Which Parts of Deposition Prep Can AI Actually Help With?
AI earns its place in the tasks that are high-volume and verifiable, not the tasks that are strategic and judgment-driven. Whether you use a general-purpose model or dedicated deposition prep software, the high-value use cases for a California deposition are the same:
- Document organization and exhibit indexing. Categorizing produced documents by topic, witness, and date, and building an exhibit list keyed to your examination topics.
- Chronology building. Extracting dates and events from the record into a timeline you can audit against source documents.
- First-draft outline generation. Producing a topic tree and candidate question sets that you then re-sequence, cut, and sharpen.
- Transcript and testimony summarization. Condensing prior depositions, declarations, and discovery responses into topic-organized digests.
- Inconsistency detection. Comparing a witness's statements across sources and flagging candidate impeachment material for attorney confirmation.
Every one of these is a drafting or organizing aid. None replaces what the State Bar guidance calls the active exercise of professional judgment. The guidance is explicit: a lawyer remains fully responsible for any outputs and work product generated with AI assistance, and the more autonomy a tool has, the more supervision and verification it requires. These are examining-attorney workflows, but AI witness preparation for your own client raises the same confidentiality analysis discussed below.
The scope of your prep, AI-assisted or not, is still framed by statute. You may depose any person, including a party or an entity (Code Civ. Proc. § 2025.010), and discovery reaches any non-privileged matter relevant to the subject matter of the action that is admissible or reasonably calculated to lead to admissible evidence (Code Civ. Proc. § 2017.010). Those two provisions define the universe your outline should cover.
How Do You Build a Deposition Outline Using AI?
An AI deposition outline works best as a three-pass process: scope, generate, refine. The attorney owns passes one and three.
Pass 1: Define the Scope Before You Prompt
Feed the tool the boundaries, not just the documents. For a percipient witness, that means your theory of the case and the relevance framework of Code Civ. Proc. § 2017.010. For an entity witness, the deposition notice controls: the notice must describe the examination matters with reasonable particularity, and the entity must then produce the persons most qualified to testify on those matters to the extent of information known or reasonably available to it (Code Civ. Proc. § 2025.230). Your outline for a PMK deposition should track the noticed topics one-for-one, and AI is well suited to mapping documents and prior testimony onto each noticed matter.
Build the time budget into the outline structure. Except for enumerated exceptions, examination by all counsel other than the witness's counsel of record is limited to seven hours of total testimony, with additional time available by court order where needed to fairly examine the deponent (Code Civ. Proc. § 2025.290). Expert witnesses, PMK deponents designated under section 2025.230, and employment cases fall outside the default limit. If your witness is subject to the seven-hour cap, have AI estimate time per topic block and force yourself to rank topics by priority.
Pass 2: Generate the Topic Tree and Candidate Questions
Ask the tool for a hierarchical topic outline first, then candidate questions within each topic. Request foundational and authentication questions for each key exhibit separately. AI tools for lawyers are reliably good at breadth here: they will surface background, timeline, document-by-document, and damages topics you might sequence differently but should not skip. If you want a repeatable structure for these prompts, the DRAFT method for legal AI prompting provides a framework that transfers directly to outline generation.
Pass 3: Attorney Re-Sequencing and the Rifkind Boundary
The attorney sequences the outline, decides where to box the witness in before showing the document, and builds in the traps. AI does not know your read of the witness or which admission matters most to your summary judgment motion.
One doctrinal boundary belongs in every California outline review. Under Rifkind v. Superior Court, legal contention questions – "state all facts, identify all witnesses, and identify all documents supporting your affirmative defenses" – are improper when posed on the spot at an oral deposition, because they force the deponent to perform a sophisticated exercise of legal reasoning live. (Rifkind v. Superior Court (1994) 22 Cal.App.4th 1255.) That work belongs in written contention interrogatories. If your AI-drafted outline includes contention-style questions aimed at a represented party, expect an instruction not to answer and a Rifkind cite. Strip them out and propound them in writing instead.
Legion's discovery workflows handle the written side of this split: contention interrogatories, propounding and responding discovery, and separate statements, all formatted on California pleading paper. If a deponent's counsel properly deflects a contention question to written discovery, the follow-up interrogatories can go out the same week.
How Can AI Analyze Deposition Transcripts and Find Impeachment Material?
Deposition transcript analysis with AI runs in two modes: compression and comparison.
