Week 1: pick one process that is frequent, rule-based and low-risk (e.g., routing inbound requests, summarizing weekly reports). Record the current baseline: volume, time per item, error rate.
Weeks 2–3: build the agent with a workflow or orchestration tool, run it in shadow mode alongside the human process, and log every disagreement.
Week 4: compare against the baseline. Decide to expand, adjust, or stop. Write down the decision and the numbers so the next pilot starts smarter.