Can callers tell it's AI?

Confident percentages get quoted around this question, usually about how few callers can tell. Chase any of them back and the trail ends at another vendor's blog post, so we do not repeat them here.

What the transcripts show is more useful anyway. Simple, transactional calls run to completion. Calls that go sideways get discovered.

The tell is usually recovery. Ask a person something odd and they laugh, or say they are not sure. An assistant handles the odd question either very smoothly or not at all.

Designing a phone line around not being caught means designing it to fail badly at the exact moment it matters, because the caller who probes hardest is usually the caller with the biggest problem.

It is also the behavior Maine's chatbot law addresses. Since October 2025, misleading a reasonable consumer into thinking they are speaking to a human, without clear notice, is actionable in trade or commerce there.

Disclosure costs you almost nothing. Callers who are told plainly that they are speaking to an assistant tend to adjust: shorter sentences, clearer requests, fewer tangents. The call gets easier.

The businesses that get the most out of this are the ones who treat the assistant as a good front desk rather than a disguise. It takes the name, the number and the reason, books what it can book, and passes on what it cannot.

Picture a heating and cooling shop on a Tuesday. At 8:12 a homeowner calls to book a fall furnace tune-up. She gives her name, her address and a preferred day, accepts Thursday at 10, and hangs up after about ninety seconds. Nothing in that call gave her a reason to wonder who she was talking to, so she didn't.

At 8:40 a second caller rings about a compressor that was replaced in March and has failed again. He wants to know whether the labor is covered, he's annoyed, and his second sentence runs for half a minute. Around the 25-second mark he stops and asks, "Am I talking to a real person?" The right reply is a plain yes-it's-an-assistant, followed by an offer to put him through to the service manager with the details already taken.

Same voice, same script, same morning. The difference was the call, and the second caller is the one whose opinion of your business was actually on the line.

You don't have to guess. After the first week, search your transcripts for a handful of phrases: "real person", "robot", "is this a recording", "representative", "human". Count the calls that contain one and divide by the week's total. Then read the line just before each hit, because that's the moment the conversation stopped feeling natural to that caller.

Often the trigger is a question the assistant had no answer for, which is a briefing gap you can close with a paragraph. Sometimes it's a caller who simply wanted a person from the start, and a faster route to a transfer is the better fix.

With the disclosure setting on, the question mostly disappears, since the caller was told in the opening line. That's the required approach in Maine and for licensed professions in Utah, and it's the sensible default everywhere else. What you're measuring then is a more useful thing: how many callers, knowing it's an assistant, still choose to finish the call with it.

GreetKeeper's greeting is yours to write, and the disclosure setting puts the notice in the opening line rather than burying it.

Every call arrives as a transcript, so you can read the ones where a caller asked whether they were talking to a person. That is the fastest way to find out what your own callers do.

When a caller asks to speak to someone, the assistant stops and transfers instead of talking around it.