Compression turns a 300-page prior transcript into a topic-organized digest with page-line citations, so you can find every passage where the witness discussed the contract term at issue. An AI deposition summary of this kind is a legitimate starting point, and the same AI legal document summarization approach works for declarations and verified discovery responses. For multi-witness cases it is often the only realistic way to get through the record before the deposition date.
Comparison is where the impeachment value lives. Feed the tool the witness's prior deposition, declaration, verified discovery responses, and relevant emails, and ask it to flag every statement that conflicts across sources. The output is a list of candidate inconsistencies, each of which the attorney then confirms against the actual transcript.
That confirmation step is mandatory, not optional. AI tools fabricate page-line citations the same way they fabricate case citations. The lesson of Noland applies with full force: counsel there admitted relying on AI and failing to manually verify quotations, and the court found nearly all of the quotations in the appellate briefing were fabricated, with some cited cases nonexistent entirely. The court held that relying on fabricated legal authority is sanctionable and violated a basic duty counsel owed to his client and the court – duties of candor and competence – found the appeal frivolous, imposed a $10,000 sanction, ordered the opinion served on the client, and referred counsel to the State Bar. A deposition outline citing invented transcript lines will not draw a published opinion, but it will collapse in the room when opposing counsel checks the passage, and any resulting motion practice built on the phantom cite carries the same candor exposure.
Confidentiality is the second constraint. Transcripts routinely contain personal identifying information, medical detail, and third-party confidential material. The State Bar's Practical Guidance (originally issued November 16, 2023; revised 2026) instructs that a lawyer must not input confidential client information into a generative AI solution that may present material risks to confidentiality or security, absent informed client consent. The Judicial Council's own rule for court staff points the same direction: Cal. Rules of Court, rule 10.430, adopted effective September 1, 2025, requires court AI policies to prohibit entering confidential, personal identifying, or other nonpublic information into public generative AI systems. Treat that as the emerging floor for your own transcript work.
How Do I Anticipate Objections and Opposing Testimony With AI?
AI is a useful sparring partner for objection strategy, with one large caveat: general-purpose tools routinely over-generalize federal practice or lay concepts of "objection." Ground every AI-generated objection plan in Code Civ. Proc. § 2025.460, which draws a sharp waiver line:
| Objection type |
Rule at deposition |
Authority |
| Privilege / work product |
Waived unless a specific objection is timely made during the deposition |
Code Civ. Proc. § 2025.460 |
| Form of question or answer (and other curable irregularities) |
Waived unless a specific objection is timely made during the deposition |
Code Civ. Proc. § 2025.460 |
| Competency of the deponent; relevance, materiality, admissibility at trial |
Not waived; objection unnecessary at the deposition |
Code Civ. Proc. § 2025.460 |
Use AI to run drills against this framework. Have it review your outline and flag questions likely to draw form objections, then draft cleaner rephrasings you can deploy without breaking rhythm. Have it generate the objections you expect defending counsel to make on your key admission sequences, so you have a second and third path to the same answer. If you are defending, have it audit the noticed topics for privilege exposure so your objections are specific and timely rather than reflexive. For the broader taxonomy of proper and improper objections in California written discovery, see our California discovery objections cheat sheet.
Also verify the mechanics AI will confidently get wrong. Notice timing is statutory: an oral deposition must be scheduled at least 10 days after service of the notice, at least 20 days after issuance of a subpoena for consumer personal records or employee employment records, and at least 5 days in unlawful detainer actions (Code Civ. Proc. § 2025.270). Do not let a tool trained mostly on federal materials talk you into the wrong clock.
Finally, plan for evasion. If the deponent refuses to answer or gives non-answers on questions you have carefully built, the enforcement path runs through a motion to compel answers. That process, including its timing traps and separate statement requirements, is covered in detail in Compelling Deposition Responses Under California Law. Building your outline with that motion in mind – clean questions, clear record, specific instructions not to answer captured verbatim – makes the eventual motion far easier to win.
What Does the California State Bar's AI Guidance Require for Deposition Prep?
The State Bar's Practical Guidance organizes the analysis around four duties, and this is the framework that governs all generative AI for litigation work in the state. All four duties apply to deposition prep even though nothing gets filed.
