Use Cases
AI Deposition and Transcript Summaries: A Litigator's Workflow
AI deposition summaries work when every quotation is checked against the transcript, the tool runs on a business tier and output is anchored to page and line.
On this page
AI can turn a 300-page deposition transcript into a page-line summary, a topic index and a contradiction table in minutes, and independent testing shows it does this better than most lawyers. The catch is verification: every quotation must be checked against the transcript, and the work has to run on a business-tier tool that keeps the transcript out of training data and inside privilege.
What the benchmark data says about transcripts
Transcript work is one of the few litigation tasks where the evidence favours the machine. The Vals Legal AI Report of 27 February 2025 tested Harvey, CoCounsel, Vincent AI and Oliver against a lawyer control group across seven tasks. AI tools beat the lawyers on four: document extraction, document Q&A, summarisation and transcript analysis. Lawyers won on redlining and EDGAR research, and chronology generation was a tie.
| Task (Vals, Feb 2025) | Result |
|---|---|
| Document Q&A | Harvey 94.8%, CoCounsel 89.6%, lawyer baseline 70.1% |
| Transcript analysis | AI tools ahead of the lawyer baseline |
| Summarisation | AI tools ahead of the lawyer baseline |
| Chronology generation | Tie |
| Redlining | Lawyers 79.7%, ahead of every tool |
| EDGAR research | Lawyers 70.1% vs Oliver 55.2% |
Two details matter more than the headline. Vincent AI was noted for refusing to answer rather than hallucinating, which is the behaviour you want from a transcript tool. And Lexis+ AI withdrew from the study; vendor absence from a benchmark is itself information.
The reason transcripts suit the technology is structural. A deposition is a closed universe: every fact the summary may contain is in the document, the document is already text, and the page-line grid is a hard anchor that makes checking cheap. Open-ended legal research is the opposite: the model has to know the law rather than read it, and the hallucination problem returns.
The five outputs worth producing
A generic "summarise this deposition" prompt produces a generic summary. Litigators need five artefacts, each with its own verification method.
| Output | Format | What it is for | How you check it |
|---|---|---|---|
| Page-line digest | One entry per topic block: pages and lines, neutral paraphrase, key quotations verbatim | Motion practice, trial prep, expert briefing | Open every quoted passage at the cited page and line |
| Topic index | Alphabetical topics, each with the page-line ranges where the witness addressed it | Finding testimony fast during cross | Spot-check ten random entries against the transcript |
| Chronology | Dated events as the witness described them, with page-line source and confidence note | Merging with the documentary timeline | Compare against exhibits and other witnesses |
| Contradiction matrix | Pairs of statements (same witness, or across witnesses) that cannot both be true, each quoted with reference | Impeachment, credibility arguments | Read the surrounding lines of both statements; many dissolve in context |
| Exhibit cross-reference | Every exhibit mentioned, the pages where it was discussed, and what the witness said about authenticity and contents | Foundation, admissibility, exhibit lists | Reconcile against the exhibit log |
Vendors have built products around this list. Legora bought Wexler, a litigation fact-analysis tool, on 29 July 2026; Harvey II (18 August 2026) added matter-aware Spaces that carry witness knowledge across sessions; Thomson Reuters' rebuilt CoCounsel Legal (US general availability 20 August 2026) advertises Tabular Analysis across 10,000 documents and 100 questions. Anthropic's open-source Claude for Legal litigation plugin tags any citation not checked against a connected source as [verify], a convention worth copying in your own prompts. The Legora review covers what the fact-analysis features cost.
Context windows and the 300-page problem
A 25-line transcript page runs to roughly 250 words, so a 300-page deposition is about 75,000 words. Anthropic's rule of thumb is that one million tokens is about 555,000 words, which puts the transcript near 135,000 tokens. As of September 2026, Claude Opus 5 and Sonnet 5, Gemini 3.1 Pro and GPT-6 Astra all offer a context window of about one million tokens, so a single deposition fits with room for the exhibits. Ten depositions do not fit at once, and forcing them in degrades the output long before the hard limit.
The working answer is two passes. Pass one summarises each witness separately into the page-line digest. Pass two receives only the digests, not the transcripts, and works across witnesses. The digests are a tenth of the source length and already carry the references, so the cross-witness pass stays inside the window and every flag can be traced back.
Three practical points follow.
- Feed text, not images. Scanned exhibits and older transcripts arrive as image PDFs; run a proper OCR pass first rather than relying on the model's vision, since a 2026 comparison reported that Claude is more likely to say "illegible" than to invent a number, while ChatGPT's OCR handled poor scans better (CounselPro, 2026). The PDF extraction playbook covers the tooling.
- Keep the line numbering. Never reflow the transcript into prose; the page-line grid is what makes the output checkable.
- Use a matter workspace. A Claude Project, a Gemini Notebook or a platform matter space keeps the digests, the exhibit log and the pleadings together. Gemini Notebook (renamed from NotebookLM on 16 July 2026) allows 50 sources per notebook free and 300 paid, works only on what you upload, and belongs in a Workspace or enterprise account, not a personal one, as Ernest Svenson's case-file workflow stresses; the Gemini Notebook review covers its limits.
