If the AI can't show you the page, it didn't read it

Someone in the meeting pointed at a figure in the summary and asked the simplest possible question: where did this come from? The source document was four hundred pages. The extraction pipeline had pulled out a number, dropped it into a tidy table, and moved on. No page, no paragraph, no trail. Three people spent the next twenty minutes scrolling a PDF trying to reverse-engineer where their own system had found its answer, and they never did agree it was right.

That is the hidden cost of extraction that returns only the answer. A clean value in a clean cell feels trustworthy precisely because the mess it came from has been hidden. You are not looking at a fact. You are looking at a claim with its evidence deleted.

An extracted value you cannot trace is a rumor in a spreadsheet.

The systems that get this right never return a bare value. Every field arrives with its provenance attached: the page, the location on that page, and the exact span of text the value was read from. The output is not "amount: $40,000." It is "amount: $40,000, found here," with a link that jumps you to the highlighted line in the source. Checking a field stops being an investigation and becomes a click.

This is a design decision, not a feature bolted on later. The pipeline is built so that a value cannot exist without a citation, the same way a serious claim cannot exist without a footnote.

Here is the part that matters more than accuracy. A model asked to extract a field will almost always produce something, because producing something is what it was trained to do. On a document that simply does not contain the answer, it invents a plausible one. The discipline that fixes this is constraining the model to extract only what it can ground in the text, and to abstain, explicitly, when the support is not there. A blank that says "not found" is worth more than a confident guess, because the blank routes to a human and the guess routes to a slide.

A wrong answer that flags itself is cheap. A wrong answer that looks certain is the expensive one.

This reframes the whole chase. Teams pour months into pushing extraction from 94% to 96% accurate and still cannot trust the output, because they cannot tell which 4% failed. Grounding and abstention solve a better problem: not "is every field right," but "can I see, in one click, which fields to check." A pipeline that grounds every value and refuses to guess is one a human can audit in minutes instead of re-reading the source.

Documents are exactly where everyone is now pointing language models, and a model reading a contract or a financial statement is fluent enough to be convincing whether or not it is correct. Confidence stopped being evidence. The only durable defense is making every answer checkable at its source, by construction, so trust comes from being able to verify rather than from the output sounding right.

We built this for a team drowning in manual document review, and the reviewers stopped reading documents end to end. They read the flagged fields and the citations, and the work went from hours to minutes because the system showed where it looked.

Make the AI show you the page and extraction stops being something you believe. It becomes something you can check, which is the only kind of automation worth trusting with a number that ends up in front of a customer.

TensorLabs builds the grounded-extraction and document-AI infrastructure behind that kind of verify by-clicking pipeline.