The Hidden Cost of AI Document Review Is Verification Time

The hidden cost of AI document review is not the subscription

Ask a litigator what AI costs and you will hear a monthly figure. That is the wrong number. The real cost is the time between reading a machine’s answer and being willing to put your name on it.

A summary that says the witness conceded the timeline was approximate is worth nothing on its own. Before it can go in a brief, a motion, or a cross-examination outline, someone has to find that concession in the transcript, read the surrounding exchange, and confirm the witness said it in the way the summary implies. If that takes twenty minutes, the tool did not save twenty minutes. It moved them.

This is why two products with identical accuracy can have completely different value. The one that tells you where it got the answer is the one that saves time.

What a page and line citation actually does

In deposition and trial practice, the record is addressed by page and line. A designation reads 47:12-48:3. A court reporter’s transcript is numbered precisely so that any statement can be located by two integers and checked by anyone in the room.

That convention exists because litigation runs on verifiability, not recall. Opposing counsel is entitled to check your characterisation. The judge is entitled to check it. If you cite a page and line, they can. If you paraphrase, they cannot, and the burden of proof quietly shifts back onto you.

When an AI system answers a question about a transcript, the same rule should apply. An answer of the form:

The witness placed the meeting in “late September, maybe October” — [doc: Hernandez_Depo][page: 47][line: 12]

can be checked in seconds. An answer that says the witness was vague about the date cannot be checked at all without re-reading the deposition, which is the work you were trying to avoid.

Why generic summarisation fails on transcripts

Transcripts are unusually hostile to naive AI summarisation, for three specific reasons.

Length. A single day of testimony runs 200 to 300 pages. A multi-week case produces thousands. Most of it is procedural noise — objections, breaks, the reading of exhibit numbers. The signal is thin and scattered.

The answer is often in the follow-up. A witness says “no” and then, four questions later, qualifies it into a “yes, but.” A system that retrieves one passage and stops will confidently report the “no.” Context has to travel with the citation.

Precision matters more than fluency. In most writing tasks, a near-miss paraphrase is fine. In a deposition designation, “approximately” and “about” are not interchangeable if the other side is going to argue about it. What was actually said, in the words actually used, is the deliverable.

What to require of any tool you let near the record

If you are evaluating software for transcript work, the questions worth asking are narrow:

1. Does every answer carry a location? Document, page, line. Not “somewhere in the Hernandez deposition.”
2. Can you jump straight to that location and see the surrounding exchange? A citation you cannot open is a footnote, not a verification.
3. Does the search return context, or just the matching line? You need the question that prompted the answer.
4. What happens when the answer is not in the record? A system that would rather invent something than say “not found” is unusable for this work, no matter how well it writes.
5. Where does the text go? Deposition transcripts contain privileged and confidential material. For some matters, the only acceptable answer is that the document never leaves your infrastructure — which is why a locally hosted, offline model option is not a luxury feature in this field.

That last point deserves emphasis. There is a real difference between a tool that sends your client’s testimony to a third-party API and one that can run the analysis on a machine you control. For a criminal matter, a sealed record, or anything under a protective order, that difference decides whether the tool can be used at all.

The arithmetic

Suppose a paralegal spends thirty hours reviewing a set of depositions to build a designation chart. Suppose a search tool cuts the locating work to six hours, but every citation still has to be opened and read before it is used — call it another four. That is a real saving of twenty hours.

Now suppose the tool summarises beautifully but does not cite. The locating work returns in full, because nothing can be used until it is found and confirmed. The saving is close to zero, and you have added a new risk: something that reads well enough to be trusted without checking.

The second tool will demo better. The first one is the one that pays.

Where Lawnova PDF sits

Lawnova PDF was built around the first pattern. Transcripts are indexed with page and line precision, search returns the surrounding exchange rather than an isolated hit, and every AI answer carries [doc][page][line] references you can open. It runs against a cloud model or against a local, airgapped model for material that must not leave your environment.

That is a narrow claim, deliberately. It does not read the case for you. It shortens the distance between an answer and a human who has confirmed it — which, in litigation, is the only place a time saving can honestly come from.