Nearly every organization we talk to has run an AI pilot. Very few have put one into production. The gap between those two facts is where most AI budgets currently go to die, and the reason is more mundane than the discourse suggests.
It is not model capability. The models are, for most business tasks, already good enough. The blocker is that the pilot was never attached to a decision anyone was willing to let it affect.
Demos are not pilots
A demo answers "can it do this?" A pilot answers "should we change how we work because of this?" Those require completely different setups, and organizations routinely run the first while believing they are running the second.
The tell is what happens when the demo succeeds. If the answer to "it worked, now what?" is a pause, the pilot was a demo. There was no decision waiting on the other side of the result, no owner prepared to change a process, no baseline to compare against. It was an impressive artifact, and impressive artifacts do not survive contact with a budget cycle.
Scope the authority, not just the task
The question that actually unblocks AI projects is not what the system can do. It is what the system is allowed to do on its own.
There is a spectrum here, and being explicit about where you are sitting resolves most of the anxiety that stalls these projects:
- Drafts for a human. The system produces something a person reviews and sends. Low risk, immediate value, and the easiest thing to put into production. Most organizations should start here and a surprising number should stop here.
- Routes and prioritizes, human decides. The system triages, classifies, or ranks, and a person acts. Valuable in high-volume work, and the failure mode is a bad ordering rather than a bad outcome.
- Acts within a defined boundary. The system takes action in cases matching explicit criteria and escalates everything else. This requires real confidence in the boundary and real logging, but it is where the compounding returns are.
Most stalled pilots are stuck because the organization was implicitly imagining the third while only being comfortable with the first, and nobody said so out loud. Naming the level turns a philosophical argument into a scoping decision.
The baseline problem
"It seems better" is not a result, but it is the most common outcome of an AI pilot. That is a measurement failure, and it is avoidable.
Before the pilot, capture how the work performs today: how long it takes, how often it is wrong, how much rework it generates, how consistent it is across staff. This is tedious and always harder than expected, because most organizations have never measured the process at all. That tedium is the point — it is also why the pilot has nowhere to land.
Without a baseline you cannot distinguish a genuine improvement from the novelty effect of a new tool, and you cannot defend the budget when someone asks what you got for it.
Where the value has actually been
The AI work we have seen deliver durable value is unglamorous. Document intake and classification. Drafting routine correspondence that a person reviews. Extracting structured data from unstructured submissions. Summarizing long records so staff can act faster. Surfacing the three cases in four hundred that need human attention.
None of that makes a compelling conference talk. All of it reduces the amount of time skilled people spend on work that does not need their skill, which is the entire proposition.
The organizations getting value from AI are not the ones with the most sophisticated implementations. They are the ones that picked a specific, high-volume, low-judgment task, defined what the system was allowed to do with it, measured what it replaced, and shipped it. Then did it again.
Start with the decision. The technology is the easy part now.