Missed calls
Does an AI receptionist sound robotic?
Does an AI receptionist sound robotic? Listeners took AI voices for human in 41% to 58% of trials. What gives AI calls away, and 5 tests to run before you buy.
A well-built AI receptionist does not sound like a robot. In a 2025 study, listeners took AI voices for human in 41% to 58% of trials, against 62% for real voices (PLOS One). The higher figure is for voice clones of real people.
The voice is rarely what gives an AI call away. Slow replies, a phone-tree script, and an agent that pretends to be a person do. For the wider picture of what these agents handle, start with AI phone answering for service businesses.
The short version
- Modern AI voices often pass as human. Listeners labelled 41% of generic AI voices and 58% of voice clones as human, against 62% for real voices (PLOS One, September 2025).
- Timing sounds robotic before tone does. In conversation, people most often reply within 0 to 200 milliseconds (Stivers et al., PNAS 2009).
- Callers still prefer people: 84.9% of consumers in the US and Canada would rather reach a human agent (Metrigy, November 2025).
- Say it is an AI. 71% of consumers in a Five9 survey call it very or extremely important to know when they are talking to one (Five9, June 2026).
- Test before you buy: interrupt it, ask if it is a robot, and ask it something off script.
Can callers tell an AI voice from a real person?
Often, they cannot. Researchers at Queen Mary University of London and UCL played listeners real and AI voices. Then they asked which ones were human (PLOS One, published 24 September 2025).
The clones, built from recordings of real people, were labelled human almost as often as the real voices. Generic AI voices, designed from scratch, were caught more often, yet still passed as human in 41% of trials. If your receptionist uses a stock voice rather than a clone of yours, 41% is the safer figure to plan on.
The study has 2 limits. Experiment 1b used 50 listeners and short recorded clips, not a live call. A real call gives the caller more chances to notice something off, through timing and content.
What makes an AI phone call sound robotic?
How the call moves gives an AI away well before its voice does. Callers notice how fast the reply comes, whether they can interrupt, and whether the answer fits the question.
Timing is the big one. Across 10 languages, the most common gap between turns is 0 to 200 milliseconds, with a median of 100 (Stivers et al., PNAS 2009). A reply that lands a full second late feels like being put on hold.
The platforms aim slower than people do. Retell's documentation says people expect a response within 300 to 500 milliseconds (Retell AI, retrieved 2026-10-07), and Vapi advertises sub-600ms response times (Vapi, retrieved 2026-10-07).
Both are vendor figures, and both are slower than a human reply. Judge the real thing on a real phone.
| What sounds robotic | What sounds natural |
|---|---|
| A pause of a second or more before each answer | A reply well under a second |
| Talks over the caller, or ignores an interruption | Stops when the caller cuts in, then answers the new question |
| "Press 1 for service, press 2 for billing" | "What can I help you with?" and a real answer |
| Repeats the same sentence word for word | Rephrases, and confirms details in its own words |
| Mangles your street names and trade words | Says your towns, products, and jobs the way you do |
| Claims to be a person when asked | Says it is an AI assistant, then keeps helping |
What our demo agent does on a call
We run a demo agent called Jessica on our AI receptionist page, and it is where we send every owner who asks this question. You press a button, talk to her in the browser, and play the customer.
Try 3 things when you test her:
- Interrupt her. She stops, takes the new question, and answers it, instead of finishing her sentence.
- Ask if she is a robot. She says she is an AI assistant. She does not dodge the question or invent a desk and a coffee mug.
- Ask something she should not answer. Try a complaint, or a job that needs a site visit. She takes the details for a callback instead of guessing.
The second test matters most. An agent that sounds human but lies about it does more damage than one that sounds slightly synthetic and tells the truth. A caller who feels tricked does not book.
Do callers mind talking to an AI receptionist?
Most callers would still rather talk to a person, and the numbers are not close. These surveys asked about customer service in general, not phone receptionists alone.
84.9% of consumers prefer a human agent over an AI agent. 80.1% still do when told either one would solve their issue (Metrigy). Metrigy surveyed 503 adults in the United States and Canada in November 2025.
