AI telephone receptionists: real-world problems and why the web works better

There are two AI receptionists. One picks up the phone and answers in a voice. The other types in a website’s chat box. This article focuses on the first: what problems arise when the enquiry is a phone call, and why those problems are not the same as those encountered in chat. Each figure is accompanied by the study on which it is based and a link to view it. No figures have been made up, and no rival companies are named.

People want to be able to reach out to someone

In February and March 2026, Gartner surveyed 3,566 business and consumer customers on the use of generative AI in customer service. 87 per cent said that, if a company uses such AI, it is essential to be able to speak to a person. At the same time, 50 per cent said that dealing with the company is easier when it uses generative AI. The takeaway is not that ‘AI is unnecessary’. Rather, people welcome the assistance but reject it when it acts as a barrier.

Gartner asked those who did not want to use AI what would make them change their minds. The most common response was the ability to be put through to a human if necessary. The same Gartner report adds that it is not advisable to make AI a mandatory step in every instance: when a customer has to go through several failed attempts before reaching a human, they are less likely to use that tool again. 58 per cent of those who use generative AI have already used it to complete a task, not just to read a response. In businesses, that figure rises to 74 per cent.

Gartner, August 2026: survey of 3,566 customers (February and March 2026)

Fear is nothing new: reaching out to someone

A previous Gartner survey, conducted in December 2023 amongst 5,728 customers, had already pointed in the same direction. 64 per cent would prefer companies not to use artificial intelligence in customer service. 53 per cent would consider switching providers if they knew the company was going to use AI for customer service. The main concern was that it would be more difficult to speak to a human. This was followed by concerns about the impact on employment and the risk of AI giving the wrong answer.

Gartner, 9 July 2024: survey of 5,728 customers (December 2023)

The person who just rang is the one who wants a voice AI the least

In Gartner’s report ‘Four Data Findings to Inform Your Customer Experience Strategy’, half of customers said they would use a generative AI digital assistant. The figure varies depending on the channel. Among those who had used the telephone in their most recent interaction, only 35 per cent were willing to use such an assistant. Gartner explains this with a practical observation: if the call had already resolved their problem by speaking to a person, they have no reason to risk wasting time on an automated voice system next time. The same report recommends not hiding the human element: make it clear that the assistant can resolve the issue and, if not, pass it on.

Gartner, Four Data Findings to Inform Your Customer Experience Strategy (PDF)

The voice is misunderstood, and that’s annoying

Ofcom carried out a study of smart speakers in the UK through qualitative fieldwork between August and October 2022: one hundred users, plus interviews with those who did not own one. This is not a survey representing the national population. Even so, the finding is clear in its report: the device failing to understand or hear what was being asked of it was, by far, the most common complaint. People had to repeat themselves or change the way they spoke, because the device had started doing something else or had ignored the command.

In the same study, some of the participants said that a regional or international accent – their own or that of someone at home – caused the device to malfunction more often. Some family members had stopped using it because they couldn’t get it to understand them. It also struggled with children, with soft voices, with people speaking quickly, from a distance, or using a conversational tone rather than a short command. Ofcom’s public summary puts it in a single sentence: most people were frustrated because the speaker did not respond well to commands, and this was more noticeable among those with a strong regional accent.

Ofcom, Smart speakers research (PDF, fieldwork August–October 2022)

Ofcom, summary: Smart speaker users won’t be lonely this Christmas

In 2019, BBC Research & Development published the results of an online survey of 330 people, recruited via a blog post. It is not a random sample from a single country: there were more responses from people aged between 25 and 54, and a higher proportion of BBC staff based in London and the north-west of England. Given this limitation, a quarter said the assistant understood them little or not at all. Thirty-eight per cent attributed the failures to their accent and 36 per cent to background noise in the room. Fifty-five per cent said the technology was not yet good enough, and 58 per cent said the problem stemmed from not knowing which command to say. When the assistant fails to understand, the report notes frustration, anger and increased stress. The complaint was not just about the failure itself, but also about the error messages, which were repetitive and unhelpful.

BBC Research & Development, Smart Speaker UX and User Emotion (PDF, 2019)

An AI telephone receptionist operates on that channel. It has to listen to a sentence – often with background noise – and decide when the person has finished speaking. A web chat does not transcribe: it reads what the visitor has typed. Accents and background noise do not alter the text. That does not make the chat infallible. If the text is incorrect, the response is incorrect. It removes the inherent error associated with human hearing.

The reply comes too late for a spoken conversation

In a conversation between people, it is usually one’s turn quite quickly. An article from SIGDIAL 2022, ‘When can I Speak?’, summarises the current state of spoken dialogue systems: they begin their turn after a long silence, lasting between 700 and 1,000 milliseconds, which is why the conversation feels slow and unnatural. People, the article notes, usually respond within about 200 milliseconds. Waiting for such a long silence is the usual way of not cutting off someone who is still thinking. It is also a way of leaving a gap which, in a phone call, is noticeable.

