What we tell boards to ask instead
Stop asking how many pilots are running. Ask which P&L lines have moved because of them, by how much, and who is accountable for the next one. That single change in the question forces every subsequent pilot to be scoped against an answer instead of a demonstration.
It also does something boards underestimate: it makes it much easier to kill the pilots that were never going to move anything, freeing the budget and attention for the two or three that actually will.
