Does an AI receptionist sound human? Can callers tell?
On voice quality alone, most callers can no longer tell an AI receptionist from a person. What gives it away on a live call is timing, interruptions, and questions it was never set up to answer. Here is what the research says and how to handle the objection honestly.
Mostly yes, and on voice quality alone most callers can no longer tell. Recent lab studies found listeners labeled cloned AI voices as human about as often as real ones. What gives an AI receptionist away on a live call is rarely the voice. It is timing, how it handles being interrupted, and what it does when someone asks a question it was never set up to answer.
That distinction matters when a prospect asks you "does it sound like a robot?" because the honest answer has two parts, and the second part is the one that closes deals.
What does the research say about AI voices passing as human?
Two peer-reviewed studies from the last year are worth knowing by name, because your more skeptical prospects will have seen a headline about one of them.
A team at Queen Mary University of London and UCL played listeners 40 real human voices, 40 AI clones of real people, and 40 AI voices generated from scratch. The clones were built in ElevenLabs from about four minutes of recordings per speaker. In the PLOS One paper, published in September 2025, listeners called the clones human on 58% of trials, against 62% for the real voices. That gap was not statistically significant. The from-scratch voices did worse, labeled human 41% of the time against 81% for real people.
The finding a local business owner will care about more sits in the same paper. The AI voices were rated more dominant than the human ones, and the generic AI voices were rated more trustworthy than the humans they were matched against. Nadine Lavan, who co-led the study, said in the university's release: making the clones "required minimal expertise, only a few minutes of voice recordings, and almost no money."
A separate UC Berkeley study in Scientific Reports, with 604 participants, found people correctly identified an AI voice as AI only about 60% of the time. Guessing gets you 50%.
So when a prospect asks whether the voice sounds human, the answer is yes. That question is settled.
Why can some callers still tell it's an AI?
Both studies share one limit. Listeners heard short clips of read speech. Nobody was on a live call, asking a follow-up, cutting in halfway through a sentence, or hearing the agent pause while it looked up a calendar slot.
A live call tests things a clip never does. Here is where the tells show up in practice.
Timing. People are very good at turn-taking. A 2009 study in PNAS looked at conversations in 10 languages, from major world languages to small indigenous communities, and found the same pattern everywhere: speakers avoid talking over each other and keep the silence between turns short. Average gaps across all ten languages sat within 250 milliseconds of each other. Callers feel a slow answer before they can name it. An agent can have a perfect voice and still sound off because it waits a beat too long.
Interruptions. A caller who talks over a half-duplex agent either gets ignored or produces a stutter as the agent stops and restarts. This is changing fast. OpenAI's GPT-Live-1 is described on its model page as a full-duplex voice model that "can listen and speak at the same time." We wrote up what full-duplex voice changes for agencies when it shipped. The short version is that interruption handling is now a model feature, not something you hack around.
The edges of what it knows. This is the big one, and it has nothing to do with audio. A receptionist that knows the practice's hours, services and price bands sounds human right up until someone asks about parking, or whether Dr. Patel is back from vacation, or if the office takes a specific insurance plan nobody put in the prompt. A human says "let me check." A poorly configured agent invents an answer or loops back to its script. Callers catch that in one exchange.
Politeness that never runs out. A real front desk gets a little short at 4:55 on a Friday. An agent that is equally cheerful to every caller on every turn starts to feel like a recording after a minute or two. Plenty of callers will not mind. A few will notice.
Does it matter if callers can tell?
Less than prospects think, and this is where you take the objection apart.
The owner asking "will people know?" is usually afraid of one of three things: customers feeling tricked, customers hanging up, or the business looking cheap. None of those depends on whether the voice passes a blind test.
Tricked is a disclosure problem, not a voice problem. Some states and some trades have rules about telling callers they are talking to an AI, and we covered what those disclosure rules look like separately. An agent that says "I'm the practice's virtual assistant" in its greeting takes that fear off the table. What the caller cares about after that is whether the question got answered and the appointment got booked.
Hanging up is a competence problem. Callers hang up on an agent that cannot help them. They hang up just as fast on voicemail, which is where a lot of after-hours calls to a small business end up. The comparison your prospect should make is not AI versus their best receptionist on a good morning. It is AI versus the phone ringing out at 7pm.
Looking cheap is a configuration problem. A receptionist that speaks in the trade's language, knows the real services and quotes inside the real price bands sounds like the business. That is why our vertical packs carry services, price bands and objections for each trade instead of a generic script.
How do you let a prospect judge it for themselves?
Stop describing the voice and put them on a call with it. Every argument above is weaker than ninety seconds of the prospect talking to an agent that knows their business.
On our live demo site, the prospect starts a call in the browser and talks to a receptionist set up for their trade. They ask the question they were about to ask you. Some will try to break it. Let them. A prospect who tried to trip it up and could not has answered their own objection, and that beats anything you could say.
A few things that help on the call itself:
- Tell them to interrupt it. Prospects expect a robot to steamroll them, and watching it stop and listen does more than any claim about latency.
- Ask them for the weirdest question their front desk got last week. Then let the agent handle it. If it says it will take a message rather than guessing, point that out. That is the behavior you want in production.
- Do not narrate over the call. Mute yourself and let the silence belong to the agent.
We walked through the full structure of that conversation in how to run the demo call.
Where this falls short
Be straight with prospects about the limits, because they will find them.
Every published demo carries a sandbox disclosure. The page states that the AI is a demo and that bookings made in it are simulated. That is deliberate and you cannot remove it. So the demo proves how the agent sounds and handles a conversation. It does not prove that bookings land in their calendar, and you should not imply it does.
Single demo calls cap at five minutes, and the $49 plan includes 100 live voice minutes a month. That is enough for a real test, not for a prospect to leave it running as their phone line.
Some callers will always know, and some will always ask for a person. Older customers calling a practice they have used for twenty years are the obvious group. An agent that hands those calls to a human, or takes a clean message, is fine. An agent built to hide that it is software and keep them on the line is a liability. If a prospect's pitch to you is "I want it to be undetectable," that is a sign to slow down, not a feature request.
And the voice will not rescue a bad setup. If the business cannot tell you its own hours, services and price ranges, the agent will not know them either, and callers will hear that immediately.
FAQ
Can callers tell they are talking to an AI receptionist? On voice quality alone, usually not. Lab studies found listeners judged AI clones human about as often as real voices. On a live call, slow responses, poor interruption handling and made-up answers to off-script questions are what give it away.
Should the AI say it is an AI? Where the law requires it, yes. Where it does not, it still costs you very little. A one-line greeting that names it as the business's virtual assistant sets expectations and removes the "tricked" objection. Check the rules for your client's state and trade.
Will customers hang up on an AI receptionist? Some will, especially callers who wanted a specific person. The fair comparison is voicemail or an unanswered line, not a human receptionist on a quiet morning. An agent that answers the question and books the slot keeps most callers on the line.
Does the demo use a real phone number? No. The prospect talks to the agent in the browser, with no phone number to buy and no API key to set up. Bookings and lookups in the demo are simulated, and the page says so.
What makes an AI receptionist sound more human? Mostly content, not audio. Real services, real price bands, the trade's own vocabulary, and a clean way to say "I'll have someone call you back" when it does not know. The voice itself is already good enough.
Next step
Pick the trade you sell into most, open the live demo, and call it yourself. Interrupt it, ask it something off-script, and time how long the gaps feel. If it passes your own test, it will pass your prospect's. When you are ready to send one under your own brand, start your workspace and read the FAQ first so you know the limits before a prospect asks.