When an AI tool extracts information from a document and presents it as output, there are two meaningfully different things it could be doing. It could be producing a claim — here is what I found, trust it or don’t. Or it could be producing auditable work product — here is what I found, at this location, which you can verify. The difference is the citation. And for professional use, the difference is not aesthetic; it determines whether the output can be used at all.

Professionals working with documents aren’t evaluating AI outputs the way a curious user evaluates a search result. They’re making decisions that carry consequences, and those decisions require a chain of justification. The accountant can’t just accept that revenue is a certain figure; they need to know where that figure came from so they can attest to it. The analyst can’t sign off on an assumption they can’t verify against the source material. The lawyer can’t rely on a representation they can’t cite. When an AI tool produces an output without telling you where it came from in the document, it’s producing something the professional can’t use — not because they distrust AI, but because their job requires them to be able to say: this number came from page 14, line 3, and I verified it.

The citation is what converts the output from a shortcut into a tool. The shortcut saves the time of finding the information; the professional still has to verify it, and without a citation they have to find it themselves, which largely negates the time saved. The tool, with a citation, saves the finding time and compresses the verification time into a single glance — you check the cited location rather than scanning the document. That’s a fundamentally different value proposition: not “here’s an answer you’ll have to verify from scratch” but “here’s an answer, pre-located for you.”

This is why provenance isn’t a polish feature to add after the core extraction is working. It’s load-bearing. The extraction without provenance produces outputs that a professional audience will treat as suggestions at best — a starting point that still requires full manual verification. The same extraction with provenance produces outputs that can be reviewed rather than re-found. For the professional who processes many documents, the difference compounds: reviewing is fast, re-finding is slow, and the tool that makes the professional’s job easier rather than merely different is the one they’ll pay for.

The practical consequence is that provenance should be designed in from the start, not retrofitted. Retrofitting is harder than it sounds — the citation requires knowing not just the extracted value but where in the source document that value was found, which means the extraction process has to preserve location metadata from the beginning. Building the extraction first and adding citations later often means rebuilding the extraction layer anyway. The citation isn’t a feature that lives on top of the extraction; it’s part of how the extraction has to work.

The cleaner principle: if the intended user needs to be able to verify the output — and in professional domains they always do — then an output without provenance is an incomplete output. Complete the output.