AI Deposition Prep: Outlines, Impeachment, and Summaries

AI does four deposition jobs well: it builds the question outline from the record, it finds contradictions across documents and prior testimony, it summarizes the transcript into takeaways you can act on, and it prepares your own witness by generating the cross they are about to face. It does not take the deposition. Nothing here replaces the lawyer in the chair reading the room.

I own 10 law firms, several of them litigation practices. Our lawyers take depositions constantly. The change over the last two years is not that AI got smarter about testimony. It’s that grounded tools arrived, meaning tools that answer only from documents you uploaded and cite the page. That’s the difference between a research toy and something you can build an outline on.

Before the deposition: load the file, then build the outline

Step one is not prompting. It’s loading. Put the complaint, the discovery responses, the deposition notices, the key documents, and any prior statements from this witness into one grounded workspace. NotebookLM is the easiest starting point because it answers only from what you uploaded and links each claim back to the source. A Claude Project works the same way if you prefer to stay in one tool. Here’s the full NotebookLM walkthrough if you have not set one up.

Then ask for the outline. Not “help me prepare for a deposition.” Something specific:

You are an experienced litigator preparing to depose a witness. Work only
from the documents I have uploaded. Do not use outside knowledge.

Case type: [e.g., breach of contract / lemon law / employment]
Witness: [name and role, e.g., service manager at the dealership]
My theory of the case: [two sentences]
Key disputed facts: [list 3 to 6]

Produce a deposition outline organized topic by topic, in the order I
should examine. For each topic:
1. A one-line statement of what I am trying to establish
2. 5 to 8 questions, foundational and background questions first, then
   the questions that lock in the admission
3. The exact documents I should have marked and ready for that topic,
   with the page or bates cite from my uploads
4. The likely evasion for that topic and one follow-up question that
   closes it off

Flag any topic where the uploaded record does not support the questions.

That last line matters more than the rest. Without it the model will happily invent a topic it has no support for. With it, you get an honest gap list, which is itself useful: those gaps are usually the documents you never got in discovery.

The impeachment hunt

This is where AI earns its keep. Cross-referencing a witness’s statements across a 900-page production, three sets of interrogatory answers, and a prior deposition is exactly the work humans do badly at hour six. The machine does not get tired at hour six.

Feed it everything the witness has ever said on the record, plus the documents that touch those statements, and ask for conflicts:

Work only from the uploaded documents. Compare every statement made by
[witness name] across all sources: prior deposition testimony, declarations,
interrogatory responses, emails, and internal records.

Output a table with these columns:
| Statement A (quote) | Source A (doc + page/line) | Statement B (quote) |
Source B (doc + page/line) | Why they conflict | Suggested question sequence |

Rules:
- Quote exactly. Never paraphrase inside the quote columns.
- If you cannot cite a page or line, leave the row out.
- Include date conflicts, quantity conflicts, and shifts in who knew what
  and when, even small ones.
- Rank rows by how damaging the conflict is to the witness's credibility.
- List separately any statement that is uncorroborated by any document.

At CPLG, our lemon law files run to hundreds of pages of repair orders, and the conflict that wins the case is usually a date: the service advisor swears the vehicle came in once for a defect, and the repair orders show four visits. That’s a table row, not a legal insight. Which is exactly why a machine should find it and a lawyer should use it.

Prepping your own witness

Flip the same tools around. Ask AI to generate the cross-examination opposing counsel will run on your client, using the documents in your own file. Your client then practices against real pressure instead of against your polite version of it.

Two other uses that people skip. First, plain-English explanations of the process for a nervous client: what a deposition is, why the court reporter pauses, why “I don’t know” is a complete answer, what an objection means and why they still have to answer most of the time. Second, a list of the three or four questions your client will most want to over-explain, so you can rehearse the short answer before the transcript records the long one.

After the deposition

The certified transcript is the official record. Nothing AI produces changes that, and no AI transcript is admissible in its place. See the transcription guide for where that line sits.

What AI does with the transcript once you have it:

1. Summarize into a page-and-line index of every meaningful admission.
2. Issue-code the testimony against your claim elements, so you can see which elements the witness just helped prove.
3. Generate the follow-up discovery list: every document, name, and date the witness mentioned that you do not already have.
4. Draft the “what did we actually get” memo for the file, including what you failed to lock down and would need a second session for.

Step three alone is worth the exercise. Witnesses name documents you never requested, and those names get buried in 200 pages of transcript.

Verification rules, non-negotiable

Every factual assertion in an AI outline gets checked against the actual document before you rely on it. Every one.

Never let AI produce a quote from testimony that you have not read yourself on the page. A fabricated quote used in impeachment is a career problem, not a workflow problem, and the lawyers who have been sanctioned over AI citations all thought they were being careful.

On confidentiality: do not upload transcripts or client documents to a consumer free tier, because free plans generally permit training on your inputs. Use a paid business or team plan with terms that say your data is not used for training, or keep the file inside your firm’s own Google or Microsoft tenant. Read the terms once, write down the answer, tell your associates.

Last rule: the outline is a checklist, not a script. Witnesses answer questions you did not ask. If you are reading from a page while a witness hands you the case, the tool cost you the deposition.

Do this with your next one

Take the file for the next deposition on your calendar. Load the complaint, the discovery responses, and the documents into NotebookLM tonight, run the outline prompt above, and compare what comes back to the outline you would have written yourself. If you want more prompts in this format, they’re in the prompt library. Ten minutes tells you whether this belongs in your practice.