Responsiveness sounds simple. Is this document relevant to the request or not. In practice it is the call that decides how much of a matter you have to hand over, and it is the one most likely to go wrong at scale.
Here is why. On a large review, thousands of documents get a yes or no from someone reading fast, late in the day, on their hundredth file. A keyword search casts too wide a net and buries the team in noise. Miss a responsive document and you risk a discovery obligation. Over-produce and you hand the other side more than you owe, and you pay to review all of it first.
Why don’t keyword searches find what’s responsive?
The old approach starts with a search string. If the term appears, the document lands in the review pile. But relevance is about what a document means, not whether it contains a word. An email that never uses the project name can still be squarely responsive. A document stuffed with the search term can be noise. So the team reads everything the search returned, and the cost climbs whether or not the search was any good.
Broad requests make this worse. A wide request makes almost everything look arguably relevant, and internal code names and project references mean nothing to a plain keyword. The reviewer has to carry all that context in their head, document after document.
What changes when an agent reads first?
Our responsive agent reads each document the way a reviewer would. It works from your request and your case context, including the code names and shorthand that a keyword search cannot know, and it decides whether the document is actually responsive. Then it gives you the reason, pointing at the text it relied on.
That does two things. It catches the responsive document that never mentioned the search term. And it clears the noise that only matched on a word, so your team is not paying to read files that were never relevant.
| Keyword search first | Agent reads first | |
|---|---|---|
| What decides relevance | Whether a word appears | What the document means, against your request |
| Code names and shorthand | Missed unless someone adds them | Read in context from your case background |
| Noise | Everything that matched gets read | Word-only matches are cleared, with a reason |
| Where the reviewer starts | A blank page | A tagged document and a short rationale |
| When someone asks why | Reconstructed from memory | The cited reason is on the record |
What does the reviewer actually do?
Every document arrives already tagged, with a short rationale attached. Your reviewer confirms or overrides, and the close calls rise to the top instead of hiding in the pile. The first pass is done. The judgment stays with your team.
Why does this hold up later?
Because each responsive call cites the text behind it, the review is explainable after the fact. When someone asks why a document was in or out, the answer is on the record. That is the difference between a defensible position and a reviewer trying to remember a decision from three weeks ago.
More on that in What “defensible AI review” actually means. For the bigger picture, start with First-pass litigation review, brought in-house.
Where to start
See how the responsive agent reasons on a real set of documents. Book a demo.