Do AI receptionists sound like real people?

Mostly, and it depends far more on the call than on the voice. On a booking, a question about your hours, a quote request or a message taken at two in the morning, an AI receptionist is hard to tell apart from a competent person on a good line. On a rambling complaint, an upset caller, or a question nobody ever wrote down, it still gives itself away inside a sentence or two. BotsDontSleep answers on the first ring in under a second, 24/7, at $0.60 CAD per answered minute, and tells callers it is an AI by default. What follows is a call-type by call-type account of where that holds up and where it does not.

What do callers actually notice in the first ten seconds?

Timing, before timbre. The thing a caller registers first is not how the voice sounds but whether the reply arrives when a reply is supposed to arrive.

A study of ordinary conversation across ten languages, from small indigenous communities to major world languages, found the same pattern everywhere: people avoid talking over each other and they minimize the silence between turns, and the average gap between turns varies across languages by only about 250 milliseconds (Stivers et al., Proceedings of the National Academy of Sciences, 2009). That is the machinery your caller is running when they phone your shop. They are not consciously grading a voice. They are checking whether the turn came back on time.

Which is why the tells that matter are timing ones. A pause of a second and a half before the answer starts. A reply that begins before the caller has finished. A voice that keeps talking through an interruption instead of stopping. Any of those reads as machine even when the diction is perfect, and none of them are fixed by a better voice model.

Three things get clocked, roughly in order: something picked up rather than ringing out, the reply came back at conversational speed, and the reply was about what was actually said. Miss the third and it does not matter how human the first two were.

Which calls does an AI receptionist sound completely natural on?

The bounded ones — where the caller arrives knowing what they want and the vocabulary is small and known in advance.

Bookings and reschedules. Someone wants Thursday afternoon, you have Thursday at 2:15, they take it. The whole exchange is four turns and every branch of it is predictable. This is the call type where a synthetic voice is genuinely indistinguishable, and it is also the highest-value call most businesses miss — a dental front desk in Toronto mid-procedure is not answering it, and voicemail does not book anything.

Hours, address, parking, and whether you do that. Pure lookup. A person reading the same answer off a screen sounds no more human than software reading it off a screen.

Quotes for standard work. Anything with a rate card behind it. The caller wants a number and a next step.

Status checks. Is my car ready, has the part come in, did you get my email. Short, factual, and constant in trades and clinics.

Wrong numbers and robocalls. Nobody is assessing realism, and the call is over in twenty seconds. At $0.60 a minute that is about twenty cents, which is roughly what it is worth.

After-hours messages. At 11 p.m. the bar is not "did this sound like a person," it is "did anyone take this down and will someone call me back." An after-hours answering service that gets the name, the number and the problem right is doing the job the caller wanted.

A salon with every chair full is the clean example: almost all of that inbound is booking, rescheduling and price questions, which is almost all of the natural-sounding column.

Which calls still sound like software?

Four kinds, and they are worth knowing before your customers find them for you.

The open-ended call. A caller who has not decided what they want, thinking aloud, changing direction twice. A person follows a conversation with no destination. An AI keeps trying to work out what is being asked of it, and the seams show.

The emotional call. Anger, distress, grief. A human hears the state and drops the script without being told to. An AI receptionist handles a distressed caller politely and competently, then hands off — which is the correct behaviour, and the first twenty seconds are still worse than a good person's.

The call with a lot of noise in it. A caller beside a running engine, on a failing mobile connection, with a second conversation happening in the room. Recognition degrades exactly where you would expect. The right response is to ask once more and then take a message with the number, not to guess.

Words nobody taught it. Your street names, a family surname, an abbreviation your trade uses. A voice model pronounces unfamiliar proper nouns by rule, and Canadian place names defeat rules constantly. That one is fixable in minutes once somebody reads the transcripts and notices, which is a different problem from the first three. There is more on that in whether your callers will understand the voice on your line.

