An AI receptionist should be tested against the calls a business actually receives, including the difficult ones. A polished demonstration is useful, but it does not show what happens when a caller is unclear, a calendar is unavailable or nobody answers a transfer.
Our AI receptionist introduction explains the broader use case. This checklist focuses on deciding whether a particular setup is ready.
Write a call boundary
List the requests the assistant may handle and those it must pass to a person. Define what it can confirm, what it can only record and what it must never promise.
Use approved business information. If a service is unavailable in a location, the assistant should not improvise an exception. If it cannot establish a booking, it should not describe a requested time as confirmed.
Give each boundary an example call and an expected outcome.
Test uncertainty deliberately
Include background noise, an unfamiliar name, a correction halfway through the call and a request involving two different services. Test a caller who declines to provide an optional detail.
For a hypothetical repair business, ask for a booking outside the service area and then change the address. Check which address reaches the staff record.
Test a failed calendar connection. The acceptable fallback might be a clearly described callback request, with the caller told that no appointment has been confirmed.
Check the human handover
Test transfers during staffed hours, after hours and when the destination does not answer. Confirm that staff receive enough context to continue without making the caller repeat everything.
Decide what to do with urgent or sensitive requests. Keep the assistant within administrative boundaries and use the business’s approved escalation instructions.
Review the notice and consent arrangements for any recording or transcription with the person responsible for those requirements.
Set acceptance criteria
Record the test, expected result, actual result, evidence, severity and owner. Correct serious failures and repeat the affected cases before launch.
Do not reduce the decision to a single accuracy percentage. One false booking confirmation can matter more than several minor wording differences.
Start with a controlled rollout and a working manual alternative. Give someone responsibility for reviewing failures and updating the approved information.
DonutLabs’ AI receptionist service can help define and test these boundaries. The useful outcome is a system whose limitations are understood by both the business and the people calling it.
For the wider planning context, read How to choose your first business AI workflow.
