Begin with a stable decision
Automation works when the team already agrees what should happen after an event. A dependency alert can request an update; a nightly check can refresh generated files. Ambiguous triage is harder because the trigger does not contain enough product judgment.
Design the failure path first
Every run needs a timeout, a visible status, and a person or queue that receives failures. Silent retries turn a small automation into background operational debt.
Measure avoided effort
Run count is not value. Track how many outputs are accepted, how much review they need, and how often a human must restart the task. A trigger earns its place when it reduces total work without hiding risk.