The prompt set
The prompts below follow the five-part structure (role, task, sources, format, constraints) that dominates 2026 legal prompt guides. The neutrality constraint is not decorative. Stanford HAI's 2026 AI Index reported that when a false premise is framed as the user's own belief, model accuracy collapses (GPT-4o fell from 98.2% to 64.4% on the relevant test), so a prompt that says "the witness admitted she saw the email" will produce a summary that says so (Stanford HAI, 2026). Ask what the witness said, never what you hope was said.
ROLE: You are a litigation associate preparing a deposition digest for the trial team.
TASK: Produce a page-line digest of the attached deposition transcript of [WITNESS], taken on [DATE] in [CASE].
SOURCES: Use only the attached transcript. Do not use outside knowledge of the case, the parties or the law. If a passage is illegible or missing, write [ILLEGIBLE p.X] rather than guessing.
FORMAT:
1. Topic blocks in transcript order. Each block: heading; page:line range; a neutral paraphrase of no more than 80 words; up to three verbatim quotations, each followed by (p.X:L.Y).
2. Topic index: alphabetical list of topics with every page:line range where the witness addressed them.
3. Exhibit table: exhibit number; pages discussed; what the witness said about authenticity; what the witness said about contents.
4. Open items: questions the witness was instructed not to answer, questions taken under advisement, and passages where the witness said "I don't know" or "I don't recall" on a material point.
CONSTRAINTS:
- Quote exactly; never paraphrase inside quotation marks.
- Do not characterise testimony as an admission, concession or contradiction; report what was said.
- Do not infer facts not stated in the transcript.
- Flag every quotation you could not locate with certainty as [verify].
The cross-witness pass takes the digests, not the transcripts:
ROLE: You are a litigation associate looking for inconsistencies between witnesses.
TASK: Compare the attached page-line digests of [WITNESS A], [WITNESS B] and [WITNESS C] on the topics listed in the attached topic index.
SOURCES: Only the attached digests. Do not use the underlying transcripts or any outside information.
FORMAT: A table with columns: topic; witness A statement (verbatim, with p:L); witness B or C statement (verbatim, with p:L); why the statements cannot both be true, in one sentence; strength (direct conflict / tension / different recollection).
CONSTRAINTS: Report only pairs where the quoted words conflict. Do not rate credibility. Do not suggest which witness is right. Mark every pair for human confirmation in the transcripts.
Expect over-inclusion. Redgrave LLP's 2026 study of generative document review found higher recall than technology-assisted review but lower precision, and the same trade-off shows up in contradiction matrices: the model flags pairs a lawyer would not bother with. That is the right failure mode; a missed contradiction costs more than a spurious one.
Privilege, work product and where the transcript goes
A deposition transcript is usually not itself privileged; the other side has a copy. The digest, the contradiction matrix and the strategy notes are work product, and the question is whether an AI tool changes that.
Two 2026 rulings frame the answer. In United States v. Heppner (S.D.N.Y., bench ruling 10 February 2026, opinion 17 February 2026), Judge Rakoff held that research memos a defendant had generated on consumer Claude were neither privileged nor work product: "Claude is not an attorney", and the consumer terms allowed logging and disclosure, so privilege never attached in the first place. The court left open that privilege might attach where counsel directs the AI use, on the Kovel model. In Warner v. Gilbarco (E.D. Mich., February 2026), reported so far by a single vendor blog, the court reportedly treated AI programs as "tools, not persons" and protected AI-assisted work product. Read together, the two cases say: lawyer-directed use on a tool whose terms keep the material confidential is defensible; a lawyer's client running the same transcript through a free chatbot is not. The privilege and AI page works through the case law, and the waiver question page gives the short answer.
The tier matters because logs are discoverable: in the OpenAI copyright litigation (S.D.N.Y.), Magistrate Judge Wang ordered 20 million de-identified ChatGPT logs produced in November 2025, and Judge Stein affirmed on 5 January 2026. As of September 2026, Anthropic's Team, Enterprise and API tiers are never used for training; OpenAI's Business, Enterprise and API tiers do not train by default but retain data 30 days for abuse monitoring; consumer tiers on both train unless the user opts out. The confidentiality checklist sets out which tier to use and what to switch off.
One further habit protects the work product. Duane Morris's February 2026 alert on transcription tools recommends adding the lawyer's own impressions to any AI-generated record so that opinion work-product protection has something to attach to; a digest annotated with "significance for the limitation defence" is stronger than a bare digest. Recording a witness interview rather than a deposition is a different problem: the New York City Bar's Formal Opinion 2026-2 (5 August 2026) requires the consent of all parties to a non-client conversation and treats not recording as the default, as the meeting notes page explains.
