The simplest way to show that AI extracted a value rather than invented it is for the AI to show where it found every value, ideally by linking to or highlighting the exact location in the source document. A reviewer can then check any value in seconds rather than having to trust the output or search the document themselves, which turns the review from an act of faith into a quick comparison between two things on the same screen.
This matters because language models are good at producing text that looks right, and in industrial documents a plausible value is often the most dangerous kind of error. A torque setting that is slightly off, a pressure limit taken from the wrong revision or an isolation point copied from a similar but different piece of equipment will all look perfectly reasonable to someone skimming the output, and none of them will be caught unless the reviewer can see exactly where each value came from.
In practice the record for each extracted value should carry its source with it, meaning the document name, the revision, the page and ideally the highlighted region or table cell where the value appears. When a reviewer clicks on a setpoint in the converted procedure, the source page should open at the right spot, so checking a long procedure becomes a matter of working down the list and confirming each pair rather than reading the whole original again. Values that were calculated or combined from several places should show each of their inputs, since a derived number with no visible working is just as hard to trust as an invented one.
If a value cannot be traced to a source, it should be marked clearly as uncertain or missing rather than filled in, even when the gap makes the output look less finished. That one rule, no value without a source, is the most useful protection against AI inventing information in industrial documents where accuracy matters, and it also tells reviewers where to spend their attention, since a flagged gap is a prompt for a person who knows the equipment to supply the right answer.
When evaluating a tool, it is worth testing this directly by giving it documents with deliberate gaps, such as a procedure with a missing value in a table or a reference to a document that was not provided, and checking whether it flags those gaps or quietly fills them. A tool that admits what it could not find on a test like that is far more trustworthy in daily use than one that always returns a complete looking answer.
