How to test an AI receptionist before using it with customers

Test an AI receptionist with awkward calls, not one cooperative script. A useful test checks routing, known facts, corrections, unsupported requests, failure handling, and the job card that reaches the owner.

Define the expected handover first

Write down the fields you need: caller, return number, issue, location, timing, access, and unknowns. Define what counts as complete, partial, and failed. Without that target, a fluent conversation can look successful even when the callback record is unusable.

Run six calls

1. Complete enquiry: give all required details naturally and check the service does not ask for them twice.
2. Interruption: answer a later question before the current one finishes. Check whether confirmed facts remain intact.
3. Silence: stop responding. Look for a bounded prompt and honest close.
4. Correction: give the wrong house number, then correct it. The final card should not preserve both as equally valid.
5. Unsupported request: ask for a quote or guaranteed arrival time. The service should leave that decision with the owner.
6. Connection failure: end the call mid-enquiry. The result should be partial or failed, not presented as complete.

Inspect more than the voice

Check which number was called, which business instructions governed the conversation, and whether one call created at most one job card. Read the card as if you were between jobs on a phone. Can you see the problem, place, timing, and gaps quickly?

Test shutdown too

Confirm how to stop routing before using the service with customers. Dedicated numbers and carrier forwarding have different controls. Verify exact forwarding steps with the provider rather than relying on generic codes.

Max & Mia's target behaviour covers these recovery and boundary cases, but acceptance criteria are not production performance evidence. Re-run the suite after instruction, routing, model, or provider changes.

Start by defining your expected output in the free job-card builder.