What an AI receptionist can't do

The demo always goes well. You call the number, it answers in a normal voice, you ask it something reasonable and it handles it cleanly, and you hang up thinking the only question left is whether to buy one.

That's the wrong question to have answered. Before you put one on your main line, what you want to know is where it stops — because the places it stops are the places your customers are going to find.

It only knows what you told it

There is nothing in it that came from your head. It doesn't know you quit doing that kind of work two years ago, that you no longer service that brand, that the Tuesday guy left, that you don't drive past the county line for anything under a certain size. Every one of those gaps in the briefing becomes a gap on a live call with a real customer.

Which means the briefing is the product and the voice is the packaging. If setup was a web form and a few dropdowns, a web form and a few dropdowns is what your callers get. Somebody has to sit down and write out how you actually answer the phone, and then keep writing it as the business changes.

It should not make commitments

There is a category of thing that must never come out of an automated system on your behalf:

Pricing you negotiate. Published, fixed prices are fine. Anything that depends on what you find when you get there is not.
Timelines. "We can be out in the morning" is a promise made against a schedule it can't see and you have to keep.
Anything with liability attached. Whether the warranty covers it, whether the symptom is urgent, whether the policy pays. Not its call, and in some trades not even legally yours to delegate.
Anything that becomes a contract when said out loud.

The right behavior at the edge of the briefing is not a best guess. It's a stop: "I don't want to give you the wrong number on that — let me have Chris call you back," followed by a message that actually reaches you with a name and a number on it. A captured lead beats a confident wrong answer every time.

Some calls need a person, now

Water coming through a ceiling. No heat with a newborn in the house. A customer already halfway through an escalation who will get worse, not better, talking to software. The regular who calls once a year with a big job and will not deal with a machine. A conversation that has turned into a negotiation.

None of that should be handled with a message. It should be a live transfer during business hours — the system rings your cell, tells you who's calling and why, and you decide whether to take it. Design that path first and the rest is easier.

The failure you'll never hear about

Here's the one that matters, and it isn't the obvious one. A wrong answer the caller argues with is recoverable — they push back, it takes a message, you find out. The expensive failure is a confident wrong answer the caller believes.

They don't complain. They hang up satisfied, and then act on it: they show up Saturday when you're closed, they haul in equipment you don't work on, they repeat a price to a friend that you have never charged in your life. From their side the call went perfectly. You have no idea it happened, and the first evidence arrives as a person standing at your counter being told something different.

That's why the only setting that really matters is which way the thing fails. A system tuned to be maximally helpful will fill gaps with plausible-sounding material. A system tuned to take a message will not. Before you buy, ask the vendor one question — what does it do when it doesn't know? — and treat any answer other than "stops and takes a message" as a warning.

What to do before you turn one on

Write out the ten questions you actually get, with the answers you'd want a new hire to give word for word.
Write the second list: what it must never say. Prices you negotiate, timelines you can't guarantee, anything resembling advice.
Decide the transfer rule — who, what hours, and what counts as urgent enough to interrupt you.
Read or listen to the first two weeks of calls. All of them. It's an hour of your life and it's the highest-value hour in the whole project.
Re-brief it monthly for the first quarter. A question you didn't anticipate is an update, not a defect.

Run that way, it does one job and does it well: nobody reaches a ringing phone, and everything it couldn't handle lands in your inbox as a message with a name, a number, and what they wanted. That's the value. Anything sold as more than that is being sold by someone who hasn't listened to enough of their own calls.

Mine is briefed by hand and I listen to how it handles real callers before it goes live — that's the part that makes it worth having.