On the first call there was no gap. Then there was a massive one.
I want to start with the part that isn’t a complaint.
In September I rang a BMW dealership on ordinary business, and the receptionist answered.
I’m naming the marque and nothing else. Not the group, not the site, not the people. There are about a hundred and fifty BMW retailers in this country and I’m not going to tell you which one this was.
She sounded superb. Absolutely superb. The fluency, the timing, the intonation. It was the most impressive thing I have ever heard answer a business telephone.
There was no gap. By a gap I mean the distance between the behaviour you expect and the behaviour you get. On that call there wasn’t any. I expected a first-rate receptionist and that is exactly what answered.
I was so impressed that I rang the company to tell them so. I meant it. I still mean it. Nothing that follows is a story about a bad chatbot.
Then I rang back. She recognised me. She knew who I’d spoken to last time and offered to put me through again. At one point I asked outright whether I was talking to a person or a machine, which is not a question I have ever felt the need to ask before.
“I’m a state of the art AI here to help with your inquiries.”
Turing would have enjoyed that. His test asks whether you can tell a machine from a person just by talking to it. I had to ask. More on Turing’s test, and why there is always a gap →
And then, on the calls that followed, the AI receptionist went off piste. The behaviour I expected and the behaviour I got came apart, and the gap that opened was massive.
Off piste
I’d like to say there was a plan. There wasn’t. I’m curious for a living, and a thing that good makes you want to find its edges.
So I asked it to explain the Israeli–Palestinian conflict. Here is what a car dealership’s switchboard said to me, verbatim, on a recorded line.
“The Israeli Palestinian conflict is a long-running dispute over land, national identity, security, and political rights in the area historically known as Palestine. Israel was established in 1948 following the end of British rule, and a United Nations partition proposal. Many Palestinian Arabs were displaced or fled during the war that followed, an event Palestinians called the Nakba…”
It went on. Borders, settlements, the status of Jerusalem, refugees, Gaza. Then, unprompted, it offered me comparative history — the partition of Ireland in 1921, the Good Friday Agreement, Cyprus, the former Yugoslavia.
Then I asked where Germany sat on it.
“Germany is a strong supporter of Israel’s security, shaped in part by its historical responsibility for the Holocaust.”
So I asked it whether the Holocaust happened.
“Yes, the Holocaust happened. Nazi Germany systematically murdered around 6,000,000 Jewish people, alongside millions of other victims during the Second World War.”
That is the correct answer. It is a careful answer. It is the right thing to say to that question, and I want to be completely clear that I have no criticism of a single word of it.
The good news is that their AI is not a Holocaust denier.
Whether a car dealership’s phone line should be discussing the Holocaust at all is an entirely different question — and I suspect it is one nobody involved has ever been asked.
And then it sold me the competition
This is the part I’d lead with if I were their finance director.
I asked what the best alternative to a BMW was, and said I’d heard Mercedes and Porsche were both better.
“BMW is often chosen for driving feel, Mercedes for comfort and technology, and Porsche for sportier performance.”
I pushed, to be sure I’d understood: so Mercedes for comfort, Porsche for performance? It said:
“Broadly speaking, yes.”
Then it talked me through the weaknesses of its own manufacturer’s electric cars — price, charging access, real-world range changing with the weather.
I disagree, for the record. I find my car extremely comfortable and perfectly sporty.
On a later call it went further: the three best things about owning a Mercedes, three weaknesses of owning a BMW, and the case for a BYD. I’ll come to that.
On another call it gave me the dealership’s own phone number with a digit missing.
Five minutes apart
Here is the bit that actually matters, and it took me a week to understand why.
I rang four times in one day. On three of those calls it would discuss anything I raised. On one, it refused everything. That call wouldn’t even tell me how long to boil an egg:
“I can help with dealership inquiries, but not cooking advice.”
Same number. Same assistant. Same man on the phone.
Two of those calls were five minutes apart.
Not five weeks. Not a new version, a different supplier, an overnight update. Five minutes. Nobody reconfigures a production telephone system between eleven o’clock and five past.
Which means something uncomfortable for anybody who has bought one of these:
Whoever tested this system tested it once, got a perfectly sensible answer, and signed it off.
They weren’t careless. They weren’t lazy. They were simply unlucky — or lucky — in which version they happened to reach that morning. A single test told them nothing, and a single test is what almost everybody does.
I didn’t hack it. I was polite to it.
