Day one on an AI program is dangerous because everyone already has an opinion about the model and nobody has the same definition of success. My first month is deliberately boring. I map reality before I touch the roadmap.
Week one is inventory: data sources, model dependencies, current eval metrics, legal constraints, and the last time leadership agreed on the goal. I write it up in one page and send it to the sponsor. Disagreements surfaced early are cheap. Disagreements surfaced at launch are expensive.
Week two is stakeholder alignment. I meet the people who can block silently: security, legal, infra, support, and the team that owns the workflow where AI will appear. I ask what would make them say no. Those answers become program requirements, not meeting notes.
Week three is the first cut of sequencing. I pick one thin vertical slice that proves value and one compliance gate that must pass before scale. I resist the portfolio slide that shows five AI bets at once. One proof point, one gate, one date.
Week four is the operating rhythm: weekly adoption and quality review, decision log started, risk register with model-specific failures included. If those exist by day thirty, the program can absorb speed. If not, every sprint will feel fast and the quarter will still slip.