Here is the workflow that holds up, where it fails, and how to choose between a general model and a deposition tool.
What does an AI deposition summary actually give you?
Three different products get called "a summary." Ask for them separately.
- Page-line digest. One line per topic with the page and line range. This is the deliverable that saves the most time.
- Topical summary. Testimony grouped by issue (liability, damages, notice, prior statements), each point tied to cites.
- Narrative summary. A readable account of the deposition. Useful for a partner or client update, and the most likely to smooth over a hedge or a correction.
A summary without page-line cites is not usable. If a tool cannot return cites you can check, treat the output as a list of questions to chase, not a digest.
How do you run it, step by step?
- Start from the certified transcript, not a rough. Under Federal Rule 30(e), when the deponent or a party asks before the deposition is completed, the deponent gets 30 days after being notified that the transcript is available to review it and sign a statement of changes with reasons. A digest built on a rough can cite testimony that was later corrected. Rule 30
- Keep the page and line numbers in the text you feed the model. Strip headers and footers, keep line numbers. If the PDF loses them on extraction, fix that first.
- Ask for a digest with quotes. Require each entry to include page:line and a short verbatim quote. The quote is what makes checking fast.
- Run a second pass for what summaries drop: objections, instructions not to answer, "I don't recall" runs, corrections made on the record, and exhibit references.
- Spot-check, then check everything you will cite. Open the transcript at every cite you will use in a motion, an outline or a client email and compare the quote to the page. See How to Verify AI Legal Citations for the same discipline applied to authority.
- Save the digest to the matter file with the model, the date and who verified it. If the digest is questioned, you can show how it was made.
Where do AI deposition summaries go wrong?
The failures are predictable and they cluster where lawyers care most.
- Wrong cites. Page or line off by a few, or a quote that blends two answers.
- Lost qualifiers. "I think it was Tuesday" becomes "It was Tuesday."
- Missed objections and instructions. The summary reports the answer and not the fight over the question.
- Speaker confusion in multi-party depositions, where counsel's statements get attributed to the witness.
- Smoothed contradictions. A model asked for a coherent summary tends to reconcile testimony that should stay inconsistent.
Make each one a checklist item in step 5. They are reasons to check, not reasons to skip the tool.
Should you use a general model or a deposition tool?
Pick by stakes and volume, not by brand.
| General model with your prompt | Deposition-specific tool or service | |
|---|---|---|
| Setup | You build the prompt and the checks | Built-in cites, templates, review interface |
| Control of output | Full | Vendor-defined |
| Data handling | Depends on your plan and settings; read the provider's current terms | Depends on the contract; ask the vendor |
| Best for | Occasional depositions, custom issue outlines | High volume, recurring formats, multi-deponent cases |
Whichever you pick, get in writing how transcripts are stored, whether they are used for training and how they are deleted. Transcripts hold client confidences, so your duties under Model Rule 1.6 and the supervision duties in Rule 5.3 reach the vendor too. ABA Formal Opinion 512 covers competence, confidentiality, supervision and fees for generative AI. For vendor choices see Best AI for Deposition Summaries. For a prompt-based setup see Claude Projects for Lawyers. For Teams recordings see Microsoft Copilot and Teams Deposition Transcript Summaries.
Can a summary be used in court?
Not as a substitute for the testimony. Depositions are used in court under Federal Rule 32, and the transcript is the record. Rule 32 A summary of voluminous material can be offered under Federal Rule of Evidence 1006 only if the underlying material is made available to the other parties for examination or copying and the rule's other conditions are met. FRE 1006 Treat an AI-generated digest as internal work product. State rules differ, and this page does not cover any one state.
What does this change about deposition prep?
The digest pays off most for the next deposition. Feed prior testimony into your outline and ask for inconsistencies by topic, with cites. It is the same workflow with a different question. See AI Deposition Analysis and AI Deposition Prep for Personal Injury. For photo exhibits see Claude Opus 4.7 Vision for Deposition Photos.
What should a firm put in writing?
A short protocol: certified transcripts only, cites required, every filed cite opened and checked by a named lawyer, vendor data terms reviewed. Start from the Law Firm AI Policy Template.
Sources
- Fed. R. Civ. P. 30 (depositions by oral examination, including 30(e))
- Fed. R. Civ. P. 32 (using depositions in court proceedings)
- Fed. R. Evid. 1006 (summaries to prove content)
- ABA Formal Opinion 512, Generative AI Tools (2024)
Questions and answers
Can AI summarize a deposition accurately?
It can produce a useful first-pass digest. It can also misquote, drop a qualifier or miss an objection. Accuracy depends on transcript quality, the prompt and the checking step. Verify every cite you plan to rely on against the certified transcript.
Is it ethical to use AI to summarize depositions?
Yes, with the same duties as any tool that touches client information: competence, confidentiality and supervision. ABA Formal Opinion 512 discusses Model Rules 1.1, 1.6 and 5.3, among others, in the context of generative AI. Check your own state bar's guidance, which can differ.
What should an AI deposition summary include?
Page and line cites, short verbatim quotes, topic headings, objections and instructions not to answer, corrections on the record, and exhibit references.
Can I upload a deposition transcript to ChatGPT or Claude?
Only if the plan's data terms, any client consent required and your firm policy allow it. Review how the provider stores and uses inputs first, because terms change.
Can the deponent change the transcript after the deposition?
Under Federal Rule 30(e), if the deponent or a party requests review before the deposition ends, the deponent has 30 days after notice that the transcript is available to review it and sign a statement of changes with reasons. State rules vary.
Operational information, not legal advice. Federal rules were checked against their published text on October 18, 2026; state rules differ and are not covered.