Why the technician still signs: lessons from AI in a vehicle inspection centre
In the SCANNO inspection platform, AI drafts the report and a technician approves it. Running it on real cars showed where AI goes wrong, and why the approval step is the product and not a formality.
Amir Diab, founder of Instanix5 min read
The short answer
AI is fast at organising what a technician found and writing it clearly. It is not the one who looked at the car. So in the platform we built for SCANNO, a vehicle inspection centre in Doha, the AI proposes and the technician decides, and no report reaches a customer without a technician’s approval.
That rule did not come from caution alone. It came from what we saw when real inspections started running through the system.
What went wrong, and what we changed
The AI mixed up left and right when reading photos. A scratch on the left door is not a small error when it is written as the right door. We stopped it from writing the side at all, and left that to the technician.
One technician’s phone was sending photos to the model rotated. No lab test had shown it, because our test photos were upright. We found it by reviewing a real inspection.
Two technicians inspected the same car at the same time and produced two different reports. The system now warns before the analysis starts.
On a car whose computer had been replaced, the diagnostic device read a chassis number different from the one the technician had typed. Comparing the two became an automatic check.
The customer assistant wrote the centre’s name wrongly in Arabic, because its instructions were written in English. We added a guard on the name.
The guardrails that matter most
A section that was not inspected is written as “not assessed”. An AI asked for a full report will happily fill an empty section with something plausible, and a plausible sentence about brakes nobody checked is the worst thing an inspection report can contain.
A photo is evidence only for its own section. Scores are suggestions that the technician can change, with a written reason. And the assistant that answers customers gives no prices, no repair times and no opinion on whether to buy the car.
What this means for your own project
If AI is going to write something your customer relies on, decide before you build who signs it. Then design the screen around that person: make it quick to accept what is right and easy to reject what is wrong.
And review real work after launch, every case at first. The problems above were all found that way, not in testing.
Keep reading
AI agent or automation: which one does your business need?
Automation follows rules you can write down. An AI agent handles work that needs reading, judging and wording. Most companies need automation first, and an agent only where the rules run out.
What to automate first: a simple way to choose
Start with the process that is repeated often, follows clear rules and hurts when it is late. Automate one process end to end before starting a second.
