Validate what the model returns
Ask for structured output (a fixed set of fields or categories) and check it before using it. If the model was supposed to return one of four labels and returned a paragraph, stop the run instead of guessing.
Give every branch a safe default
When the automation is unsure, the safest action is usually to do less: save a draft instead of sending, mark for review instead of filing, notify you instead of replying.
Retry the flaky parts, not the whole run
Network calls and rate limits fail temporarily. Retry those steps a few times with a pause. Do not retry steps that already sent something — that is how customers get the same email three times.
Alert on failure and on silence
- Send yourself a message when a run errors.
- Also alert when nothing has run for longer than expected — a broken trigger fails silently.
- Include the input that caused the problem so you can reproduce it.
Test with ugly inputs
Before going live, run the workflow with an empty form, a very long message, another language, and a message that tries to give the AI instructions. Those are the inputs real life will send you.
Before you call it done
- Structured output, validated before use
- A safe default on every uncertain branch
- Retries only on steps that are safe to repeat
- Alerts for errors and for unexpected silence
- Tested with empty, long, foreign and adversarial inputs
A free master guide and resource pack for people starting with AI automation: the systems behind voice agents, lead workflows and a small automation practice.
Get it free on Gumroad