The Legal Stack
Independent LegalTech Analysis
← Research Briefings
Research BriefingNo. 087 · August 16, 2026 · 10 min read
Legal AI · Research Report

The Legal AI 'Second Chair' Deployment Report 2026: How Law Firms and Legal Departments Are — and Are Not — Using AI in Real-Time During Depositions, Hearings, and Negotiations — and What the Capability Gap Actually Looks Like

Real-time AI tools for depositions exist, but their accuracy and uptake are largely unmeasured. What the products do, where errors arise, and ethics rules.

Research briefing — a review of public sources. The Legal Stack did not conduct a survey, benchmark or interviews for this piece.

Executive summary

Can AI serve as a live "second chair" during depositions, hearings and negotiations, processing what is happening in the room quickly and reliably enough to act on? Public evidence on how many firms actually do this is essentially nonexistent. We found no published survey that measures real-time AI use in proceedings, and no independent latency or accuracy testing of real-time deposition tools.

The public record does show four things. Commercial real-time deposition products exist. Both layers they rely on, speech-to-text and language-model analysis, have documented error rates. Bar guidance on AI-recorded conversations is arriving. And courts have begun ruling on whether AI-generated material is privileged. This briefing sets out that record and gives labelled analysis of where the capability gap actually lies.

What is commercially available

The clearest documented example is Filevine's Depo CoPilot. LawSites reported its launch in September 2024 as a standalone product that "transcribe[s] and analyze[s] your depositions in real time, as they are happening." According to that coverage, it records through the user's microphone in person or joins an online deposition, labels each utterance as a question, answer or third-party statement, and checks attorney-set goals every few minutes. It flags inconsistencies, suggests follow-up questions and points out unanswered or ambiguous answers, each with transcript references. At launch, access was limited to conference attendees and a waitlist.

On 4 September 2025, Filevine announced "Depositions by Filevine." The package combines scheduling and hosting, Depo CoPilot as an AI "second chair" during the deposition, and post-deposition review, with rough-draft transcripts delivered to both sides within 24 hours.

Asynchronous deposition tools are more common. NetDocuments' ndMAX Studio, launched in February 2025, includes a "Deposition Prep" app among its 12 AI applications, according to IT Brief.

Analysis. The marketing claim that real-time deposition support exists is accurate: there are products. What is missing is public evidence of how widely they are used, how fast they are under real conditions, and how often their flags are wrong. Vendor-specific latency figures and adoption estimates in earlier versions of this briefing were not verifiable and have been removed.

Two layers of error

A real-time tool can fail at two points. It can mishear what was said, and it can misanalyse what it heard.

Transcription. In "Careless Whisper," presented at the ACM FAccT conference in June 2024, Cornell-led researchers found that roughly 1% of transcriptions from OpenAI's Whisper speech recognition system contained entire hallucinated phrases. The test set was more than 13,000 audio clips. Some of the hallucinations included violent language or invented personal information. Cornell notes that OpenAI has since improved the model and that the lead author reports lower rates. The study drew partly on speech from people with aphasia, so its rates should not be read as deposition rates. The point is that speech-to-text can invent content, not just mishear it.

Analysis. Stanford RegLab and HAI researchers tested legal research tools in May 2024. Lexis+ AI and Ask Practical Law AI produced incorrect answers more than 17% of the time, and Westlaw AI-Assisted Research hallucinated more than 34% of the time. These are research tools, not deposition tools, but they show that retrieval-grounded legal AI still errs at material rates.

The consequences of unverified AI output are well documented. In Mata v. Avianca, Inc. (S.D.N.Y. 22 June 2023), the court imposed a $5,000 sanction under Rule 11, jointly and severally, over fabricated case citations produced with ChatGPT, according to Reed Smith. By the start of June 2026, Damien Charlotin's public database listed 1,058 U.S. cases involving improper use of generative AI in court filings, according to UNC Law Library. Practitioners are aware of the risk: 75% of respondents to the ABA's 2024 Legal Technology Survey named accuracy as their top AI concern, according to LawSites.

Analysis. In asynchronous work, review catches errors before they matter. In live examination there is no review step, so a false inconsistency flag can lead an examiner down a line of questioning on the record. Accuracy requirements for real-time use are therefore higher than for overnight review, not lower. In our view, that is the categorical difference between the two modes.

