The contradiction is the finding

A request for quotation arrives with three documents: an email, a drawing and the client's specification. The drawing says PP-T20. The specification says PP. Which one is right?

A system that reads documents and produces a summary has to do something with that line. The tempting behaviour, and the one many AI tools fall back on, is to pick the most plausible answer and move on. The summary looks complete, and nobody notices that a decision has been made.

Why picking is the wrong answer

In a quote, a contradiction is rarely a typo you can safely correct. The two materials shrink differently, need different steel and give different part weights. Picking one quietly changes the price, and the mistake shows up weeks later, when the first samples don't fit.

The person who should resolve it is the client, and the earlier they're asked, the cheaper the answer. That makes the contradiction itself the most valuable line on the summary sheet. It's a finding, not noise to be smoothed over.

How we build for it

• Every value on the sheet is traced to the document it came from, so a disagreement between two documents stays visible instead of being averaged away.
• Where sources disagree, the sheet shows both values and marks the line as contradictory. It never chooses.
• Only the contradictions and gaps that block the quote go to the client. Questions your own team should answer, like steel grade or gate position, stay off the list.
• The reply asks for exactly those points, in the client's language, and waits for a person to send it.

A test you can run

If you're evaluating any AI tool that reads technical documents, give it a request where two attachments disagree. Then look at what it produced. If the summary is confident and clean, ask where the other value went. The tools worth trusting are the ones that tell you what they don't know.