Five concepts. Then back to the canvas.

The specific event, schedule, request or condition that makes the workflow start. If you cannot name it precisely, you do not have a workflow yet.

Example: A new email lands in the shared support inbox — not 'when customers contact us'.

Everything the workflow must read before it can reason: records, documents, messages, policies and prior state. Missing context is the most common cause of bad automation.

Example: The email body plus the customer's plan and open tickets from the CRM, plus the refund policy document.

The judgement you are delegating to AI — classify, extract, compare, summarise, prioritise, draft or predict. State it as a job, not as a prompt.

Example: Classify the request, extract the order number, and judge whether the policy fully answers it.

What actually changes in a system when the workflow runs: a record created, a message drafted, a task routed, a report stored. No action means no automation.

Example: Create the ticket, label the mail, draft the reply, alert the team lead in Slack.

The approvals, thresholds, failure paths and measures that keep the workflow safe and improvable. Controls are designed at the same time as the action, never afterwards.

Example: Refunds above 100 USD need lead approval; confidence under 80% goes to a human; first response time is tracked weekly.