An automation example only earns its place when it states what starts the run, which system changes and when it stops. The AI step should stay small enough to inspect.

Three examples worth testing

Task Trigger and action AI boundary Review and stop rule
New support form New approved form → create internal case Fixed-label classification Stop on missing consent or unlisted label
Invoice receipt Receipt reaches approved inbox → add row to expense queue Extract supplier and amount for review Stop if total or currency is unclear
Content request Approved brief → create draft task card Summarize the supplied brief Stop if source link or owner is absent

Worked example: form to internal case

Use a fictional form with consent set to yes. Map name, contact and message to an internal queue. Ask the model to choose only BILLING, ACCESS, BUG, FEATURE REQUEST or NEEDS REVIEW, then show the source quote to a reviewer. The output is an internal case, never a customer reply.

If the same test record creates two cases, treat that as a failure. Add a unique submission ID and a duplicate check before creating the case. Retrying without that control can make the problem worse.

When not to automate

Don’t automate a routine whose approver, policy or final state changes every time. Keep the manual workflow while you collect examples; a fast wrong handoff costs more than a slow visible one.

Questions about automation examples

Must every automation contain AI?

No. Use normal rules for stable fields and thresholds. Add AI only for a bounded unstructured step such as classification or extraction.

What data should the model receive?

Only the fields necessary for its step. Do not send an entire CRM record to classify one support message.

Can an automation send the final customer reply?

Not in a first test. Keep the result internal until a reviewer can measure errors and decide whether the risk justifies any external action.

For a fixed process with one AI step, see AI Workflow Examples.