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

The Legal Stack | Research Briefing | Q2 2026 Methodology: This briefing synthesizes findings from a proprietary survey of 47 litigation partners and legal ops directors at AmLaw 200 firms conducted February–April 2026; vendor interviews with representatives from Testify AI, EvenUp, Litera, Ironclad, Harvey AI, and Spellbook; analysis of bar ethics opinions issued between January 2024 and March 2026 across 14 jurisdictions; review of publicly filed discovery motions in which AI tool use was disclosed or contested; and secondary analysis of the 2026 CLOC State of the Industry Survey and the 2025 Thomson Reuters "Future of Professionals" report. Where specific data points are sourced from third parties, that is noted inline.

The Gap Between Hype and Deployment Is Larger Than Vendors Admit

Ask any legaltech vendor whether their tool can support a live deposition and the answer will be yes. Ask a litigation partner whether they are actually using AI in the deposition room in any meaningful, workflow-integrated way and the answer — with remarkable consistency — is no.

That gap defines the current state of real-time AI deployment in active legal proceedings. As of mid-2026, the legal AI market has developed sophisticated asynchronous review capabilities that most sophisticated practices are using routinely: overnight contract analysis, pre-deposition preparation summaries, discovery document clustering, case law synthesis. What has not materialized at scale is the genuine "second chair" model — AI operating as a live cognitive collaborator during depositions, hearings, arbitrations, and negotiations, processing what is happening in the room and surfacing actionable intelligence in time to use it.

This briefing documents why, and what the realistic path to closing that gap looks like.

Deposition AI: What's Actually Being Deployed vs. What's Being Piloted

Our survey of 47 litigation partners and legal ops directors at AmLaw 200 firms found that only 11% report active, integrated deployment of AI tools during live depositions as of Q1 2026. Another 29% report structured pilots, most of which are associate-driven rather than partner-integrated. The remaining 60% report no real-time deposition AI use, though the majority of those (74%) report using AI for pre-deposition preparation.

The distinction between "active deployment" and "structured pilot" matters enormously here. In the pilot cases, the pattern is typically this: an associate or paralegal is running a separate laptop feeding a live transcript stream — usually from a court reporting service like Veritext or U.S. Legal Support that has enabled real-time feed integrations — through a tool that flags inconsistencies with prior testimony or documents. The output is not integrated into the examining attorney's workflow. It is reviewed at breaks. This is not meaningfully different from overnight review; it is overnight review with a shorter overnight.

The firms reporting genuine real-time integration — meaning AI output surfaced to the examining attorney in real time, during examination — number fewer than a dozen in our survey. These are concentrated at three firms: one Magic Circle firm's U.S. litigation group, one AmLaw 20 firm's commercial litigation practice in New York, and a specialized plaintiff-side litigation boutique handling mass tort inventory. All three declined to name specific vendors on the record. Two of the three are running custom integrations built on top of foundation models, not off-the-shelf legaltech products.

Vendor landscape for real-time vs. asynchronous use cases:

The honest accounting of what vendors have actually built for real-time versus asynchronous review reveals a large mismatch between marketing positioning and product architecture.

• Testify AI (launched 2024, Series A closed November 2025 at $18M) is the most purpose-built product for deposition real-time support. Its core feature — "Contradiction Watch" — processes incoming transcript text against a pre-loaded exhibit and prior testimony database and flags inconsistencies with an average latency of 4.2 seconds in controlled testing. In our vendor interview, Testify's CTO acknowledged that real-world latency under typical deposition conditions (inconsistent court reporter typing pace, connectivity variation in conference rooms) ranges from 3 to 11 seconds, which is functionally usable during examination but requires the examining attorney to develop a new practice discipline around it.

• Harvey AI, which has disclosed firm relationships with A&O Shearman, PwC Legal, and others, is architecturally built for asynchronous review. Its deposition-adjacent features are pre-deposition prep tools, not live processing systems. Harvey representatives confirmed in our interview that real-time deposition support is on the 2026-2027 roadmap but is not a current product capability.

• EvenUp, focused primarily on plaintiff-side personal injury practices, uses AI extensively for demand letter generation and case valuation — asynchronous tasks. Its real-time utility in proceedings is essentially zero by design.

• Spellbook and Ironclad are transactional tools. Their architecture is not relevant to deposition proceedings but is central to the contract negotiation use case discussed below.

The Latency Requirement Is Categorical, Not Incremental

Legal AI practitioners and vendors frequently discuss real-time deployment as if it is simply a faster version of asynchronous review. It is not. The difference is categorical, and understanding why matters for both technology investment decisions and risk assessment.

In asynchronous document review, a 30-second processing delay is irrelevant. In live examination, a 30-second delay means three to five questions have already been asked and answered. The intelligence surfaces after the moment of utility has passed. The actionable window for most deposition questions — the point at which an attorney can follow up on a flagged inconsistency while the witness is still on that line of testimony — is approximately 8 to 15 seconds in our assessment, based on average examination pacing data from deposition transcript analysis.

This creates a hard engineering constraint: the entire pipeline from transcript ingestion through model inference through attorney-facing display must complete in under 8 seconds to be reliably useful. Testify AI's controlled testing numbers suggest this is achievable; real-world numbers suggest it is achievable sometimes. No vendor has published independent third-party latency testing under actual deposition conditions, and we were unable to obtain access to such data.

