Introducing agents to a regulated team: a 90-day plan
Every team we deploy asks the same question in week one: how much should we let the agents do? The honest answer is "less than the demo suggested, then more than you expect" — and the difference is a deliberate 90-day ramp.
Days 0–30: one case type, suggest-only
Pick a single, high-volume, low-variance case type and run agents in suggest-only mode. Nothing ships without a person; the point is to build a corpus of agent recommendations your team has graded. By day 30 you know exactly where the agents are already right.
Days 31–60: gate the decisions, free the admin
Let agents act on the coordination work — chasing, filing, diarising, drafting — while every consequential decision keeps a human gate. This is where cycle time falls, because queue time was never about the decisions.
Days 61–90: autonomy where it’s earned
Widen autonomy per action, using the grading data from the first month: actions with sustained agreement go autonomous, the rest stay gated. Write the thresholds down — the regulator will ask, and so will your own risk team.