SIGDIAL 2022: When can I speak? Predicting initiation points for spoken dialogue agents

A study presented at SemDial measured real-life interactions with voice assistants, rather than a single-command laboratory setting. It analysed 1,077 turn-taking exchanges involving ten people using the assistant for the first time, across six households. The modelled average time from when the person finished speaking until the assistant began speaking was 1,366 milliseconds, with a standard error of 30 milliseconds. The authors compare this with turn-taking between people, which is around 300 milliseconds, citing Levinson and Torreira (2015), Stivers et al. (2009) and Heldner and Edlund (2010). The sample is small and these are initial findings: it is not an average across all countries or all products. It does, however, show that in real conversations, the pause is several times longer than in human conversation, and without the filler words a person uses to signal that they are about to reply.

SemDial: Alexa, you’re too slow! (Barthel), shift changes with voice assistants

In a chat, that delay doesn’t cut anyone off mid-sentence. The visitor types, sends their message and can continue browsing the page. The reply is displayed in writing. There is no voice that comes in late and cuts off the next sentence – which is another effect that the same SemDial study describes in the chatbot’s turns. Hostly does not publish a guaranteed response time for the model: the chat goes through a queue and the time depends on the workload. What it doesn’t have is the silence of a phone call whilst a system decides you’ve finished speaking.

Loops, a synthetic voice and the question of whether there’s anyone out there

Gartner’s 2026 commentary does not give a percentage of calls stuck in a loop. It describes the mechanism: several failed attempts with the AI before being able to speak to a human make it less likely that the customer will use the tool again. During a call, that loop can be heard. You have to repeat yourself, wait for the next question from the menu and ask to speak to a person again. In a chat, the conversation is recorded in writing. The team opens it in the dashboard and can reply directly within that same thread.

The voice that answers an automated call is a computer-generated voice. This article does not quote a percentage of people who find it ‘robotic’: none of the sources listed above includes a survey that measures this perception using a clear sample. What has been measured, however, is the frustration when the system fails to understand – as reported by the BBC – and Ofcom’s concern when users have to repeat their instructions. A voice that does not allow for nuance and faulty recognition both occur within the same minute. On the website, the response is text. You can read it again. You don’t have to guess at a word that sounded wrong.

Salesforce commissioned YouGov to carry out a survey of 15,015 consumers between 16 July and 16 August 2024 across 18 countries, including Spain. Almost three in four wanted to know whether they were speaking to an AI agent. Forty-five per cent said they would use an AI agent more often if there were a clear way to be put through to a human. Only 30 per cent of consumers would work with an AI agent rather than a human in exchange for a faster service. The Salesforce PDF report, based on the same research, puts it this way: 72 per cent believe it is important to know when they are speaking to an AI agent. Transparency and the option to switch to a human agent are not just nice-to-haves. They are what this sample is asking for.

Salesforce, State of the AI Connected Customer, YouGov survey, July–August 2024

Salesforce, State of the Connected Customer (PDF from the same research series)

Why a web-based AI receptionist does not inherit those flaws

Hostly is the website’s AI receptionist, not a telephone-based service. Visitors type into the chat on your website. There’s no voice recognition, no accents, no background noise, and no synthetic voice to respond. The response uses information that the business has approved. If the information isn’t available, it says it doesn’t know. It doesn’t make up a price, a time or an available slot. If bookings are enabled, it only offers times that the calendar shows as available.

Reaching a person on Hostly does not involve an automatic call transfer. The team sees the thread in the dashboard and takes control. Whilst a person is typing, the AI does not respond in that thread. Requesting a person does not automatically transfer the conversation: it is the person on the dashboard who does so. This is a genuine limitation of the product, and it’s better to state this than to promise a ‘transfer me to someone’ feature that the chat system cannot execute on its own. What it does prevent, however, is leaving someone talking to themselves against a menu: the text remains, and a member of staff can pick up the conversation from exactly where it left off.

AI handles enquiries via your website’s chat feature. If you enable the WhatsApp button, it opens the business’s WhatsApp account and a person will reply: the AI does not get involved in that conversation. Hostly does not read those WhatsApp messages or use a model to process them. Anyone who wants to message the business, rather than using the website’s chat, can use that button. Visitors to the website can continue using the chat.

The entry-level price is €29.90 per month, including VAT. This page does not state that a chat feature reduces cancellations, nor does it state that payments processed via Stripe incur a specific commission: this information is not included in the sources cited above and is not a product promise made in this text. Payment, if the business enables this feature, is either a deposit or the full amount at the time of booking. The card payment does not go through Hostly.

What this article does not prove

It does not prove that every AI telephone receptionist fails in the same way. The studies measure voice assistants, smart speakers and public opinion on AI in customer service, not a comprehensive survey of products. SemDial’s sample consists of ten people. The BBC’s sample is not random. Ofcom’s study on smart speakers is qualitative. The Gartner and Salesforce surveys are large-scale surveys of customers and consumers, not a listening test of a specific number.

Nor does it prove that a chatbot is never wrong. If the profile is out of date, so is the response. If the business switches off the AI, nobody will reply until a human steps in. If the visitor needs a phone call, the business’s phone number is still the phone number: Hostly does not replace it. What the chatbot prevents are the errors that only arise because the conversation is spoken.

The product page, which includes installation details, languages, bookings and price, is the AI receptionist.

Sources

The product details page is the AI receptionist.