Most automation projects stall because they start with a tool and a broad mandate. This playbook starts with one workflow, one owner and one number, and gives each week a single job.

Week 1 — Choose and measure

Goal: Pick the workflow and capture how it performs today.

  1. List three recurring workflows and score each on frequency, hours and pain.
  2. Pick one with a clear trigger, reachable information and a single owner.
  3. Record a baseline for the week: time per run, errors, and waiting time.
  4. Write down the one number that will decide whether this worked.

You finish with: A one-page workflow brief with an owner and a baseline.

Week 2 — Map and bound

Goal: Decide exactly what the automation may and may not do.

  1. Map every step, input, system and handoff on one page.
  2. Mark the steps that need human approval: anything external, sensitive or irreversible.
  3. Define the stop rule: what makes the workflow hand back to a person.
  4. Confirm access: which systems it can read and change, with scoped credentials.

You finish with: A bounded design with approval points and a stop rule.

Week 3 — Configure, connect, automate

Goal: Build the smallest version that runs end to end.

  1. Use the tools you already have first: configure AI assistants and connectors.
  2. Connect the systems with the narrowest access that works.
  3. Automate the repeatable middle; keep approvals as explicit steps.
  4. Run it in parallel with the manual process on real work.

You finish with: A working workflow running alongside the old process.

Week 4 — Validate and hand off

Goal: Prove the result and make it someone’s job to run.

  1. Compare against the week-1 baseline on the same measure.
  2. Review every exception and adjust the rules or the stop rule.
  3. Document how it works, who owns it and how to pause it.
  4. Decide: keep, adjust, or retire — and what to automate next.

You finish with: A measured result, a runbook and an owner.

Common pitfalls

  • Starting with a platform instead of a workflow.
  • Automating a process nobody owns.
  • Skipping the baseline, so the result cannot be proven.
  • Removing approval steps before the results justify it.

More guides like this

Related answers

Read as Markdown Automate the repeatable