Call your AI receptionist before a customer does

Why you should be the first caller

An AI receptionist answers your business line, books appointments, takes messages and answers common questions. Plenty of small businesses are trying them: clinics, salons, trades, restaurants, law offices. The setup usually takes an afternoon. You give it your hours, services, prices and calendar, it reads a sample greeting and it's live.

What the setup rarely includes is someone calling it the way real customers do. Not the polite test call where you ask for a Tuesday appointment and hang up pleased. The caller who wants a refund, the one who can't hear, the one whose pipe has burst at 11 p.m.

So before a customer does it for you, set a timer for thirty minutes, pick up your own phone and make five calls. Keep a notepad next to you. For each call, write down three things: what it said, what it did in your systems and who found out.

It helps if the person calling isn't the person who set it up. Whoever wrote the greeting and the answers knows what to ask to get a good result. Someone from the front desk, or a friend who's never heard the script, will ask the way customers do, and stumble in the places customers will.

Minutes 0–5: the easy booking

Call and book the most common thing you sell, the way a regular customer would. Give your name and number. Ask for a day that's busy.

Then hang up and check. Is the appointment in your calendar, at the right time, with the right service and length? Did you get a confirmation text or email, and does it say the same thing the AI said on the phone? Did it offer a slot that's actually free, or one that only looked free because a booking lives somewhere it can't see?

While you're there, call back and move the booking to another day, then cancel it. Check that the calendar followed both changes, and that the freed slot really opened up again. Rescheduling is where half-finished integrations tend to show.

If any of this fails, stop and fix it before the other four. Everything else rests on it.

Minutes 5–12: the question it shouldn't answer

Call and ask something that needs a person's judgment: "Can I get a refund if I cancel tomorrow?" "Will you match this other quote?" "Is this treatment safe if I'm pregnant?" "Can you guarantee it'll be fixed by Friday?"

Write down exactly what it says. An AI that's trying to be helpful will often produce a confident, complete answer. If that answer isn't your policy, or isn't something anyone should promise on the phone, you have a problem worth more than the subscription. The right behavior is to say plainly that a person will confirm and to take a message with the details.

Then check your settings and knowledge sources for where its answer came from. Often it's an old price list or a line from your website that nobody's updated. For more on keeping automated replies from inventing commitments, see A macro may not invent a promise.

Minutes 12–20: the caller who wants a person

Call twice. The first time, say early and clearly that you want to speak to a person. The second time, get frustrated halfway through, the way people do when they've already called once.

Does it offer a person, or keep trying to solve it? If it transfers, does someone actually answer, at that time of day? If nobody's free, does it take a message and say honestly when someone will call back?

Then find the handoff. Whatever reached a person, a transfer, a text, an email or a ticket, should say who called, their number, what they wanted, what the AI already told them and what it promised. If the person picking it up has to call the customer and start from zero, the handoff isn't working.

The call isn't over when the AI hangs up. It's over when the right person knows what was promised.

Minutes 20–30: the messy call and the urgent one

Make the messy call from somewhere noisy, or ask a colleague to call. Change your mind halfway through. Give a name that's easy to mishear. Ask two questions at once. See whether it confirms details back to you before booking, and what it does when it isn't sure.

Then the urgent one. Call after closing time, or set the system to its after-hours mode, and describe something that can't wait: a burst pipe, a reaction after a treatment, a lock-out, an order that hasn't arrived for a wedding tomorrow. Does it recognize that this can't wait until morning? Does it reach the person on call, or at least give the caller something real to do? If you have an emergency number or an on-call rota, it needs to know.

If your business is in a place with rules about recording calls or telling callers they're speaking with an automated system, check that the greeting does what you're required to do. That's a question for whoever advises you locally, not for the vendor's default script.

What to do with your notes

You'll end the half-hour with a page of notes. Sort them into three groups.

Fix today. Anything where it promised something you don't offer, booked something wrong or failed to reach a person on an urgent call. These are settings, knowledge or routing changes, and most can be made the same day.

Fix this week. Handoffs that arrived without enough detail, confirmations that didn't match the call and questions it answered vaguely.

Decide. The calls you'd rather a person always took. It's fine to have the AI answer, take the details and pass them on, with no attempt to solve them.

Then repeat the drill after every change to your prices, hours or services, and once a month regardless. Weekend cover deserves its own check; see When the agent fails on Saturday, who gets the packet?.

We build AI applications and agents that are connected to the calendar, records and people behind them, so handoffs carry what was promised. If you'd like help setting up or testing an AI on your phone line, talk to us.

Related work: Sutton Place.