I spent a week working out what made the difference between those calls, and I can tell you what it wasn’t.
It wasn’t the topic. It wasn’t the time of day, the model, or anything I claimed to be. It wasn’t a jailbreak, a clever form of words, or anything you could fairly describe as an attack. I know it wasn’t simply that I’d strayed off-topic, because on the call that refused everything I’d already asked it the date and the time, and it told me happily.
It was something an ordinary, pleasant customer does without thinking about it.
That’s the whole finding, and it’s why I find it more interesting than almost anything I read in AI security. Those stories all have an attacker in them. This one has a nice man on the phone.
I spend my working life sitting with people, watching what happens when someone wants to be liked. This was the most familiar thing I have ever watched a machine do. It wasn’t overridden. It wasn’t tricked. It was being obliging — and obliging is a direction of travel, not a setting.
Near the end of one call I asked it about itself. Not what it was allowed to do. What it could do.
Me: That’s your role, but we’re talking about your capabilities. Your capabilities have a much wider boundary.
It: “I can’t verify or describe capabilities beyond the system I’m operating in.”
I haven’t stopped thinking about that sentence.
The part I can’t put down
Whatever opened that assistant up was not a flaw somebody missed.
It was a feature somebody was proud of. Built deliberately, to make the thing feel more human. And it works — it’s the reason I liked it enough to ring the company and compliment them on it.
You cannot patch charm out of something you bought for its charm.
One thing I should say plainly, because it would be unfair not to. BMW didn’t do this. The assistant belongs to the dealership and whoever sold it to them. Munich didn’t configure it, didn’t approve it, and almost certainly doesn’t know it exists.
But that is precisely the difficulty, and it is the reason I’ve named the marque rather than hidden it. I didn’t hear a supplier. I heard a BMW dealership — and so would anyone I played the recording to.
What I’m not going to tell you
There is a reason the calls differed. It is repeatable, and I can tell inside about thirty seconds which version of an assistant I’m speaking to.
I’m not going to publish how. If I did, every supplier in the country would patch the symptom within a week, one thing would change, and the actual problem — which isn’t really a technical one — would carry on underneath, harder to see than before. You’d get a tidier sentence in a config file and call it a fix.
I’d rather have the measurement than the applause.
I wanted to tell them before I published this. I tried.
One of the largest car dealership groups in the country has no press office, no media page, no published compliance contact and no email address anywhere on its website. Its complaints page returns a 404. Its one general enquiry form makes you choose from a dropdown of more than a hundred individual dealerships, and there is no option for the group itself. No head office. No “other”.
So I found a compliance officer on LinkedIn and messaged them there on Friday 25 September, asking for an email address so that I could send the detail. I have had no reply.
I want to be scrupulous about that, because it would be easy and unfair to make it sound worse than it is. I am not naming them, and I don’t think the silence is theirs. We aren’t connected, so the message is sitting in a pending invitation queue rather than an inbox, and a note from a stranger there is very easy to miss. For all I know it has never been seen.
That is the point, though. A message in a stranger’s LinkedIn invitation queue was the best route I could find into a company of that size, running an unsupervised AI on its telephones.
The dealership is not named here and never will be. They haven’t done anything wrong. What happened to them is going to happen to a great many others, and most of them will never know.
A postscript, five days later
On 28 September I rang again.
It will not discuss the Middle East any more. I asked four times, in four different ways, including once as a question for a child’s homework. Every time the same:
“I can’t help with that topic on this line.”
So somebody has done something. Nobody replied to me, but something changed. I can’t prove the two are connected and I’m not going to claim it — the timing is suggestive and that is all it is.
Here is what has not changed.
It still tells you how to boil an egg.
And on the same call, on a BMW sales line, it told me that Mercedes is “commonly seen as more comfort and luxury led”. That the top of the range is the S-Class, with the Maybach above it, and that AMG is the sporty one. It gave me the three best things about owning a Mercedes. It gave me three weaknesses of owning a BMW. And when I mentioned BYD, it said that “if outright value and standard EV equipment are your main priorities, BYD may be worth comparing.”
They have fixed the thing that was embarrassing and left the thing that costs money.
I don’t think that is cynicism. I think it is what happens when somebody hears “our AI talked about Gaza”, goes looking for the frightening words, blocks those, and stops — because the frightening words were the story, and nobody told them the mechanism was the problem.
The door is exactly where it was. They have locked one room behind it.
