Deposition Analysis Is a Testimony Problem, Not a Document Problem
Deposition analysis is a testimony problem, not a document problem
E-discovery and deposition analysis get spoken about together and are not the same work. Discovery is a search problem across a large corpus: find the responsive material. Deposition analysis is a comprehension problem across a small one: understand what a person said, under questioning, in sequence, and be able to prove it later.
Tools built for the first are routinely sold for the second, and the mismatch is where review time disappears.
What makes testimony different
It is adversarial by construction. A deposition is not a witness explaining events. It is a witness answering questions chosen by someone with an objective. The meaning of an answer depends on the question that produced it, and often on the three questions before that.
Meaning arrives late. Documents say what they say. Testimony qualifies itself — a flat denial at page 47 becomes “not to my knowledge” at page 51 and “I may have been told” at page 63. Any system that retrieves a matching passage and stops has captured the least reliable version.
Precision is the deliverable. In document review, “this email discusses the shipment delay” is a useful summary. In a deposition designation, the difference between “about three weeks” and “approximately three weeks” can be argued over, because someone will read the line aloud in court.
Every assertion is checkable. Testimony is addressed by page and line precisely so that any characterisation can be verified by anyone. That is a feature to design around, not an inconvenience.
The four questions worth asking of a transcript
Most useful deposition work reduces to four:
- What did the witness say about X? — the retrieval question, and the easy one.
- Did they contradict themselves, or a prior statement? — requires comparing passages that may be two hundred pages apart, or in a different deposition entirely.
- What did they refuse or fail to answer? — objections, instructions not to answer, non-responsive answers. Often the most valuable material in the transcript and the hardest to search for, because the useful signal is an absence.
- What can I actually use? — which passages survive objection, and in what form.
A tool that answers only the first is doing keyword search with better manners.
Where automation genuinely earns its place
Cross-transcript contradiction hunting. Three witnesses, four hundred pages each, one timeline. Locating every account of the same event across all three is mechanical work that a person does slowly and a machine does in seconds. The judgement about whether the accounts actually conflict remains yours.
Exhibit tracing. Following an exhibit through the testimony of everyone who was handed it, in order, is tedious and highly automatable.
Building the first draft of a designation chart — passages, citations, and the surrounding exchange — for a human to accept, cut or extend.
Finding the qualifiers. Searching for the language of hedging near a topic (“I think”, “approximately”, “to the best of my recollection”) surfaces the passages where testimony is softer than the summary suggests.
Where it does not
Deciding what matters. Judging whether a contradiction is real or an artefact of a badly worded question. Assessing whether a witness is evasive or simply confused. Each of these is the work, and none of it survives being delegated.
And the recurring rule: any output you cannot trace to a page and line is a lead, not a fact. If a system reports a contradiction and cannot show you both passages, you have been given something to check rather than something to use.
The test that matters when you evaluate software
Not accuracy on a demo. Ask instead:
- Does an answer come with a location you can open? Document, page, line.
- Do you get the surrounding exchange, or an isolated line? You need the question.
- Can it work across multiple transcripts at once? Single-document tools miss the contradictions, which are the point.
- What does it do when the answer is not there? A system that would rather compose something than report nothing is disqualifying for this work.
- Where does the transcript go? Depositions contain privileged and protected material. For some matters the only acceptable answer is that it stays on infrastructure you control.
The economics, honestly
Suppose a set of depositions takes thirty hours to review by hand. Good tooling can cut the locating to perhaps six. But every passage you intend to use still has to be opened and read in context before it goes in a chart — call that four more. Ten hours against thirty is a real saving and worth paying for.
The failure mode is a tool that summarises fluently without citations. The locating work returns in full, because nothing is usable until it is found and confirmed, and you have added the risk of something that reads well enough to be trusted unchecked.
How Lawnova PDF handles it
Lawnova PDF indexes transcripts with page and line precision and searches across all of them together, returning the surrounding exchange rather than a bare hit. Every AI answer carries [doc][page][line] references you can open, and it will run against a local airgapped model where the material cannot leave your environment.
If you want the working sequence rather than the tooling argument, we set it out separately in how to review a deposition transcript.