It is tempting to start an automation project by picking the tool. Resist it. Tools are interchangeable; the workflow is not.
Draw the work as it is
Write down every step, who does it, and what it touches. Include the copy-paste, the “let me just check” and the spreadsheet nobody owns. The drawing is usually the first useful output of the project.
Delete before you automate
Every step you remove is a step that can’t break. Ask of each one: does anything downstream depend on it? If not, delete it. Automating a pointless step makes the pointless step faster.
Automate the boring, assist the judgement
Deterministic work — moving data, notifying people, creating records — belongs in a workflow with retries and monitoring. Judgement work can use an AI step, but with limits: clear inputs, reviewed outputs and a human who can override.
Make failure visible
A workflow nobody watches fails silently. Log every run, alert on errors and give the team one place to see what happened. Reliability is a feature of the automation, not an afterthought.