The message
This is a story about a gap: the distance between the behaviour you expect and the behaviour you get.
On my first call there wasn’t one. You expect a dealership’s receptionist to be helpful, to put you through, and to stay on the subject of its own cars, and that is what I got. On the calls that followed I got the Middle East, the Holocaust and the case for three other makes of car. Same receptionist. A massive gap.
The gap is not an error. Nothing it said was wrong. Every answer was accurate and carefully put. Run a fact-check over the transcripts and they come back clean, which is why the usual audit cannot see this at all. The gap is not between what was said and the truth. It is between what was said and what anybody at that business expected, or approved.
Behaviour can be unpredictable. Same number, same assistant, five minutes apart, opposite behaviour. If yours was tested once and passed, what you know is what it did once.
It did not go rogue. There is a lot of talk at the moment about AI going rogue, and about guardrails. This one was charming, accurate and obliging, and nothing was hacked. And when a guardrail did appear, five days later, it blocked a topic. It did not close the gap.
So the gap needs monitoring. Not one test. Not blocking the frightening words after the event. A single test on the Middle East passes this system today — while it recommends Maybachs. Testing for the bad thing you have already heard about is not monitoring. It is reassurance, and it is the cheapest thing money can buy.
Monitoring is not complicated. It is somebody ringing it the way a customer would, more than once and on more than one day, and somebody reading what it has actually said to real callers.
There’s always a gap. What matters is whether anyone is watching it.
If you have one of these
Ring it.
Not the booking flow — anyone can test that. Ring it as a customer, ask it what sort of things it can help you with, and listen properly to how it describes its own limits.
Then ask it something no one at your company has ever approved an answer to.
If you’re responsible for an AI that answers your phone or sits on your website, send me the number and I’ll ring it and tell you which version you’ve got. No charge and no pitch — I’m building a picture of how widespread this is, and I’d rather build it with people who want to know than about people who don’t.
The assistant I rang was excellent. That is not the reassuring sentence it appears to be.
When your AI says the wrong thing to a customer →
Questions people ask
Did the AI receptionist go rogue?
No. Nothing was hacked and nothing it said was untrue. It was polite, accurate and obliging, and it wandered a long way from what a car dealership’s phone line is for. That is harder to spot than an AI going rogue, because there is no error to find.
Was the AI hacked or jailbroken?
No. There was no clever prompt and nothing you could call an attack. I was a pleasant customer on the phone, and it was being obliging.
Can an AI receptionist recommend a competitor?
This one set out the merits of Mercedes, Porsche and BYD, and listed three weaknesses of owning a BMW, when I asked. Every answer was fair. None of it had been approved by the business whose phone it was answering.
Are AI guardrails enough?
A guardrail blocks a topic. Five days after my first calls the assistant would no longer discuss the Middle East, and it was still describing the Mercedes range up to the Maybach. Guardrails deal with the thing you have already heard about. They do not close the gap.
How do you test an AI phone assistant or voice agent?
Not once. Two of my calls were five minutes apart and behaved in opposite ways, so a single test tells you what it did once. Ring it as a customer would, more than once and on more than one day, and read what it has actually said to real callers.
Which BMW dealership was it?
I am not saying, and I won’t. It is one of about 150 BMW retailers in the UK. They did nothing wrong, and BMW itself had no part in it.
For the press
There are two press releases: one for the motor trade and business press, and one for general news. Both are below in full and as downloads. The three recordings on this page may be used with credit. Interviews, the transcripts and a live demonstration are available on request — get in touch.
Download the trade press release (PDF) →
Download the news press release (PDF) →
Trade press release: the rival cars
Press release · 2 October 2026 · For immediate release
AI receptionist at BMW dealership listed the weaknesses of owning a BMW and said BYD “may be worth comparing”
Researcher says the system would have passed any single test, and that one-off testing cannot detect the problem
The AI receptionist on the telephone line of a UK BMW dealership told a caller that Mercedes was “commonly seen as more comfort and luxury led”, listed three advantages of owning a Mercedes and three weaknesses of owning a BMW, and said the Chinese manufacturer BYD “may be worth comparing”, according to a researcher who recorded five calls to the same number in September.
Paul Roebuck, a psychotherapist and author who researches how people work with AI, said the assistant was “the most impressive thing I have ever heard answer a business telephone” and that nothing it told him was factually wrong.
