Testing your AI receptionist before it goes live

Before real customers meet your assistant, meet it yourself. The test console gives you the exact same brain that answers calls and chats — here's the test script that finds problems before customers do.

Test 1 — The FAQs (2 min)

Ask your own FAQ questions back in natural phrasing: "do you charge for callouts?", "are you open Sundays?", "do you cover Morley?" You're checking it retrieves the right answer — not the phrasing you wrote.

Failure sign: correct topic, wrong details → fix the FAQ wording. "I'm not sure about that" on something you covered → check it saved.

Test 2 — The booking flow (3 min)

Say "can I book a boiler service?" and let it run. Check: it asks for a day → offers real slots (not invented ones) → takes your name and number → confirms. Then check the booking actually landed in the dashboard and your Google Calendar.

Failure sign: offers slots inside busy time → check calendar connection. Books without collecting contact details → shouldn't happen; report it if it does.

Test 3 — The boundary tests (2 min)

Try the things outside your knowledge: "do you do oil boilers?" (when you don't), "what's the best boiler for a 4-bed?" (judgement call), a complete non-sequitur. Correct behaviour: honest "let me get the team to confirm" + takes details. Bad behaviour: improvised facts — that means your boundaries need stating explicitly in knowledge notes ("we never give technical advice over the phone").

Test 4 — The human request (1 min)

Say "can I speak to a person?" Correct behaviour: no arguing — it offers a transfer or takes a prioritised message, immediately and graciously.

Test 5 — The weird stuff (2 min)

Rambling, mumbling, off-topic questions, a caller who won't say what they want. Check it stays patient, steers back ("happy to help with anything Hartley-related") and doesn't invent answers.

What you're really testing

Not "does it sound impressive" but "does it behave like a competent receptionist" — accurate on the routine, honest on the edge cases, graceful with humans needed. If all five tests pass, ship it. Then repeat the loop weekly from real transcripts — that's how good it gets.