An automation is defined as much by what it does when inputs are strange as by its happy path. If every missing field triggers a confident default, the workflow can scale mistakes faster than routine work.
Map exceptions before connecting tools
For each step, list expected inputs, valid range, and owner. Then define what happens when the input is missing, duplicated, stale, contradictory, out of scope, or sensitive. Assign one of four outcomes: continue with a safe default; ask the user; send to a human queue; or stop and alert an owner.
Example: routing a campaign inquiry
- Missing market → ask which country; do not infer from dialect or IP.
- Service outside coverage → explain the limit or route to an approved alternative; do not create a booking.
- Duplicate submission → link to the existing case and prevent repeated follow-up.
- Unsupported claim in free text → preserve it as a customer statement, not an approved product fact.
- System timeout → retry within a defined limit, then alert a person; do not silently drop the inquiry.
Make the fallback observable
Log exception type, source, workflow version, action taken, and whether the case was resolved. Minimize personal data in logs. Give people a clear route to correct information or reach a human. Test each exception with synthetic or anonymized cases before enabling external actions.
Review exception frequency and false alarms after launch. If one branch dominates, improve the form or rules rather than adding more model instructions. Keep sending, spending, deleting, and account-permission changes behind explicit human approval unless the organization has a separately approved control.
AI disclosure: AI tools assisted with drafting. The routing examples are illustrative, not tested automation results.