Start with the constraint, not the shiny demo
Great pilots live in the real bottleneck that leadership already loses sleep over. We map the workflow, the systems of record, and the people who approve the work today, then decide where an AI copilot can relieve the bottleneck without requiring permission from five extra departments.
That usually means cutting scope aggressively: automate one approval, summarize one dataset, or draft one outbound sequence. Everything else is a “later” column on the wall.
Instrument on day one so success is obvious
A pilot that cannot show improved cycle time, lower handle time, or higher throughput is just a cool demo. Before we open Cursor or Replit we wire basic telemetry: start and stop timestamps, volume counts, error codes, and a short human satisfaction pulse.
- →Dashboards go into the tools leaders already refresh (usually Notion, Coda, or Looker).
- →Each metric has a human owner who agrees to update or green-light it weekly.
- →We log every handoff where the copilot punted to a human so we can tune prompts with real data.
Protect the humans who keep the pilot alive
Pilots die when they create more emotional labor for operators. We write playbooks for what to do when the copilot stalls, how to escalate, and how to roll back without paging engineering.
Weekly office hours with the operators become mandatory. They show us the messy edge cases and we fold that back into prompt libraries or lightweight automations before the next run.