$ cat ~/blog/ai-deposition-summaries-guide.md
The Paralegal's Guide to AI Deposition Summaries (What's Safe, What Isn't)
AI can draft a deposition summary in seconds — and fail in specific, avoidable ways. A field guide to using it without getting burned, and keeping every cite honest.
An AI model will draft a deposition summary in about the time it takes to read this sentence. That’s genuinely useful. It’s also where careers get dinged, because the ways these summaries fail aren’t random — they’re predictable, and predictable means avoidable.
Here’s the field guide I’d hand a paralegal or a litigation-support team today.
What’s safe
- First-pass structure. Turning 200 pages into a topic outline, a witness timeline, or a list of exhibits referenced. The model is great at organizing what’s there.
- Surfacing, not deciding. “Find every place the witness discusses the maintenance log” is a search task with a high recall ceiling. Verify the hits; don’t trust the misses.
- Drafts a human will edit. Anything where a competent reviewer reads the output against the transcript before it leaves the building.
What isn’t
- Uncited claims. If the summary asserts something the witness said, it needs a page:line. No cite, no claim.
- Characterizations. “The witness was evasive” is argument, not record. Keep the model describing testimony, not editorializing about it.
- Silent inference. The dangerous failure isn’t a wrong fact — it’s a plausible fact the witness never said, written in confident prose. This is the one that gets quoted back to you.
The one rule that prevents most of it
Make the tool cite the record, every time, and admit when it can’t. If a line in the summary can’t be traced to a page:line in the transcript, it doesn’t ship. That single discipline turns “AI summary” from a liability into a time-saver.
It’s the entire idea behind CasePrompts: prompts built on the CITE method, where every output points to page:line — and if it’s not in the record, the AI says so instead of inferring. And because they’re vendor-neutral, the same prompt works whether your firm is on one model this quarter and a different one the next.
The professionals who produce the record are still the product here. AI just makes the first 80% faster — as long as you never let it do the last 20% unsupervised.
What’s the failure mode you’ve actually hit with an AI summary? I’m collecting the real ones.