Guardrails & fallbacks
An automation without a fallback is a liability. Add the rails before you add the throttle.
A workflow that succeeds in a demo can still fail on an unexpected input or a busy day. Guardrails catch problems before they spread. Fallbacks give the work somewhere useful to go when the normal route fails.
Four checks to build in
- Input validation: check required fields, formats and allowed values before processing the request.
- Output validation: check the result against the required structure and the task’s acceptance criteria. A correctly formatted answer can still be wrong.
- Rate limits: control your own request volume and handle the service’s limits explicitly.
- Human checkpoints: route ambiguous or sensitive work, and failed checks, to a person. Set these around the cost of a wrong action rather than a model’s confidence score alone.
Choose the right fallback
The response should match the failure. These options form a useful starting point:
- Retry: limit attempts and increase the delay between them for temporary failures. Before repeating an action that sends a message or changes a record, check whether it already succeeded.
- Simplify: use a reduced version of the task only when its result remains useful and acceptable.
- Hand over: send the original input, attempted steps and failure details to a named person or queue.
- Stop: provide a way to halt further actions and notify the owner.
These are options, not a sequence every failure must follow. A sensitive or irreversible action may need to stop immediately for review.
Log enough to investigate
Record decision points, failure details, retry counts and response times. Keep only the input details needed to investigate, and avoid unnecessary personal or sensitive information. Alert on rising failure rates, repeated retries and unexpected outputs.
Ask who would notice
Before release, ask: “If this failed silently for a week, who would notice, and what would the impact be?” Name the owner, add a check that would detect the failure, and decide when a person must intervene.
Trust comes from failures you can detect, explain and recover from.