Confidentiality: Business and Professions Code § 6068(e) and Rule 1.6
The statutory duty is absolute in phrasing: an attorney must maintain inviolate the confidence, and at every peril to himself or herself preserve the secrets, of the client (Bus. & Prof. Code § 6068). The guidance applies it to AI inputs directly: prompts and uploaded documents can be used to train models or shared with third parties, and even tools that do not train on inputs may lack adequate security. Deposition prep materials – strategy outlines, witness vulnerability assessments, privileged communications – sit at the top of the sensitivity scale. Legal AI confidentiality starts with the tool's terms: use enterprise tools with contractual no-training and no-retention terms, or anonymize before prompting. The guidance also recommends consulting IT or cybersecurity professionals to confirm a tool is secure.
Free and consumer AI tools are dangerous for legal work. Free tiers of general-purpose chatbots train on user inputs by default, which means pasting a transcript excerpt or case fact pattern into one can breach the confidentiality duty under Bus. & Prof. Code § 6068(e) and Rule 1.6. These tools also fabricate citations, creating the sanctions exposure realized in Noland v. Land of the Free, L.P. (2025) 114 Cal.App.5th 426. Never use a free or consumer AI tool for client work, filings, research, or anything touching confidential facts. Their only acceptable use is non-confidential language tasks, such as tightening generic prose that contains no case information.
Competence: Rule 1.1 and the Two-Part Test
The guidance frames competence as two duties. First, baseline technological competence: a reasonable understanding of the tool's capabilities, data sources, limitations, and material risks before using it in legal services. Second, independent professional judgment: reviewing, verifying, and correcting AI outputs. The California Lawyers Association's summary puts the point sharply: it is not enough to detect and eliminate false AI results; the attorney must engage in the active practice of law, and the duty of professional judgment is non-delegable.
Candor: Rule 3.3 and the Verification Duty
The duty of candor cannot be delegated to AI, and every citation to authority must be verified before submission to a court. Deposition prep feeds motion practice constantly – motions to compel, protective orders, sanctions motions – so any fabricated authority or transcript cite in your prep materials will eventually surface in a filing. The Noland panel decision itself illustrates the consequences: the court found that 21 of 23 case quotations in the opening brief were fabricated and imposed the $10,000 sanction on the responsible attorney, Amir Mostafavi. Proposed amendments to the Rules of Professional Conduct would make the verification duty explicit in a new comment to Rule 3.3, covering fabricated, misstated, or out-of-context authorities including AI-generated ones. Those amendments remain proposals, not adopted rules, but they show the direction of travel.
Communication and Billing: Rule 1.4
Rule 1.4 requires keeping the client reasonably informed about significant developments in the representation. Whether AI use requires disclosure is context-dependent; the cleanest practice is addressing it in the engagement agreement, and our guide to adopting AI safely in a small firm covers policy language and client disclosure in detail. On fees, the guidance is concrete: you may charge for time actually spent refining prompts and reviewing outputs, but you must not charge for the time AI saved you.
Legion is built around these duties rather than in tension with them. Client materials are never used to train AI models, every draft is delivered for attorney review and signature, and each document carries a time-saved estimate with the equivalent hourly cost, which maps directly onto the guidance's bill-for-time-spent rule. Usage-based pricing broken down by matter lets the cost pass through as a case cost.
What Belongs on Your California AI Deposition Preparation Checklist?
Ethics before you prompt
- Confirm the tool has adequate security and no-training / no-retention terms (Rule 1.6; Bus. & Prof. Code § 6068(e))
- Anonymize client and witness identifiers if the tool is not enterprise-secured
- Check your firm's AI policy and the local superior court's AI policy (courts must adopt policies under Rule 10.430 by December 15, 2025)
- Confirm you understand this specific tool's limits and hallucination risk (Rule 1.1)
Deposition mechanics
- Notice timing correct: 10 days standard; 20 days for consumer or employment records subpoenas; 5 days in unlawful detainer (Code Civ. Proc. § 2025.270)
- PMK topics described with reasonable particularity, outline mapped to each topic (Code Civ. Proc. § 2025.230)
- Time budget built to the seven-hour limit and its exceptions (Code Civ. Proc. § 2025.290)
- Objection plan tracks the section 2025.460 waiver lines: form and privilege objections made on the record; relevance and admissibility preserved automatically
- Contention questions moved out of the outline and into written interrogatories (Rifkind v. Superior Court (1994) 22 Cal.App.4th 1255)
Verification before you rely
- Every AI-drafted case citation independently confirmed against the actual authority
- Every AI-asserted transcript citation checked against the actual page and line
- Outline re-sequenced and edited by the attorney, not adopted as generated
Client-facing
- AI use addressed in the engagement agreement or otherwise disclosed as appropriate (Rule 1.4)
- Billing reflects time actually spent, not time saved
What Steps Still Require the Attorney's Own Judgment?