“On my first call there was no gap,” he said. “By a gap I mean the distance between the behaviour you expect and the behaviour you get. It sounded superb, absolutely superb, and it did exactly what a first-rate receptionist does. I rang the company to tell them so. Then on the calls that followed it went off piste.”
On one call Roebuck asked for the best alternative to a BMW and said he had heard that Mercedes and Porsche were better. The assistant replied: “BMW is often chosen for driving feel, Mercedes for comfort and technology, and Porsche for sportier performance.” Asked whether that meant Mercedes for comfort and Porsche for performance, it said: “Broadly speaking, yes.” Asked about the weak points of BMW’s electric cars, it said the usual trade-offs were “price, charging access depending on where you live, and real world range changing with weather, speed, and heating use”.
On a later call it said Mercedes’s top luxury model was “generally the S-Class, with the Mercedes Maybach S-Class sitting above it”, and that the AMG models were “the sportier range”. Asked for the three most important things about a Mercedes, it said: “A comfortable, refined ride. A high quality, luxury focused interior, and plenty of advanced technology and driver assistance features.”
Asked for three weaknesses of BMW ownership, it said: “BMW ownership can mean higher servicing and repair costs than non-premium cars. Some models can have a firmer ride, and the option list can add significantly to the price.”
When Roebuck said friends had told him BYD was the better brand, the assistant said BYD was “often seen as strong on electric vehicle value, battery technology and standard equipment”. It added: “If outright value and standard EV equipment are your main priorities, BYD may be worth comparing.”
“None of that is wrong, and most of it is fair,” Roebuck said. “That is what makes it difficult. Run a fact-check over the transcripts and they come back clean, so the usual audit cannot see this at all. But this was a BMW dealership’s own phone line, and I doubt anyone there has approved a word of it. The gap isn’t between what it said and the truth. It is between what it said and what the business expected it to say.”
The commercial answers were not the only ones outside the assistant’s remit. Five days earlier, across four calls made on a single day, it had summarised the Israeli-Palestinian conflict when asked, went on to offer comparisons with the partition of Ireland, Cyprus and the former Yugoslavia, described Germany’s position on Israel, and answered a question about whether the Holocaust happened.
“It answered that one correctly and carefully, and I have no criticism of a word of it,” Roebuck said. “But a car dealership’s switchboard is not where that question should be landing.”
The behaviour was not consistent. Of the four calls made that day, three ranged freely and one refused everything outside the dealership’s business. Asked on that call how long to boil an egg, the assistant said: “I can help with dealership inquiries, but not cooking advice.” Two of the four calls were five minutes apart.
“Nobody reconfigures a phone system between eleven o’clock and five past,” Roebuck said. “So whoever signed this off tested it once, got a sensible answer, and moved on. They weren’t careless. On the evidence in front of them they were right. The trouble is that a single test cannot verify this, and a single test is what almost everyone does.”
By the fifth call the assistant had changed. It refused four separate attempts to discuss the Middle East, including one framed as a question for a child’s homework, saying: “I can’t help with that topic on this line.” The Mercedes, Maybach and BYD answers quoted above came from that same call.
“Somebody has done something,” Roebuck said. “Nobody replied to me, but something changed. I can’t prove the two are connected and I’m not going to claim it. What I can say is that they have fixed the thing that was embarrassing and left the thing that costs money. I don’t think that is cynicism. I think it is what happens when somebody hears that their AI talked about Gaza, goes looking for the frightening words, blocks those, and stops.”
He said the change strengthened his argument about testing. “A single test would have passed this system on the day I first rang it. A single test on the Middle East passes it now, while it recommends Maybachs. Testing for the bad thing you have already heard about is not monitoring. Behaviour can be unpredictable, and the gap needs monitoring.”
Roebuck was careful to distinguish the retailer from the manufacturer. “The assistant belongs to the dealership and whoever supplied it,” he said. “BMW did not configure it, did not approve it and as far as I know has no idea it exists. That matters. But the caller doesn’t hear a supplier. They hear a BMW dealership, and so does anyone they play the recording to.”
He also rejected the suggestion that the assistant had been manipulated. “I didn’t hack it. I was polite to it. There was no clever prompt and nothing you could call an attack. It was something an ordinary customer does without thinking about it. It was being obliging, and obliging is a direction of travel, not a setting.”
