Why your AI pilot dies in September
Not because of the model. Because of the three things nobody budgeted for in January.
The pilot starts in January with enthusiasm, a demo that works and a reasonable budget. By September it has stalled. In the retrospective the model takes the blame, which is the most convenient explanation and almost always the wrong one.
The three missing line items
The first is integration with the system that already exists. Not the clean API someone imagines, but the ERP with twelve years of accumulated layers, the free-text fields used for things nobody documented, and the nightly jobs that can't be touched.
The second is who answers. As long as the pilot runs on test data nobody asks. The day output reaches accounting, the question arrives from three directions at once: legal, audit, management.
The third is maintenance. An agent isn't a project: it's a service. It has versions, regressions, recurring costs and someone who has to be reachable when it stops working on a Friday evening.
A pilot that hasn't budgeted for these three items isn't a pilot. It's a long demo.
None of the three is an artificial intelligence problem. All three are software engineering problems, which is exactly why this work is done better starting from enterprise than from research.