Three hours in the garage: a multimeter, a mechanic, and an AI
Yesterday I was at a friend's auto shop — my car was getting new shock absorbers. While it hung on the lift, next to it stood a W207 coupe that wouldn't start: diagnostic port dead, head unit dead, fuel pump silent. And yet the starter cranked and the cluster lit up.
The three of us sat down over it: the mechanic, me, and an AI assistant on my phone.
We started badly. The AI produced three confident hypotheses — a short from the steering rack, a relay in the fuse box, a pyrotechnic battery disconnect. All three turned out wrong. Had the mechanic acted on them, he'd have lost an hour tearing down a healthy unit.
Then it got interesting. The multimeter showed 7 volts where there should be 12: looked like sagging power. The AI pushed back — that's phantom voltage on a high-impedance meter, test with a bulb instead. A false trail died in a minute. Then it drew the map: what's dead, what's alive. And it turned out this wasn't five separate faults but one shared power rail. The mechanic backfed 12 volts through the diagnostic port — the whole "dead" group came alive, and the car started on the backfeed. So the front fuse box simply isn't receiving its main power feed: the section is localized, the exact break point in the run is still to be found. The repair is ahead.
An honest tally. The AI didn't "find the fault" — it was confidently wrong three times, and the decisive test was the mechanic's idea and the mechanic's hands. But it held the method: correct measurements, a power map built from symptoms, reading photos of fuse boxes and diagnostic screens. The whole thing took three hours; by my estimate, without that map the car would have sat in the bay another day — and an occupied bay is direct money for a shop.
What I drove home with is a conclusion about AI in physical businesses: what's useful isn't an oracle but a second diagnostician — one that keeps the measurement list, doesn't tire, and is never taken at its word until a measurement confirms it.