The first-90-days plan I used as a new product leader five years ago is mostly obsolete, and I say that having run it successfully more than once. It was a good plan for its era. The era changed. The old plan centered on people and roadmap because execution was the constraint. Execution is no longer the constraint. Judgment, cost, and quality are. So the plan changes.
Days 1 to 30: audit reality, not the deck
You get handed a roadmap, an org chart, and a stack of decks. Read them, then set them aside, because they describe the company as it presents itself, not as it is. Spend the first month closing the gap.
Use the product yourself on real tasks, the way a customer would, not a demo. You learn more about the true state of the product in five real sessions than in fifty slides. Sit in raw customer calls, unedited, before you adopt anyone's synthesis. Map what is actually produced by agents versus people, because you cannot reason about cost or quality until you know what generates the work. And find out whether anyone reads evals by asking the single most clarifying question available: what is the eval score on our most important AI feature, and which way is it trending?
Days 30 to 60: instrument what matters
You cannot lead what you cannot see, and most product orgs run on instruments that lag reality. Make cost per outcome visible by workflow. Pull compute and agent spend out of the aggregate cloud bill and attribute it to the workflows that incur it. You will almost certainly find at least one workflow quietly underwater, and that finding pays for the whole exercise.
Put eval scores with trend lines on every production AI feature, because that is the instrument that catches quality drift before revenue does, and revenue will not warn you in time. Then replace the status meeting with one that forces decisions. The fastest way to change what an org optimizes is to change what its leadership meeting is about.
Days 60 to 90: kill and defend
Now you have a true picture and working instruments. Kill the ceremonies that survive only by inertia, using one test: does this change a decision? If not, retire it and reinvest the time. Defend the narrow set that is truly yours: problem selection, the quality bar, and the few expensive-to-reverse decisions. Then run the exercise: if eight people started today with no legacy, which segment would they attack? Pick it and decide deliberately whether you defend it or cede it, because leaving it undecided is the same as ceding it.
The throughline: stop optimizing execution, which is cheap, and start owning judgment, cost, and quality, which is the job. Before your second week ends, write down the answer to one question: what is the eval score on our most important AI feature, and which way is it trending?