Start narrow enough to prove ROI, then expand wide enough to become infrastructure.
The accelerator should not launch as “AI services.” It should launch as a proof-driven implementation engine with a first wedge, a pilot package, and a compounding monthly path.