Inside the New Deposition Prep: How AI Is Changing the Way Litigators Build a Case
It's Sunday night, forty-eight hours before a deposition, and a litigator is staring at three banker's boxes and a shared drive stuffed with documents. The witness has been deposed twice before in unrelated matters. There are two prior sworn statements, a stack of medical records, and an interrogatory response that contradicts an email nobody has read yet.
In the old workflow, an associate would spend the weekend building a chronology and hoping to catch the contradiction. More and more, that first pass now happens inside a piece of software before anyone opens a binder.
Generative AI has moved from novelty demo to daily tool for a lot of litigators, and depositions are where you can see the shift most clearly. The interesting question isn't whether to use it. It's which parts of the workflow you hand over, and which you keep in your own hands. Every firm now faces a set of concrete choices about how far to push the technology and where to hold the line.
Decide What the Machine Reads First
The first real decision is which documents you feed to an AI system before you touch them yourself. Practitioner guides now walk through using AI to review pleadings, interrogatory responses, prior deposition transcripts, and witness statements as the opening move of prep, not the closing one. A Thomson Reuters guide lays out that sequence in detail, and it reflects how a lot of teams already work.
The trade-off is subtle. Reading the record yourself first builds intuition about the witness: how they talk, what they dodge, where the story bends. Letting the machine go first saves a weekend and surfaces cross-transcript contradictions you'd probably miss on your own.
Most experienced litigators land somewhere in the middle. They skim the key exhibits themselves, then push the full corpus through a summarizer and a search index. The order matters more than people admit.
Choose Between a General Model and a Legal-Grade One
Working inside a browser tab with a consumer chatbot is fast and cheap. It is also the wrong tool for privileged material. The choice between a general-purpose model and a purpose-built legal platform comes down to three things:
- Confidentiality. Consumer tools may retain prompts or use them for training. Legal-grade platforms run in closed environments with contractual data controls, which is usually what your engagement letter and your bar rules assume.
- Grounding. A general model will happily invent a case cite. A retrieval-based legal tool answers from documents you loaded, with a pointer back to the source page. That's the difference between a summary you can verify and one you have to believe.
- Workflow fit. Deposition-specific tools produce chronologies, exhibit indexes, and cross-transcript comparisons in formats a litigation team already uses. General models produce prose you then have to reformat.
For anyone weighing how deeply to build AI into a practice, the broader research from research from Law.co on AI in legal work in legal work is a useful frame. It reads the technology less as a plug-in and more as infrastructure the firm has to design around.
Decide How Far to Push the Drafting
Outlines are the natural entry point. Feed the model the complaint, the answer, key exhibits, and prior testimony, then ask for a topic-by-topic outline with proposed lines of inquiry. What you do with the output is the decision.
Some litigators use it as a first draft they mark up heavily. Others treat it as a checklist they compare against their own outline, so they can see what the machine flagged that they missed, and what they thought of that it didn't. A third camp uses AI as a sparring partner, running mock objections and anticipated answers before ever sitting across from the witness.
A recent Mayer Brown analysis walks through eight practical uses across the litigation lifecycle, including that mooting role. The common thread: the lawyer's judgment is the product, and the draft is scaffolding.
Set a Verification Rule and Follow It
Every piece of AI-generated output that touches a filing or a witness needs a verification step, and the standard has to be explicit. Not "we'll double-check anything that looks off." A written standard: quotes checked against the transcript, cites checked against the reporter, exhibits checked against the original.
Judges have started codifying this. Some now require lawyers to certify whether generative AI was used in a filing and to confirm that any AI-generated content was verified. There is no substitute for reading the case yourself before you cite it.
The Lawyer Still Has to Be in the Room
AI is good at pattern recognition across text. It is not good at reading a witness who pauses a half-second too long before answering, or deciding when to burn a document on cross versus save it for trial.
It doesn't know your judge. It doesn't know which opposing counsel will fold on a motion in limine and which will fight to the last breath. Those calls are the job.
The litigators getting the most out of these tools treat them like a very fast, very literal junior associate: fluent, tireless, occasionally wrong, and never the one signing the pleading. Make the decisions above deliberately and the technology earns its place. Skip them and it becomes one more thing to manage on the Sunday night before the deposition.
