01Start from one funded workflow, baseline metrics, expected value, guardrail assumptions, and an agreed decision rule for expansion.
02Document the inputs, outputs, source systems, owner group, handoffs, exception types, and actions that AI can prepare or recommend.
03List which inboxes, CRMs, ERPs, documents, spreadsheets, vertical systems, and data sources the workflow reads from or updates.
04Separate draft-only work, reviewer-approved actions, blocked actions, source-evidence requirements, and escalation owners before build starts.
05Plan prompt setup, tool permissions, integration work, historical test cases, messy examples, fallback states, audit logs, and reviewer signoff.
06Sequence discovery, design, build, testing, guarded launch, monitoring, and the first expansion review so the project does not drift.
07Train reviewers, tell users what changes, keep support paths visible, and measure adoption before expanding to more workflows or teams.
08Define post-launch monitoring, incident response, prompt changes, integration fixes, support ownership, ROI reporting, and managed optimization.