From dictation to a finished expert report for a used-car inspector

The starting point

The expert inspects used cars on behalf of buyers, roughly fifty a month. After the on-site inspection the customer receives a one-page expert report, and that document is the tangible product of the service. Until now it was produced with a general-purpose AI chatbot: the expert typed or dictated what he found, and the chatbot wrote it up.

It worked, but it had to be done from scratch every time, and the document was regenerated on every run. Its layout and colours changed from inspection to inspection, typos stayed in, and some passages repeated from car to car. At fifty inspections a month every document needed separate rework. The bigger problem was that there was no guarantee that only what the expert had actually found on that particular car made it into the report.

The solution

A phone app that produces the expert’s own document and leaves every decision in his hands.

• Dictation at the car. One large Record button, pausable. A finished audio file or typed text can be submitted too, if the inspection was not recorded on the phone.
• Sorting into his own structure. In about a minute the AI sorts the dictation into the expert’s 17 sections, from suspension to battery, in the order he works in. Processing runs on the server, so the phone can be locked in the meantime.
• A source for every statement. One tap shows which dictated sentence a finding came from. Anything the expert did not talk about is marked “Not inspected” in the document, instead of being filled with invented text.
• Review and photos. Each section’s status (OK, Attention, Fault) is set with a tap, the text is editable, and up to two photos of the car can go into the document.
• A consistent, branded document. After approval a one-page PDF or image is produced, always with the same layout and the expert’s logo. It can be shared directly on Viber and Messenger, because that is where customers ask for it. Hybrid cars get a separate battery-health box.

Deliberate limits. The AI does not decide and does not rate: the report is always approved by the expert, after a separate confirmation. There is no automatic sending. It does not fill empty sections, it takes measured values verbatim only, and it adds no professional explanation of its own. A purchase recommendation goes into the document only if the expert said it.

Why a general chatbot is not enough

The difference is not in the quality of the prose but in three things a general chatbot cannot guarantee.

1. A fixed structure. The 17 sections and the one-page layout are not the model’s choice but part of the system. Two inspection reports look the same, and both the customer and the expert find the same thing in the same place.
2. Traceability. Every statement has the dictated sentence behind it. If something was not said, that is visible. This is what lets the expert put his own name on the document.
3. An approval gate. The document does not finish itself. Before approval the system warns about unsourced or uncertain statements, and without confirmation it issues neither a PDF nor an image.

How we worked

1. Samples and questions. Previous documents from the expert and a short questionnaire about volume, devices and branding.
2. Demo first. A working demo with fictional dictations, before the first meeting. We did not talk about what it would be; we talked about a clickable system.
3. His own template. Switching to the expert’s 17-section, one-page document and his own branding, instead of the earlier generic structure.
4. Test rounds. Browser and automated tests and fixes before anything touches live data. All data in the demo is fictional: sample dictation, made-up customer name, plate and make.

Next steps

During the pilot the expert uses the system with his own dictations, his own logo, signature and stamp. That is also when we measure how long a document takes before and after. Business results are measured after live use, so this study contains no estimated savings; it will be updated with the results.

All data in the video and screenshots is fictional. The client’s name is withheld at the client’s request.