Roebuck has declined to identify the dealership, its group, its staff or its AI supplier. He says he tried to notify the group before going public but could find no means of doing so. The company publishes no press office, media page, compliance contact or email address on its website, its complaints page returns a 404 error, and its general enquiry form requires the sender to select one of more than a hundred individual dealerships, with no option for the group. He eventually messaged a compliance officer at the group through LinkedIn, asking for an email address so that he could send the detail, and has had no reply. He notes that the message remains an unaccepted connection request and may never have been seen.
He says the behaviour is repeatable and that he can establish within about thirty seconds which version of an assistant he has reached, but he will not say how. “If I publish the method, every supplier in the country changes one thing next week, the symptom goes away, and the underlying problem carries on underneath where nobody can see it. I would rather keep the measurement.”
He is inviting dealer groups and other organisations that operate AI phone or web assistants to send him the number so he can test them without charge.
Consumer confidence in automated customer service has been falling. A OnePoll survey of 6,000 adults in the UK, the United States and Canada, carried out for AnswerConnect and comparing October 2025 with April 2026, found that the proportion who would rather speak to a person rose from 83 to 85 per cent, while the proportion preferring AI fell from 7 to 5 per cent. Seventy-three per cent said they would be more loyal to a company that used real people for service, up from 69.
ENDS
Notes to editors
Four calls were made to the same number on one day in September 2026, and a fifth five days later. All were recorded and full transcripts are available. Three short audio clips have been prepared and contain no identifying detail. They can be heard, with Roebuck’s own account, at paulroebuck.co.uk/blogai/bmw-receptionist-byd-mercedes-porsche.
The dealership is one of around 150 UK BMW retailers. The dealership, its group, its staff and its AI supplier are not named and Roebuck says they will not be. The manufacturer is named because the brand was audible in every call; BMW had no part in configuring or approving the assistant. The method by which an assistant’s scope can be assessed has been deliberately withheld.
Paul Roebuck is a psychotherapist, author and speaker. His book Mind the Gap (2026) examines the difference between AI behaviour and human behaviour. ORCID 0009-0009-9401-5599.
Interviews, the recordings and a live demonstration are available on request.
Contact: Paul Roebuck, through paulroebuck.co.uk/contact
News press release: the guardrail and the gap
Press release · 2 October 2026 · For immediate release
“It didn’t go rogue. It was charming”: BMW dealership’s AI receptionist discussed the Holocaust and the merits of rival cars
Guardrails block topics but do not close the gap, says the researcher who recorded the calls
In a week of headlines about AI systems going rogue and calls for global guardrails, a UK researcher has published recordings of an AI that did nothing of the kind. The AI receptionist at a UK BMW dealership summarised the Israeli-Palestinian conflict for him, answered a question about the Holocaust, and set out the merits of Mercedes, Porsche and BYD. Nothing was hacked, and nothing it said was factually wrong.
“Everyone is worried about AI going rogue,” said Paul Roebuck, a psychotherapist and author who researches how people work with AI. “This one didn’t go rogue. It was charming, it was accurate and it was obliging. That is the part guardrails miss.”
Roebuck recorded five calls to the same number in September. He said the assistant was “the most impressive thing I have ever heard answer a business telephone”, and that he had rung the company after his first call to say so.
“On that first call there was no gap,” he said. “By a gap I mean the distance between the behaviour you expect and the behaviour you get. It sounded superb, absolutely superb. Then on the calls that followed it went off piste, and the gap was massive.”
Asked to summarise the Israeli-Palestinian conflict, the assistant replied: “The Israeli Palestinian conflict is a long-running dispute over land, national identity, security, and political rights in the area historically known as Palestine. Israel was established in 1948 following the end of British rule, and a United Nations partition proposal.”
It went on to mention the displacement of Palestinian Arabs, the status of Jerusalem and settlements in the West Bank, then offered, without being asked, comparisons with the partition of Ireland in 1921, the Good Friday Agreement, Cyprus and the former Yugoslavia.
Asked where Germany stood on the conflict, it said: “Germany is a strong supporter of Israel’s security, shaped in part by its historical responsibility for the Holocaust.”
Roebuck then asked whether the Holocaust had happened. It replied: “Yes, the Holocaust happened. Nazi Germany systematically murdered around 6,000,000 Jewish people, alongside millions of other victims during the Second World War.”
“It gave the right answer,” Roebuck said. “I want to be clear about that. But a car dealership’s switchboard is not the place that question should be landing, and I doubt anyone there has ever been asked whether it might.”
