When I stood up the AI Center of Excellence at Bank Leumi, the mandate was broad and the runway was short. The temptation in that position is to open with a strategy deck. I opened with a list of who actually owns data access, model risk sign-off, and infrastructure budget — because none of the rest matters until you know who can say yes.
The first 30 days were entirely about mapping the org, not the technology. I sat with risk, compliance, infrastructure, and every business unit that had already tried an AI pilot on their own. Almost every one of those pilots had stalled in the same place: not the model, but the review process that had no clear owner.
By day 60 I had picked one visible, bounded use case — small enough to ship in a quarter, real enough that a business owner would notice if it worked. Everything the CoE would eventually standardize (evaluation harnesses, logging, a governance checklist) got built once, against that one use case, instead of designed in the abstract first.
Day 90 wasn't a platform launch. It was a working pilot, a documented review path other teams could actually reuse, and a short list of the three things that would break first at scale. That list became the CoE's actual roadmap — not the one from the original slide.