An AI employee answers. A person still decides.

A call comes in at seven in the evening. Nobody is in the office. Without something answering, the call reaches voicemail, and a voicemail from a stranger competes with everything else in somebody’s morning. Instead something answers, asks what the job is, offers Thursday at ten, and puts it in the calendar. By morning there is a booking and a transcript. That is an AI employee doing exactly what it is sold to do.

The second call is the interesting one. The caller has a burst geyser, a landlord, and a question about who pays for what. There is no slot to offer and no script that fits. It will still answer, still sound calm, and still get it wrong, because nothing inside it is judging anything.

What is an AI employee?

A set of tools that handle the routine end of a conversation. It answers calls and messages. It asks a fixed set of questions, books time that is already free, and sends the confirmation. Then it writes what happened into the record. The word employee is marketing. Nothing is being hired, and nothing is exercising judgement.

Underneath it are four ordinary parts. A voice that sounds human. A set of rules about what to ask. A calendar it may write to, and somewhere to file the outcome. Each part has existed for years. What changed is that the voice stopped sounding like a menu.

The distinction worth holding is between answering and deciding. Answering is repeating what it was told, in the right order, at three in the morning. Deciding is weighing two things nobody told it about. Every convincing demonstration you will be shown is the first one, because the first one is what it does well.

Where it earns its keep, and where it does not

• The evening call. A leak at seven that can wait until Thursday, from somebody who would otherwise reach voicemail.
• The repeated questions. The same handful, asked by every new enquiry.
• The booking. A slot that is already free, put in the diary while the caller is still on the line.
• The write-up. What was said goes into the record, so nobody retypes it in the morning.
• The chase. A quote that has gone quiet, followed up without anyone remembering to.

It fails on exceptions, and exceptions are where your reputation lives. An angry caller, a job that fits no category, a price that depends on something only you can see. Each of those needs a person.

So the test before it speaks to a customer is not a script run-through. It is a handful of awkward calls made by somebody on your team, deliberately trying to break it. An unclear address. A job with no category. A caller who asks what it will cost. You are watching for the moment it should stop, and whether it does. A system that guesses politely is worse than one that says somebody will call back.

The handover matters as much as the trigger. If it hands over into an inbox nobody opens until Monday, the exception it correctly spotted still becomes a lost customer. The hand-off needs a destination with a person attached, and a time attached to that person.

The thing to test is not how well it answers. It is how quickly it admits it should not.

So the buying question is narrower than the marketing suggests. Not what can it handle, but where does it hand over, and does the handover reach somebody. Built that way, it closes the gap after hours without putting your name in the mouth of something that cannot read a room. We scope that line before anything goes live, as part of the work that runs on its own.

Set one up for the first minute of the conversation and it books the calls nobody was there to take. Set it up to run the whole conversation and one day it tells a customer something you would never have said.