Can you trust your AI workflow?

Eight questions to ask

1. Is the answer supported? Compare it with the current approved information. Check what happens when the answer is missing or uncertain.
2. Did it do the right thing? Check the actual booking, message or record—not just what the AI says it did.
3. Did it use the right information? Check that the correct customer, order or document was selected.
4. Was approval respected? Check that a person is asked before any action your business requires them to approve.
5. What happens with an unusual request? Try missing details, conflicting instructions and requests outside the agreed scope.
6. Can a repeated request cause a duplicate? Check whether the same message, booking or update could happen twice.
7. Can the team recover from an interruption? If a connected service stops halfway, can someone see what happened and decide what to do next?
8. Does it still work after a change? Repeat the important checks after changing the model, instructions or connected tools.

Record what you find

For each check, note the request, expected result, actual result, evidence and next action. Record the system version and test conditions, and distinguish an issue from a check you could not complete.

Download the editable checklist (CSV). It includes two illustrative scenarios and space for your own checks.

Use an authorised test environment and fictional or appropriately protected data. Avoid sending real messages or changing live records unless those actions are explicitly agreed.

Understand the limits

A checklist is a starting point. A small set of checks cannot guarantee future results or replace a security, privacy or compliance assessment. Choose tests and review steps that fit the consequences of an error.