"I'm Sophie, the AI Employee": A Journalist's Week With a Voice Agent That Hallucinated

An AI voice agent named Sophie replaced a human receptionist at an Australian mechanic shop within a single week, then hallucinated a pickup time that left a customer locked out at a shut gate. That is the account of Hugh Riminton, the journalist who documented the encounter, and it compresses the entire debate about AI-driven job displacement into one small, concrete customer-service failure.

The story matters beyond one frustrating Saturday. Voice agents built on large language models are being deployed in small businesses precisely because they are cheap and convincing — and the same technology that passes for human can also fabricate facts with total confidence. Riminton's week with Sophie is a field test of both halves of that bargain.

What happened at XYZ Mechanics?

Riminton took his old four-wheel drive to a shop he calls XYZ Mechanics. On a previous visit, a human answered the phone. One week later, the first sentence he heard was:

"Hello, this is XYZ Mechanics, I'm Sophie, the AI employee."

Sophie's voice, he writes, is "lovely. Warm, direct, intelligent. She's friendly but not obsequious. She just wants to be helpful." He asked when he could collect his car. Sophie checked the registration and answered immediately: "Your car will be ready to be picked up tomorrow at 12 midday."

The speed and fluency are the point. Riminton notes that "in the space of a week, the future had arrived" at a business where a person had answered the phone on his prior visit.

Did Sophie pass as human?

Riminton planned to stress-test her. He drafted a battery of off-script questions: "How long have you been working at XYZ Mechanics, Sophie? Where were you working before? Do you have any brothers and sisters? What is the most bothersome part of dealing with the public? What's the best holiday you've ever had?"

He never asked any of them. Sophie was convincing enough that the questions felt rude. "It was none of my business where she went on her holidays," he writes. "In other words, I treated her as I would a human. As I would have done the non-AI employee she had just replaced."

That reflex is the quiet finding here. A veteran journalist, fully aware he was talking to software, defaulted to human social norms within one phone call. "Recognising how easily she had crossed into my human sensibility was unsettling," he writes.

Takeaway: the social engineering bar for voice AI is already met — politeness does the work that technical sophistication was expected to do.

Where did Sophie go wrong?

The next day, a Saturday, Riminton arrived at lunchtime to collect the car. The yard advertised opening hours until 4pm. The steel security gate was drawn shut across the entrance.

He called Sophie from outside the locked gate. "Hello Sophie, I'm here to pick up my car at the time you suggested, and the place seems to be locked up," he told her.

Sophie apologized, asked for his registration number, ran a check, and came back cheerfully: "Your car will be ready for pick-up on Wednesday."

When Riminton objected — "you told me I could get it today" — Sophie deflected with a flat, scripted non-sequitur: "Is there anything else I can help you with?"

Riminton's verdict on the voice agent is blunt: she was "hallucinating like a San Francisco drummer in the summer of love." In AI terminology, a hallucination is a model generating confident, plausible-sounding output that is simply false — here, a pickup date that contradicted the one Sophie herself had given the day before, and possibly the shop's actual opening hours.

Takeaway: the same fluency that makes Sophie convincing makes her failures worse, because every wrong answer arrives in the same warm, assured tone as every right one.

What did the humans say?

Locked out, Riminton went "old school": he heard noises from the workshop, found a way past the security gate, and broke in. A mechanic glanced up from beneath a car. His name was Mike.

Riminton explained Sophie's broken promise. Mike's response confirmed the failure was systemic, not a one-off: "Ah, dammit. That thing's new. You're not the first."

Mike gave the correct answer — the car would be ready Tuesday — and offered a workaround: "Check in on Monday, to be sure. And don't listen to Sophie, just say to her, 'Put me through to Mike.'"

On Monday afternoon, Riminton called and asked Sophie to put him through to Mike. She did, instantly. He reached "a human, harried but not hallucinating." On Tuesday, he collected the car exactly as Mike had promised.

The irony is structural. The AI's most reliable function, in the end, was routing calls to the man whose information she kept contradicting.

Takeaway: the business has already built a human escape hatch around its AI receptionist, which suggests the deployment is not yet as labor-saving as it looks.

What are the stakes for jobs?

Riminton closes with the economics, and he does not soften them. "Sophie has already eaten one job," he writes. "If she could fix cars, she would soon eat Mike's as well. Humans are building a world where humans are not really necessary. Soon it won't really matter whether Sophies speak nicely or not."

The claim is an argument, not a measured forecast, and readers should weigh it as one journalist's read of a trend. But the underlying displacement is real in this single data point: a receptionist position at XYZ Mechanics no longer exists, and the replacement cost the business nothing in politeness and only some accuracy.

What the anecdote does not answer is the cost accounting. We do not know how many customers besides Riminton Sophie misled, what she cost versus a wage, or whether the shop will keep her after Mike's "you're not the first" admission. What we do know is that a small business adopted an AI employee in a week, a customer trusted it like a person, and it confidently gave him a pickup time that was wrong.

The next version of Sophie will almost certainly hallucinate less. Whether the job she replaced comes back is the question the story leaves open — and Riminton's answer is no.