$ cat ~/blog/ai-deposition-summaries-guide.md
The Paralegal's Guide to AI Deposition Summaries (What's Safe, What Isn't)
A practical workflow for drafting and reviewing AI deposition summaries with page-line citations, failure checks, and accountable human sign-off.
Story overview 6 main sections
Visual story map
How this long read moves
- 01 Section 1 of 6: Pick the summary before picking the tool
- 02 Section 2 of 6: Prepare the source
- 04 Section 4 of 6: A synthetic example
- 05 Section 5 of 6: Predictable failure modes
- 06 Section 6 of 6: Reviewer checklist
An AI model can turn a long deposition into a tidy summary in minutes. That speed is useful, but it creates a dangerous temptation: reviewing the prose instead of reviewing the work.
An AI summary is a draft generated from a source. It is not the record, it is not self-verifying, and it is not ready merely because it sounds careful. The safe workflow makes every material statement traceable to the transcript and gives a named human reviewer the final decision.
This is an operational guide, not legal or ethics advice. Firm policy, client instructions, protective orders, professional rules, and the requirements of your jurisdiction control. Lawyers should consult the rules and guidance that apply to their work; ABA Formal Opinion 512 discusses competence, confidentiality, supervision, candor, communication, and fees when lawyers use generative AI.
Pick the summary before picking the tool
“Summarize this deposition” is underspecified. Decide what the reader needs:
| Summary type | Useful for | Required structure |
|---|---|---|
| Chronological | Reconstructing a sequence of events | Date or period, event, witness testimony, page:line |
| Topical | Evaluating allegations, defenses, or subject areas | Topic headings, testimony on each side, page:line |
| Narrative | Fast orientation for a case team | Neutral prose with citations after every factual sentence |
| Testimony index | Returning to the source quickly | Issue, short description, page:line range |
| Exhibit index | Tracking documents discussed in testimony | Exhibit number, description, authentication/use, page:line |
A topic outline and a witness-credibility assessment are not the same assignment. Keep factual extraction separate from advocacy and characterization.
Prepare the source
Before a transcript enters any AI workflow:
- Confirm that the file is complete, searchable, and has stable page and line numbering.
- Keep the original transcript unchanged and work from an authorized copy.
- Check the protective order, client instructions, and your organization’s approved-tool policy.
- Determine whether the tool stores prompts, retains files, permits provider access, or uses inputs for training.
- When policy permits minimization, create a controlled redacted derivative rather than altering the source. Preserve the original page and line numbering, retain enough surrounding context to review each statement honestly, and document every omission or redaction.
- Tell the reviewer which transcript version was used.
Confidentiality is not solved by a checkbox that says “enterprise.” Understand the actual data flow. If the task can run locally, on-device legal AI can reduce exposure, but local processing does not replace access controls, retention rules, or professional judgment.
A reviewable workflow
1. Write the assignment like an assignment memo
Specify the source, scope, output type, neutral tone, citation format, prohibited inferences, and what to say when the record is silent. The CITE method is a reusable structure: context, instructions, template, and evidence.
At minimum, require:
- Use only the supplied transcript.
- Cite every factual statement with page and line.
- Quote before characterizing important testimony.
- Do not combine testimony from different speakers.
- Mark unclear or conflicting passages for review.
- Say “not found in the record” instead of inferring.
2. Generate in sections
Long documents are easier to audit when the work is bounded. Process a defined page range or topic, preserve the same template, then reconcile the sections. A single enormous prompt can silently omit the middle of a transcript even when the output looks complete.
3. Check the citations before polishing the prose
For each material statement:
- Open the cited page and line.
- Confirm the correct witness is speaking.
- Read enough surrounding testimony to catch qualifications, objections, and corrections.
- Confirm that the summary is no broader than the cited passage.
- Mark the statement verified, revise it, or remove it.
Citation presence is not citation accuracy. A fabricated page-line reference and a real citation that does not support the sentence are both failures.
4. Run a second pass for omissions
Ask what the summary left out, not only what it got wrong. Compare the topic list, exhibit references, and key examination sections against the source. Search important names and terms independently. High recall is a human review objective, not something fluent prose proves.
5. Record human sign-off
The reviewer should be identifiable, qualified for the task, and able to explain what was checked. Save the approved summary separately from the raw AI output. Record the source version, tool and model if relevant, prompt or template version, review date, reviewer, and disposition of flagged items.
A synthetic example
Imagine this fictional transcript excerpt:
47:12 Q. When did you first inspect the stairway?
47:14 A. I don’t remember the exact day. It was after the first complaint, maybe that week.
47:17 Q. Did you write an inspection report?
47:19 A. Not that I recall.
An unsafe summary says:
The witness inspected the stairway during the week of the first complaint and failed to prepare a report.
That sentence turns uncertainty into a date and “not that I recall” into a definitive failure.
A supportable version says:
The witness could not identify the exact inspection date, stating it was after the first complaint and possibly during that week (47:12–16). The witness did not recall preparing an inspection report (47:17–19).
The difference is not style. It is whether the summary preserves the limits of the testimony.
Predictable failure modes
| Failure | What it looks like | Review response |
|---|---|---|
| Unsupported fact | Plausible statement with no source | Remove unless the transcript supports it |
| Citation mismatch | Real page-line that discusses something else | Re-cite or rewrite after reading the passage |
| Lost qualification | “May,” “usually,” or “I don’t recall” becomes certainty | Restore the witness’s limitation |
| Speaker merger | Counsel’s question becomes witness testimony | Attribute each statement correctly |
| Omission | A contrary answer or later correction disappears | Compare the full examination and errata |
| Characterization | “Evasive,” “admitted,” or “contradicted” appears as fact | Separate record summary from attorney analysis |
| Exhibit confusion | Exhibit description, number, or use is wrong | Verify the exhibit index and cited discussion |
| Truncation | Summary covers the beginning and end but skips a range | Reconcile against page ranges and topic checklist |
The NIST Generative AI Profile treats generative-AI risk as something organizations should govern, map, measure, and manage across the system lifecycle. For a litigation team, that means a repeatable review process—not hoping the next model is better.
Reviewer checklist
- The approved source transcript and version are identified.
- The tool and data handling comply with applicable policy and instructions.
- The requested summary type and scope are explicit.
- Every material factual sentence has a verified page-line citation.
- Quotations, speakers, exhibits, and page ranges match the source.
- Qualifications, objections, corrections, and uncertainty remain visible.
- Important topics and examination sections were checked for omissions.
- Advocacy or credibility analysis is labeled and kept separate from factual summary.
- All flagged passages were resolved or left clearly flagged.
- A named human reviewer approved the final work product.
For the shorter office-level controls, use the AI Reliability Checklist for Law Firms. For the page-line discipline behind this workflow, read Page:Line or It Didn’t Happen and explore the prompts in the CasePrompts case study.