Willingness is a different number. Five9, which sells contact centre software, found 80% of consumers willing to use AI-powered customer service.
Two-thirds still prefer a human (Five9, June 2026). That survey covered 3,000 consumers in the US, the UK, and Germany.
Put together, the caller's position is clear. They would like a person, they will accept an AI, and they want to know which one they got.
At 8 pm, while you are on a roof, the real choice is rarely AI or a person. It is AI, a paid answering service, or voicemail, as our comparison of AI receptionists, answering services, and voicemail sets out.
Should an AI receptionist say it is an AI?
Yes. Say it in the greeting, and again whenever a caller asks. In the same Five9 survey, 71% rated knowing they face an AI agent as very or extremely important (Five9, June 2026).
The rules point the same way. In the United States, the FCC has confirmed that robocall limits on an "artificial or prerecorded voice" cover human-sounding AI voices (Federal Register, published 10 September 2024). Those rules cover calls you place, not calls you answer, but they show where regulators are heading.
At least one state already covers answered calls. Utah's SB 226, in effect since 7 May 2025, requires disclosure of generative AI when a consumer asks (Alston & Bird, June 2025). Rules differ by state and province, and this is a summary, not legal advice.
Check the rules where you operate before you go live. A disclosure line in the greeting costs nothing and meets the asks-based rule in Utah.
How to test an AI receptionist before it answers your line
Make 5 calls in 10 minutes from your own cell phone before you sign anything. You are testing the call, not the demo video.
- Call from a noisy place. A truck, a job site, or a kitchen. It should still catch your address and the job.
- Interrupt it mid-sentence. It should stop and answer what you just asked.
- Ask, "Am I talking to a robot?" It should say it is an AI, without dodging, and keep helping.
- Ask your 3 most common questions. Price, service area, and the next open slot. The answers should be yours, not generic.
- Ask for something it cannot do. A complaint or an odd job. It should take the details and send them to you.
If it fails 2 of these, your callers will notice too. Few small businesses use AI to serve customers yet, so a good agent still stands out. In Canada, 12.2% of businesses used AI to produce goods or deliver services in the year to spring 2025 (Statistics Canada).
In the US, firms with fewer than 20 employees used AI at rates under 20% (US Census Bureau, May 2026). The flip side: a bad agent may be the first AI phone call your customer remembers.
Where to hear one for yourself
Samro Studios builds quoting agents for service businesses. By phone, the agent answers when you cannot pick up, quotes from your rates, and books the job. You get the details by text and email, and it works in English and French.
Talk to Jessica on the AI receptionist page and judge the voice with the 5 tests above. Builds for other businesses are on the portfolio. When you want one built around your trade, tell us about it on the contact form.
Frequently asked questions
Does an AI receptionist sound like a real person?
A good one often does. In a 2025 PLOS One study, listeners took generic AI voices for human in 41% of trials and clones in 58%, against 62% for real voices. On a live call, slow replies and stiff scripts are often what give an AI away.
Why do some AI receptionists still sound robotic?
Mostly timing and scripts. People most often answer each other within 0 to 200 milliseconds, so a 1-second delay feels mechanical. Weaker AI receptionists run fixed scripts and ignore interruptions, while a good one replies quickly, stops when cut off, and answers the actual question.
Will customers hang up when they hear an AI?
Most say they are willing to use one, if it helps them. In a Metrigy survey, 84.9% of US and Canadian consumers would rather have a person (Metrigy, November 2025), but Five9 found 80% willing to use AI customer service. A caller with a burst pipe wants a booked time more than a human voice.
Do I have to tell callers they are talking to an AI?
It depends where you operate, and the rules are changing. Utah's SB 226 requires disclosure when a consumer asks whether AI is being used. Even where no rule applies, say it. In a Five9 survey, 71% of consumers said knowing they are talking to an AI is very or extremely important.
Can an AI receptionist handle noisy job sites?
Some handle noise well and some do not, so test it yourself. Call from a truck, a job site, or a busy street, and ask a colleague with a different accent to call too. If it mishears an address twice in 5 test calls, it will mishear your customers.
Sami Bdeira is the founder of Samro Studios, which builds quoting agents and custom tools for service businesses in the United States and Canada.