The ethics framework

Supervision and consultation. ABA Formal Opinion 512 (29 July 2024) applies the duties of competence, confidentiality, communication, candor and supervision to generative AI. It says client consultation about a generative AI tool "is also necessary when its output will influence a significant decision." It also says informed consent is required before client information is put into a self-learning tool, and that boilerplate engagement-letter language is not sufficient. Opinion 512 does not address real-time use specifically. Its verification expectations are, however, hard to meet live.

Recording and transcription. NYC Bar Formal Opinion 2025-6 (22 December 2025) addresses AI tools that record, transcribe and summarise conversations with clients. Clients must be notified and their consent obtained before an AI system records their calls. Lawyers should independently review recordings, transcripts and summaries soon after they are prepared. And lawyers should evaluate a tool's security, retention, discoverability, and use of data for training. The opinion expressly does not cover conversations with opposing counsel, witnesses or courts. The NYC Bar's site links a separate Formal Opinion 2026-2 on non-client conversations, which Law360 reported under the headline "NYCBA Cautions Against Recording Nonclient Calls." For any tool that listens to a negotiation call or a remote deposition, these opinions are the most direct guidance currently available.

Privilege and work product. Two February 2026 federal decisions point in different directions on AI-generated material, according to Wilson Elser. In United States v. Heppner (S.D.N.Y.), Judge Jed Rakoff held that a defendant's own conversations with a consumer AI chatbot were neither privileged nor work product. In Warner v. Gilbarco, Inc. (E.D. Mich.), Magistrate Judge Anthony Patti treated a pro se plaintiff's ChatGPT prompts and outputs as protected work product. Neither case involved a lawyer's real-time deposition tool. But both show that courts are now examining AI-generated litigation material and the platform terms behind it.

Negotiation: less documented than claimed

Earlier versions of this briefing described live clause benchmarking as widespread in in-house negotiations and attributed adoption figures to a CLOC survey. We could not verify those figures, and they have been removed.

What is documented is a fast-growing contract AI market built around the document, not the call. When Spellbook announced its $50 million Series B on 9 October 2025, it reported nearly 4,000 law firm and in-house customers in 80 countries. It described a product that works directly in Microsoft Word, with "Market Comparison" features, which ground suggestions in market data, still in beta. Tools like this support drafting and redlining that a lawyer can do during a negotiation without the tool hearing the call. A tool that listens to the call itself would fall within the recording and consent questions raised by the NYC Bar opinions.

Where the gap actually is (analysis)

  • Functional today: AI-assisted preparation before depositions and negotiations, and AI-assisted redlining during negotiations, with a human reading the output.
  • Available but unproven at scale: live deposition analysis such as Depo CoPilot. Products exist, but there is no public data on adoption, latency or false-flag rates.
  • Largely undocumented: AI support in hearings and oral argument, and AI listening to negotiation calls.

The binding constraints are reliability, verification and consent, not model speed.

Recommendations (analysis)

  1. Pilot in low-stakes settings first, such as internal mock depositions, and log every AI flag against what the transcript actually shows to build your own false-positive rate.
  2. Require vendors to show evidence, not demonstrations: error rates on real audio, data-retention terms, and whether inputs train models.
  3. Treat AI flags as prompts to check, not facts to put to a witness. Verify against the source document before using a flag in examination.
  4. Address consent up front. Follow NYC Bar 2025-6 for client calls. For depositions and negotiations involving other parties, get agreement on the record or in writing before any AI recording.
  5. Prefer enterprise terms. After Heppner, avoid consumer AI tools for anything touching privileged strategy.

Sources

Filed under Legal AI → · The Legal Stack accepts no vendor funding for its research.

More Research

View all →
No. 116
10 min
What Published SEC Enforcement Actions, Staff Guidance, and No-Action Letters Through October 2026 Actually Say About AI-Assisted Legal Opinions, Disclosure Drafting, and the Attorney Liability Exposure When AI-Generated Securities Filings Contain Material Misstatements
10 min
No. 115
10 min
What Published IRS Guidance, Tax Court Opinions, and Treasury Proposed Rulemaking Through October 2026 Actually Say About AI-Assisted Tax Legal Work Product — Privilege, Disclosure Obligations, and the Penalty Exposure When AI-Generated Positions Turn Out to Be Wrong
10 min
No. 114
10 min
What Published FTC Guidance, State AG Enforcement Actions, and Bar Ethics Opinions Through October 2026 Actually Say About AI-Generated Legal Intake Tools That Collect Consumer Data Before Any Attorney-Client Relationship Is Formed
10 min
© 2026 The Legal Stack — Independent LegalTech Analysis