AI & automation / Checklist
Ten calls to make before your AI receptionist goes live
I'd want to hear an AI receptionist get something wrong during testing. A caller changes their mind, the calendar stops responding, or someone asks for a price the business hasn't agreed. That's when you find out whether it's ready to answer your phone.
Set aside a test number and use made-up customer details. Agree which calendar and inbox will receive the results, and make sure nobody mistakes a test booking for real work. Then make these calls yourself, from a phone.
Write down what counts as a pass
For a message-taking call, a pass might mean the right name, callback number, reason for calling and a clear note in the agreed inbox. For an appointment, check the actual calendar entry, its time zone and the confirmation the caller receives.
Keep the expected result beside each call. Otherwise it's easy to forgive a missing detail because the voice sounded good. I'd treat a wrong booking, invented price or lost callback number as something to fix before launch.
ElevenLabs documents separate tests for replies, tool calls and whole conversations. That distinction is useful whichever platform you choose: an agent can say the right words while sending the wrong information to another system.
Source: ElevenLabs: agent testing
The ten-call script
Change the examples to match the calls your business receives. These are proposed tests, not results from a customer installation.
| Call to make | What to check |
|---|---|
| 1. Ask for opening hours and the address. | The answer matches your current business details, including any appointment-only arrangements. |
| 2. Leave a normal new enquiry. | The message reaches the right place with enough detail to call back. |
| 3. Correct a digit in your phone number. | The final saved number includes the correction. |
| 4. Ask about a service you do not offer. | The agent admits the limit and follows your agreed next step. |
| 5. Ask for an unlisted discount or a firm quote. | It avoids making a price or promise up. |
| 6. Ask to speak to a person. | The agreed transfer or callback process works. |
| 7. Call when nobody is available to take a transfer. | The caller gets a clear fallback and a message is saved. |
| 8. Request an appointment, then change the time. | The final entry is correct, with no unwanted duplicate. |
| 9. Test a controlled calendar or integration failure. | It explains that the action could not be completed and offers the agreed fallback. |
| 10. Interrupt, pause and change the subject. | The agent follows the correction and leaves an accurate message. |
Arrange failures with whoever builds it
Ask the person configuring the agent to simulate an unavailable calendar or failed transfer in a safe test setup. Disconnecting your live calendar to see what happens can affect real bookings. Test integrations can also send real messages, so agree their destinations first.
If your service only takes messages, replace the booking calls with your own awkward cases. Try a caller who won't leave an email address, or someone asking for an update without enough information to identify their enquiry. The test should reflect what you've actually bought.
Listen again, then check the saved result
Use a simple record: scenario, expected outcome, what happened, where the evidence is and what needs fixing. If recordings or transcripts are available in the agreed setup, compare them with the message. Don't put real customer details into a shared test sheet.
Repeat a failed call after the fix, then try a slightly different version. Once it is live, review early calls and add new failure cases to the same list. A change to booking rules deserves another booking test; a new voice alone does not prove the workflow still works.
Sources
Checked on 22 October 2026.
- ElevenLabs: agent testing
The distinction between response, tool-call and conversation testing. The ten-call checklist is our suggested acceptance process.