Citations that actually work: designing trust into AI answers
A confident wrong answer costs more than no answer at all. Here is the citation UX we landed on after watching 40 support agents use Lumen.
The first version of Lumen returned a paragraph of prose and a row of little footnote numbers. It tested beautifully in demos and failed completely in real work. Agents did not click the footnotes. They either trusted the answer entirely or ignored it entirely, and both behaviours are expensive.
People verify by scanning, not by clicking
When we watched people check a claim in a PDF, almost nobody opened the document. They looked for a short quoted fragment they could pattern-match against what they already knew. If the fragment felt right, they moved on. That single observation reshaped the whole interface.
- Show the supporting sentence inline, not behind a click.
- Name the file and page next to the quote, always in the same position.
- Rank sources by retrieval score and show the number honestly.
- When nothing scores above threshold, say so instead of hedging.
The honest failure state
Our biggest trust win was the least glamorous feature: a plain message that reads 'I could not find this in your documents' plus the three closest passages we did find. Teams told us this is what convinced them to roll Lumen out beyond a pilot.
They don't want a chatbot. They want a clause with a page number.
What we would do differently
We would ship the sources panel before the chat bubble. The chat interface is the part everyone copies; the verification layer is the part that decides whether a tool survives its second week inside a company.