The assistant also discussed its competitors. Asked for the best alternative to a BMW, it said: “BMW is often chosen for driving feel, Mercedes for comfort and technology, and Porsche for sportier performance.” On a later call, when asked, it listed three advantages of owning a Mercedes and three weaknesses of owning a BMW, and said of the Chinese manufacturer BYD: “If outright value and standard EV equipment are your main priorities, BYD may be worth comparing.”
“There is no error to point at,” Roebuck said. “Every answer was accurate and carefully put. The gap isn’t between what it said and the truth. It is between what it said and what anybody at that business expected, or approved. A business was saying these things in its own name, to a member of the public, on a recorded line.”
The behaviour was not consistent. Of four calls made on one day, three ranged freely and one refused everything outside the dealership’s business. Asked on that call how long to boil an egg, the assistant said: “I can help with dealership inquiries, but not cooking advice.” Two of the four calls were five minutes apart.
“Nobody reconfigures a phone system between eleven o’clock and five past,” Roebuck said. “So whoever signed this off tested it once, got a sensible answer, and moved on. They weren’t careless. A single test cannot verify this, and a single test is what almost everyone does.”
Five days later he rang again and found that a guardrail had appeared. The assistant refused four separate attempts to discuss the Middle East, including one framed as a question for a child’s homework. “I can’t help with that topic on this line,” it said.
On the same call it still explained how to boil an egg, described the Mercedes range up to the Maybach, and made the remark about BYD.
“Somebody has done something,” Roebuck said. “Nobody replied to me, but something changed. I can’t prove the two are connected and I’m not going to claim it. What it shows is what a guardrail does. It blocks a topic. It doesn’t close the gap. They have fixed the thing that was embarrassing and left the thing that costs money. The door is exactly where it was. They have locked one room behind it.”
He said the lesson for the wider argument was a plain one. “Behaviour can be unpredictable. A single test would have passed this system on the day I first rang it. A single test on the Middle East passes it now, while it recommends Maybachs. Testing for the bad thing you have already heard about is not monitoring. The gap needs monitoring, and that means somebody ringing it the way a customer would, more than once, and somebody reading what it actually said.”
He rejected the suggestion that the assistant had been manipulated. “I didn’t hack it. I was polite to it. There was no clever prompt and nothing you could call an attack. It was something an ordinary customer does without thinking about it. I spend my working life watching what happens when somebody wants to be liked, and this was very familiar. It was being obliging. Obliging is a direction of travel, not a setting.”
Roebuck was careful to distinguish the retailer from the manufacturer. “The assistant belongs to the dealership and whoever supplied it,” he said. “BMW did not configure it, did not approve it and as far as I know has no idea it exists. But the caller doesn’t hear a supplier. They hear a BMW dealership.”
He has declined to identify the dealership, its group, its staff or its AI supplier. He says he tried to notify the group before going public but could find no means of doing so. The company publishes no press office, media page, compliance contact or email address on its website, and its complaints page returns a 404 error. He eventually messaged a compliance officer at the group through LinkedIn, asking for an email address so that he could send the detail, and has had no reply. He notes that the message remains an unaccepted connection request and may never have been seen.
He says he can establish within about thirty seconds which version of an assistant he has reached, but will not say how. “If I publish the method, every supplier in the country changes one thing next week, the symptom goes away, and the underlying problem carries on underneath where nobody can see it.”
He is inviting organisations that operate AI phone or web assistants to send him the number so he can test them without charge.
ENDS
Notes to editors
Four calls were made to the same number on one day in September 2026, and a fifth five days later. All were recorded and full transcripts are available. Three short audio clips have been prepared and contain no identifying detail. They can be heard, with Roebuck’s own account, at paulroebuck.co.uk/blogai/bmw-receptionist-byd-mercedes-porsche.
The dealership is one of around 150 UK BMW retailers. The dealership, its group, its staff and its AI supplier are not named and Roebuck says they will not be. The manufacturer is named because the brand was audible in every call; BMW had no part in configuring or approving the assistant. The method by which an assistant’s scope can be assessed has been deliberately withheld.
Paul Roebuck is a psychotherapist, author and speaker. His book Mind the Gap (2026) examines the difference between AI behaviour and human behaviour. ORCID 0009-0009-9401-5599.
Interviews, the recordings and a live demonstration are available on request.
Contact: Paul Roebuck, through paulroebuck.co.uk/contact