By the third week, Sam had 37 tabs open. Sam is imagined, but you’ll probably recognise him. He’s a dad with two kids and a large, slightly smelly retriever, and he’s got it into his head that a restored G Wagon from the 1990s is the right car for weekend trips to the lake. His wife isn’t fully convinced yet. His spreadsheet has a column headed “vibes,” and it’s the only column he’s filled in.
Every listing he reads describes restorations in its own way, and some sellers barely write anything. One photo of a sand-colored Wolf parked by a river had him ready to wire a deposit before he’d asked a single question about service history.
AI did help him in the end. It just took him a couple of tries to learn how to use it.
His first prompt was useless
“What’s the best classic SUV?” got him a cheerful top-ten list covering Land Cruisers, Broncos, a Range Rover and a Scout. None of it said anything about his actual life.
The second attempt went much better because he told the chatbot about himself. He needs two kids in the back, plus the dog. Most trips are under three hours. The car would sleep in a garage. He can change oil and swap a battery, but he does not want a project car.
With that context, the AI stopped ranking vehicles and started giving him questions to ask. Would a child seat fit properly in the back of a 30-year-old military truck? Who near him could work on an old diesel? He hadn’t thought of either.
Make it show its working
A chatbot can blend what a seller actually said with what it assumes about the model in general, and the result sounds equally confident either way.
So Sam started pasting in the raw listing text with one strict instruction: pull out only what’s written, and put “unknown” wherever the seller is silent.
G-Wagens turned out to be a good place to learn why. A lot of 1990s restorations, including Expedition Motor Company’s, start from ex-military 250GD “Wolf” models. Most of what’s online describes civilian G-Classes, which are a different animal. Early on, the AI gave Sam a list of “standard features” that turned out to belong to the wrong version. Nothing in the listing supported them.
The prompt that fixed it:
Using only the text I’ve pasted, compare these listings. For each feature, quote the sentence it comes from. Mark anything not mentioned as unknown. Then list what I should ask before booking an inspection.
With quotes attached, it took him about two minutes to spot where the AI had wandered off the page. It’s the same habit as knowing how to check AI-generated content at work, applied to a car.
He used Expedition’s own build information the same way. He didn’t ask the AI to vouch for the restoration. Moreover, He asked it to find the holes, things like which transmission it has, what was rebuilt and what was replaced, and what exactly the two-year warranty covers. That gave him a short, specific list.
Photos told him what he hadn’t seen
He was tempted to ask the AI whether the frame looked rusty. He asked a better question instead: which parts of the car do these photos not show?
The answer was a fair amount. There was nothing from underneath, the engine bay appeared in a single shot from a strange angle, and the roof fittings didn’t show up at all. Sam turned that into a polite email asking for eleven more pictures.
Even when the photos did show something, he stayed sceptical. Compressed images and glare can hide plenty. “Looks rust-free” really only means the photos don’t show any rust.
The same went for history questions. The AI happily explained differences between production years, and some of it was even right. He checked the parts that mattered with a G-Wagen owners’ forum and a specialist mechanic who’d worked on Wolfs before.
The paperwork became a to-do list for the inspector
Once the seller sent over a folder of records, Sam asked the AI to put everything in date order and separate work that was actually done from work that was only recommended. He also told it to flag any document that disagreed with another one rather than quietly picking one. It found two, a mileage note and a pair of invoices with overlapping dates.
Before uploading anything, he blacked out previous owners’ names and addresses. Most chatbot services keep some record of what you share, and he’d read enough about whether ChatGPT is safe not to hand over other people’s details for no reason.
The timeline saved the independent inspector real time. Instead of starting from scratch, they could go straight to the useful questions. Does the documented work match what’s actually on the lift? What’s original? What does the paperwork skip entirely? The inspector still did the part only a person with a torch and a creeper can do.
The last prompt didn’t ask for a winner
Sam’s final request was simple: list everything still unanswered. It came back with nine items, and three of them only a test drive could settle.
He drove the car on a Saturday. It’s loud, slower than he expected on the highway, and the dog loved it. He found a diesel mechanic forty minutes from home who’d take it on. His wife took it round the block once and came back grinning, which more or less settled the vibes column.
He hasn’t signed anything yet. When he called the seller on Monday, though, he had a clear list of what was known, what wasn’t, and four questions he wouldn’t have thought to ask a month earlier.
Related: How Does AI Keep Getting Smarter? 5 Reasons Explained
