The point of this playbook
Two measured workflows, a real operating cadence, and a clear decision about what deserves to continue.
01
Days 1–10: choose two workflows
Do not launch another broad AI initiative. Pick two recurring workflows with visible friction, an accountable owner, and enough volume to measure.
- Write the current workflow in five to seven steps.
- Choose one baseline: time, cycle time, cost, quality, or rework.
- Name the risk that must remain under human review.
- Define a successful result before anyone changes the process.
02
Days 11–45: run the smallest useful test
Change one part of each workflow, not everything at once. Weekly evidence reviews matter more than a large launch meeting.
- Run the new process on representative work.
- Capture the baseline and result for every attempt.
- Log failures, workarounds, and the human review required.
- Meet for twenty minutes each week to adjust or stop.
03
Days 46–70: make it repeatable
A successful experiment becomes useful only when someone other than the inventor can run it reliably.
- Document the trigger, inputs, steps, quality check, and owner.
- Test with at least two people who did not design the workflow.
- Remove unnecessary prompts and manual handoffs.
- Confirm the process still meets security and data rules.
04
Days 71–90: promote or kill
At week ten, force the decision. Continue only when the evidence and ownership are strong enough to survive normal work.
- Promote: the measure improved, quality held, and an owner exists.
- Revise: the signal is promising but one known constraint remains.
- Kill: results are weak, risk is too high, or the process depends on heroics.
- Publish the decision and what the organization learned.