The QC step nobody skips twice
Summaries hallucinate less than research does, but not zero. Vectara's summarisation hallucination leaderboard put the best models at roughly 1.8% to 3.3% in May 2026 (leaderboard). On a 150-statement digest, that is three to five unsupported ones, and you do not know which. Damien Charlotin's AI Hallucination Cases database counted 548 false quotations among 2,041 decisions as of 14 September 2026; a fabricated witness quotation in a motion is exactly that category.
The check is mechanical and takes an associate about an hour per transcript.
- Every quotation: open the page and line, confirm the words match character for character, including ellipses.
- Every paraphrase: read the cited range; confirm nothing has been added and no qualifier ("I think", "as far as I recall") has been dropped.
- Every "I don't recall": confirm the model has not converted a non-answer into a denial or an admission.
- Every contradiction pair: read ten lines either side of each statement; strike pairs that dissolve in context.
- Every exhibit reference: reconcile against the exhibit log; models transpose exhibit numbers.
- Sign and date the digest, listing the checker, the tool and the tier. Over 30 of the 94 US federal district courts had AI disclosure or certification requirements as of April 2026; once a quotation reaches a motion, the court's standing order applies.
Public defenders have arrived at the same discipline from the other direction: the NACDL's August 2026 white paper argued that defenders have a duty to use AI because prosecutors already do, and Kentucky public defenders use JusticeText to review body-camera footage (Minnesota Lawyer, 4 August 2026). Video carries the same verification duty, plus a check of the transcription against the recording.
A setup that works
For a solo litigator or a small team: a Claude Team or ChatGPT Business seat, one project per matter, the two prompts above saved as a template, and a rule that no digest leaves the project unchecked. For a litigation department: a platform with matter-level memory and tabular review, the same prompts adapted to its workflow builder, and the check built in as an approval gate before export. Module 2 of the AI Lab for Lawyers course builds the page-line digest and the cross-witness pass live on a sample transcript. The litigators' practice page places transcript work alongside e-discovery and brief drafting; the red-teaming playbook is the natural next step once the contradiction matrix exists.
This week: pick one closed deposition from a finished matter, run the digest prompt, and time the QC step. If the check takes longer than producing the digest by hand would have, the prompt needs tightening; if it takes an hour and finds two errors, you have your workflow.
Frequently asked questions
Can AI summarise a deposition transcript accurately?
Largely yes, with a verification step. Vals AI's February 2025 legal benchmark found AI tools ahead of a lawyer control group on transcript analysis and summarisation. But the best summarisation models still produce unsupported statements at a rate of roughly two to three per hundred, so every quotation and every page-line reference has to be checked against the transcript before the summary goes into a motion or an outline.
Is an AI-generated deposition summary protected as work product?
It can be, if a lawyer directs the work and the tool's terms keep the material confidential. A February 2026 Michigan ruling reportedly treated AI programs as tools rather than persons and protected AI-assisted work product, while Judge Rakoff refused privilege for memos a defendant made on consumer Claude because the terms allowed logging and disclosure. Use a business tier and add the lawyer's own impressions.
Which AI tool is best for deposition summaries?
For a single transcript, any frontier model on a business tier with a one-million-token context works: Claude Opus 5 or Sonnet 5, Gemini 3.1 Pro or GPT-6 Astra as of September 2026. For a whole case file, litigation platforms such as Legora, Harvey, CoCounsel Legal or Everlaw add matter-level memory, tabular review across witnesses and a citation trail. The tool matters less than the page-line anchoring and the quotation check.
How do I find contradictions between witnesses with AI?
Summarise each transcript separately into a topic-indexed digest with page-line references, then run a cross-witness pass that receives only those digests and asks for statements on the same topic that cannot both be true, quoting each side with its reference. Confirm every flagged pair in the transcripts themselves; models over-include and some flagged contradictions dissolve on reading the surrounding lines.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Practice Areas
AI for Litigators: Research, Drafting, Discovery and the Sanctions Trap
AI for litigators pays off in chronologies, discovery and deposition prep; it costs $110,204 when a brief cites invented cases. Where to gain and the one rule.
8 min read
Regulation & Ethics
Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots
Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.
10 min read
Use Cases
AI in E-Discovery: From TAR to Generative Review
Generative AI review beats TAR on recall (87–88% vs 64% in Redgrave's 2026 study) but loses precision. Tools, privilege, defensibility and a sampling plan.
11 min read
Tools
Gemini Notebook (NotebookLM) for Lawyers: A Grounded Case-File Assistant
Gemini Notebook (ex-NotebookLM) for lawyers as of September 2026. How source-grounded answers work on a case file, three workflows, source caps and data terms.
6 min read
Use Cases
Drafting Briefs and Motions with AI Without Ending Up in Charlotin's Database
AI brief drafting is safe only as a five-stage workflow: grounded research, outline, drafting from verified authorities, every case read, court rules checked.
10 min read
Questions
Does Using ChatGPT Waive Attorney–Client Privilege?
Using ChatGPT rarely waives privilege in the technical sense; the courts' answer is worse. In Heppner, consumer-chatbot memos were never privileged at all.
6 min read