Ask a general-purpose chatbot about your retention policy and it will give you a fluent, confident, well-structured answer. The problem is that fluency and accuracy are unrelated. The model would produce the same confident tone whether it was right, subtly out of date, or inventing section numbers.
For a marketing team drafting social posts, that’s a tolerable trade. For a compliance officer, a clinician, or a lawyer, it’s disqualifying. In regulated work, an answer you can’t verify isn’t an answer — it’s a rumour with good grammar.
The mechanism that fixes it
A citation, properly done, is a link from a claim to the exact passage it rests on — not to a document, to a passage. Click it and the source opens beside the answer with the relevant text highlighted. Verification stops being a research task and becomes part of reading.
This changes the economics of trust:
Checking is cheap. When verification costs one click, people actually do it. When it costs a search through a 200-page PDF, they either don’t check (risk) or don’t use the tool (waste).
Errors surface fast. If a citation doesn’t support the claim, you see it immediately — and you’ve learned something about the answer, the document, or both.
“The documents don’t say” becomes possible. A system built around evidence can decline to answer when the evidence isn’t there. That honesty is only meaningful because the alternative — showing the supporting passage — exists.
The audit question answers itself. “How do we know this?” is a permanent question in regulated organisations. With cited answers, every response arrives already carrying its own paper trail.
What this looks like in practice
When we built the document analysis system for BAiSICS — legal documents, many of them large, complex and poor quality — source verification wasn’t a feature request, it was the requirement. Lawyers won’t act on an extraction they can’t check against the instrument itself. With verification built in, review time fell from around two hours per document to about ten minutes, and the accuracy claims survived because anyone could test them against the source in seconds.
That experience shaped Marella: citations aren’t an add-on panel, they’re how every answer is constructed. The retrieval that produced the answer produces the receipts at the same time.
The buying test
If you’re evaluating AI for an organisation where wrong answers have consequences, make this the first demo request: ask a question, then make the vendor show you the exact source passage for each claim in the answer — in one click, not a scavenger hunt.
Systems built on evidence handle that request without flinching. Systems built on confidence change the subject.