The list is short but it is the whole ballgame:
- Theory of the case and witness sequencing. Which witnesses to depose, in what order, and what each deposition must accomplish for summary judgment or trial.
- The examination itself. Reading the witness, deciding when to press and when to move on, and adapting the outline in real time. No tool does this.
- Objection calls on the record. The section 2025.460 waiver clock runs during the deposition. A privilege objection missed live is waived, and no post-hoc AI review restores it.
- Privilege assertions. Deciding what is privileged and instructing accordingly is legal judgment applied to facts under pressure.
- Legal reasoning about contentions. Rifkind adopted the observation that sorting facts into legal contentions "is what lawyers are for." That line is a fair summary of the whole framework: AI organizes the raw material; the lawyer supplies the reasoning.
When the deposition produces evasion or improper instructions, the follow-through is also attorney work, though the drafting is not. The meet-and-confer letter, the motion to compel, the separate statement, and the supporting declaration are all workflows Legion produces in court-ready format, and the strategic groundwork for that motion is laid out in our guide to compelling deposition responses under CCP § 2025.480.
Frequently Asked Questions
Is it ethical to use AI for deposition preparation in California?
Yes, subject to the existing Rules of Professional Conduct as applied by the State Bar's November 16, 2023 Practical Guidance. The core requirements are keeping confidential information out of insecure tools (Rule 1.6; Bus. & Prof. Code § 6068(e)), understanding the tool's limits, and verifying every output (Rule 1.1). The guidance is guiding principles, not a standalone rule, but the underlying duties are mandatory.
Can I upload deposition transcripts to ChatGPT?
Not to a free or consumer tier. Transcripts contain confidential and personal identifying information, and consumer tools may train on inputs or lack adequate security, which the State Bar guidance treats as disqualifying. Consumer tools also fabricate citations and quotations – the failure sanctioned in Noland v. Land of the Free, L.P. (2025) 114 Cal.App.5th 426 – so anything they generate carries candor exposure on top of the confidentiality breach. Use an enterprise tool with no-training and no-retention terms, or anonymize thoroughly before any prompt.
Uploading a transcript to a free or consumer AI tool risks breaching the confidentiality duty under Bus. & Prof. Code § 6068(e) and Rule 1.6, and any output it generates carries the fabrication and sanctions risk realized in Noland. See the full warning in the confidentiality section above before using any tool for client material.
Do I have to tell the court I used AI to prepare for a deposition?
No statewide rule currently requires disclosure of AI use in deposition preparation. Rule 10.430 governs the courts' own internal AI use, not litigants. Check the specific court's local rules and standing orders, since individual policies vary and some address AI in filings.
Can AI-generated deposition questions get me sanctioned?
The questions themselves will not, but fabricated authority flowing from AI prep into filed papers will. Noland v. Land of the Free, L.P. (2025) 114 Cal.App.5th 426 resulted in a $10,000 sanction and a State Bar referral for AI-fabricated quotations and citations in appellate briefs. Verify every citation, in every document, before you rely on it.
Can I ask contention questions at a deposition if AI drafted them?
The drafting source does not matter; the question type does. Under Rifkind v. Superior Court (1994) 22 Cal.App.4th 1255, questions demanding that a deponent state all facts supporting legal contentions are improper at oral deposition. Propound them as written contention interrogatories instead.
Does the seven-hour deposition limit apply to PMK depositions?
No. Code Civ. Proc. § 2025.290 excludes deponents designated as most qualified under section 2025.230 from the default seven-hour limit, along with expert witnesses and employment cases. For witnesses who are covered, courts must allow additional time where needed to fairly examine the deponent or where the examination is impeded or delayed.
Can I bill the client for AI-assisted deposition prep?
You may bill for time actually spent, including refining prompts and reviewing outputs. The State Bar guidance states you must not charge for the time saved by using generative AI, and it recommends a fee agreement that explains AI-related fees and costs.
Take the Next Step
AI deposition preparation in California works when the tool handles volume and the lawyer handles judgment. Legion was built on that division of labor by a California litigator who still litigates: no prompt engineering, court-ready output on California pleading paper, client materials never used for training, and every draft delivered for attorney review and signature. Get started with Legion and see how the discovery and motion workflows around your next deposition can move faster without moving outside the rules.
