Does the field report explain what was checked?

Installation complete, and the test result?

A company sets up printers in its clients' offices. After the visit, the technician has to record which device they installed and how the test print turned out. The coordinator opens the report and sees: „Installation complete.” The field is filled in, but someone still has to call the technician for the test result.

Help is worth most while the technician still remembers the visit: „The description has no test print result.” The author adds: „Printer installed at reception. The test page printed in full.” The suggestion points to the gap and leaves the description to the technician. If the print had failed, that is exactly what the report should say.

What a simpler change gives, and what training adds

In this count, a separate result question in the form gives most of the effect for a fraction of the cost. Training makes sense at a larger scale or when the remaining errors are expensive. Unneeded notes cost something too: if the technician gave the result in their own words and the system still asks for more, it encourages a stock phrase. In the trial we count gaps found and false notes separately.

The set of training examples has to be maintained: when the process, the naming or the base model changes, the evaluation is repeated on the same test set. We put that cost into the quote together with the training itself.

A separate question can solve the problem

If every setup needs the result of the same test, add it as a separate question in the form, with a check for an empty answer; service apps usually have required and conditional questions. A difference between an answer and a description of the result remains: you can type „done” into the „Test print result” field. A model has work to do when the company needs the content judged. First, the people who read the reports agree what needs completing, because a model tuned on conflicting judgements will repeat the conflict. Google describes supervised tuning of models on examples with a correct judgement.

Where to start

The comparison covers reports not used for customisation. A technician and a coordinator judge whether the notes lead to a needed addition, and how many notes were unneeded. Bringing many notes together is covered in the article on a local model for service notes. We run model customisation as part of private LLMs, and the application in which the employee sees the result beside its source as part of AI applications. For the free analysis, a few examples of an error that keeps coming back and a note on who corrects it today are enough.