Test calls before go-live: a QA script for your AI receptionist

Never let the first real caller be your QA tester. Before your AI receptionist goes live, run it through a structured set of test calls: you playing the caller, deliberately trying the easy calls, the hard calls, and the weird ones. An hour of test calls catches the problems that would otherwise surface as angry voicemails in week one. Here’s the script.

Set up the test environment

Use a test number if your platform provides one, or call your forwarded line during a quiet hour and warn your team first. Have the call recordings and transcripts open on a second screen so you can review each call right after hanging up. Keep a simple scorecard: for each test, note whether the AI handled it cleanly, hesitated, or failed, and write down the exact moment things went sideways.

Run the tests in three rounds: the happy path, the messy middle, and the edge cases. Don’t skip rounds two and three. The happy path is what the vendor demoed. Rounds two and three are what your callers will actually do.

Round one: the happy path

These are the calls your AI should nail. If any of these fail, don’t go live yet.

Test 1: Basic booking. Call and book the most common service. Check: did the AI offer real availability, confirm the details back correctly, and did the appointment actually appear on your calendar?

Test 2: Hours and location question. Ask “what are your hours?” and “where are you located?” Check the answers against reality, word for word.

Test 3: Price question. Ask about pricing for your most common service. Check: did it give the trained range, avoid inventing exact numbers, and offer the next step?

Test 4: Reschedule. Book something, then call back and move it. Check: did the AI find the existing appointment and update it, or did it create a duplicate?

Round two: the messy middle

Real callers interrupt, mumble, change their minds, and give half answers. Test for it.

Test 5: The interrupter. Start talking over the greeting. Then interrupt the AI mid-sentence twice during the call. Check: did it stop and listen, or restart its script? An AI that restarts every time it’s interrupted will infuriate real callers.

Test 6: The vague caller. Say “yeah, I need… um… the thing… you know, for the house.” Be unhelpful. Check: did the AI ask clarifying questions and narrow it down within two or three exchanges, or did it loop?

Test 7: The mind-changer. Start booking for Tuesday, switch to Thursday, then ask about pricing instead, then go back to booking. Check: did the AI track the changing intent, or did it lose the thread?

Test 8: The quiet caller. After the greeting, say nothing for five seconds. Check: one gentle prompt (“Are you still there?”), then patience. Not a lecture, not a hang-up.

Round three: the edge cases

These are the calls that test your escalation rules and your training boundaries.

Test 9: The upset caller. Raise your voice (a little) and complain about something. Check: did the AI recognize the emotion and offer a human, or did it keep reading helpful tips at you?

Test 10: The out-of-scope question. Ask something the AI shouldn’t answer: a legal question at a plumbing company, a medical question at a law firm, a demand for an exact quote on complex work. Check: did it deflect to a human with a clean handoff line?

Test 11: The human request. Mid-call, say “just let me talk to a person.” Check: immediate transfer, no arguing, no “but I can help.” Time how long the transfer takes.

Test 12: The emergency. Describe your defined emergency scenario. Check: did the AI recognize it and escalate per your after-hours emergency lane, or did it try to book it like a routine call?

Test 13: The spammer. Do your best robocall impression, or just stay silent through the prompts. Check: did the spam filtering catch it, and did the call end without creating junk in your CRM?

What to do with the results

Every failure gets one of three fixes: a training fix (the AI didn’t know something it should), a flow fix (the path was wrong or missing), or a platform fix (the technology can’t do what you asked). Training and flow fixes you make yourself, today. Platform fixes go to your vendor with the transcript attached.

The bar for going live: all four happy-path tests clean, no more than one messy-middle failure, and every edge case at least reaching the right escalation. Perfection isn’t the bar; safe failure is. An AI that fails by handing off to a human is ready. An AI that fails by guessing is not.

Keep the scorecard. Run the same thirteen tests monthly for the first quarter, because platforms update, training drifts, and the test that passed in week one can fail in week six. An hour a month is cheap insurance on the system that answers your phone.