The failure mode we design for is boring on purpose: when the call leaves what it knows, it says so and passes the caller to a person with a summary. It never invents an answer. An AI that improvises to stay in character is worse than one that admits the edge of what it knows, and the caller can tell the difference between the two immediately.

Do callers ask whether they are talking to a person?

Some do, and they get a straight answer, because disclosure is on by default.

Our agents identify as an AI at the start of the call unless a client turns that off, and it stays a client decision rather than ours. The reaction is less dramatic than owners expect. A caller who wants an appointment at 4:30 mostly wants an appointment at 4:30 — being told what they are speaking to at the start costs a second and removes the thing they would otherwise be quietly wondering about for the rest of the call.

Some callers do not want to deal with software, and they will say so. That is what transfer rules are for: any caller who asks for a person should reach one on your terms, and you decide whether that means a ring-through during business hours or a callback promise after them. The mechanics of the voice itself — pacing, filler words, breathing, the four places it gives itself away — are covered in can an AI receptionist actually sound human.

How does each way of answering your phone sound to the person calling?

The row that decides most of these purchases is not "sounds like a person." It is the transcript row. A recording of what was actually said settles a complaint in thirty seconds, and it is how a mispronunciation or a bad handoff gets found in week one rather than repeating quietly for a year.

Can you change the way it sounds?

Yes, and most of the levers are wording rather than voice.

You choose the voice and the pace. You write the greeting — the difference between "Thanks for calling, how can I help" and naming the business and the person is the difference between a call centre and your front desk. You set how much it says before its first question, which drives whether a call feels brisk or scripted. You give it the pronunciation list for your streets, your staff and your services. And you set what it never says.

Tuning happens during the build, which takes about a week, and changes after that are a message rather than a project. There is no per-change fee and no configuration charge. The reason to keep adjusting is not vanity about realism — every transcript from the first fortnight shows a real caller reacting to a real line of your script, which beats any amount of pre-launch guessing. The same applies to a line answering across Canada, where the pronunciation list is different in every province.

Is sounding human the thing you are actually buying?

No, and we would rather say so plainly than sell you on realism.

The variable that moves money is whether the phone gets answered at all. A call answered in under a second by a voice the caller can tell is software beats a call that rings out — the caller who reaches voicemail is gone to the next listing, and no amount of human warmth on a line nobody picks up recovers them.

That cuts the other way too. Three cases where a person is the better purchase, and we will tell you so on a sales call:

When the conversation is the product. High-touch sales, negotiation, consultations where the caller has to be persuaded rather than served. If the phone call is the pitch, put a person on it.

When your callers are mostly unfamiliar with synthetic voices. Familiarity does most of the work in how natural a voice feels. If a large share of your callers are over eighty, the gap is real today, even though it narrows every year.

When you take fewer than about twenty calls a month and you answer nearly all of them. The phone in your pocket is already doing the job and there is nothing here worth paying for.

If you want a human voice on every single call and you are prepared to pay per minute for it, a staffed service is the right answer and our comparison with Posh sets out where a human-answered service genuinely wins. Everything on that page is still true when we are not the one being chosen.

What does it cost to have your line answered this way?

$0.60 CAD per answered minute, billed by the second, with $0 setup and no contract.

Voice selection, script tuning, the pronunciation list, and every later change to how something is said are included. Three hundred answered minutes a month is $180. A two-minute booking call is about $1.20. That twenty-second wrong number is about twenty cents.

For a comparison point on the human side: a February 2026 pricing guide from Polaris Voice put Canadian live answering at roughly $0.75 to $1.50 a minute for message-taking and $1.25 to $2.00 a minute for a virtual receptionist tier. Rates move, and any provider's own site carries the current figure — but that gap is why this question gets asked at all. Nobody would wonder how close software gets if it were not substantially cheaper.

The honest way to settle it is to stop reading and call one. Maryann, our demo agent, answers at 437-494-9110 at any hour. She is trained on a business that is not yours, which is exactly what makes the test fair — ask her something a real caller would ask, then interrupt her halfway through the answer and see what she does.