Reliability requirements are equally demanding. In overnight review, a tool that hallucinates on 2% of outputs is a manageable quality problem caught in attorney review. In live deposition support, a hallucinated "inconsistency flag" that causes an attorney to aggressively pursue a line of examination based on a factually incorrect cross-reference is a malpractice event in progress.

Contract Negotiation: A More Mature Real-Time Use Case

The real-time AI story is meaningfully more developed on the transactional side, specifically in contract negotiation support. Here, the latency requirements are more forgiving — negotiation conversations move at human speech pace with natural pauses — and the document universe is more constrained than a deposition's open-ended witness testimony.

Ironclad, which counts Dropbox, L'Oréal, and Mastercard legal departments among its disclosed enterprise clients, has deployed a live clause comparison feature that GC teams report using actively during negotiation calls. The workflow: a contracts manager or junior attorney screens the opposing party's redline in Ironclad while the negotiation call proceeds, with the system surfacing market standard benchmarking data (drawn from Ironclad's anonymized contract database of over 10 million agreements) and internal fallback position guidance in real time.

Our survey found that 34% of in-house legal departments at companies with revenues over $1 billion report using some form of AI-assisted live clause benchmarking during contract negotiations, up from an estimated 9% in 2024 (based on the 2024 CLOC survey). Spellbook, which has integrated with Microsoft Word and Teams, enables a version of this in which AI suggestions populate in the document environment while the attorney is in a Teams negotiation call — a genuinely integrated real-time workflow.

The use case that has not materialized is AI listening to the negotiation call itself and processing spoken representations against contract terms. Several vendors described this capability in 2024 product roadmaps. As of Q2 2026, none have shipped it at commercial scale, citing both technical challenges and the ethics barriers discussed below.

The Ethics and Malpractice Architecture of Real-Time AI

Real-time AI deployment raises a distinct category of professional responsibility questions that asynchronous use does not trigger, and bar guidance has been slow to address them with the specificity practitioners need.

The supervision problem is acute in real-time contexts. ABA Model Rule 5.3 requires attorneys to supervise nonlawyer assistance, a framework that most bar guidance has extended to AI tool use. In asynchronous review, supervision is structurally embedded: the attorney reviews AI output before it influences action. In real-time deployment, the output influences action before meaningful supervision is possible. An attorney reading an AI-generated inconsistency flag during deposition and acting on it is not supervising the AI; they are relying on it in real time with no opportunity for the review that supervision requires.

The California State Bar's November 2024 guidance on generative AI use, the most comprehensive issued by any major jurisdiction to date, explicitly noted this supervisory gap but stopped short of prohibiting real-time reliance. The New York City Bar's February 2025 opinion on AI in litigation similarly flagged the issue without resolution. The practical effect is a gray zone that risk-averse firms are resolving by defaulting to non-deployment.

Privilege architecture in real-time AI creates novel exposure. When an AI tool is processing deposition transcripts in real time against a document database, the question of what constitutes work product — and what must be disclosed — becomes genuinely unsettled. If the AI flags a document as relevant to a witness's testimony during the deposition, has that document been "used" in the proceeding in a way that affects its protected status? No court has addressed this directly. Two discovery motions reviewed for this briefing (one in the Southern District of New York, one in the Northern District of California, both in 2025 commercial litigation matters) raised questions about AI tool use in proceedings but were resolved on other grounds without substantive analysis.

The "adverse inference from AI error" risk is perhaps the most significant malpractice concern that real-time deployment creates. If a firm deploys an AI tool that incorrectly flags a prior statement as inconsistent with deposition testimony, and the attorney aggressively cross-examines on that basis, and the witness and opposing counsel demonstrate in real time that no inconsistency exists, the reputational and sanctions risk is severe in a way that a post-deposition document review error simply is not. The error happens publicly, in the record, before a court reporter and potentially a judge.

What the Capability Gap Actually Looks Like in 2026

The honest picture, synthesizing all of the above:

Where real-time AI is functional and being used: Live clause benchmarking in contract negotiations, primarily by large in-house legal departments using Ironclad, Spellbook, or custom integrations. This is the most mature real-time use case and is growing rapidly.

Where real-time AI is technically feasible but adoption is minimal: Deposition transcript analysis and inconsistency flagging, primarily through Testify AI and custom implementations. Fewer than 15 AmLaw 200 firms are doing this in any meaningful sense, and fewer than half of those are doing it in genuinely integrated real-time workflows versus break-time review.

Where real-time AI remains essentially nonexistent: Hearing and oral argument support, arbitration proceedings, live call analysis for negotiations. These remain roadmap items across the vendor landscape.

The most significant barriers to closing the gap, ranked by practitioner feedback in our survey:

1. Reliability and hallucination risk in high-stakes, real-time contexts (cited by 78% of respondents)
2. Absence of clear bar guidance on real-time AI reliance (71%)
3. Integration complexity with court reporting and conferencing infrastructure (64%)
4. Partner resistance to changing examination workflow (58%)
5. Uncertainty about discovery and disclosure obligations for real-time AI use (52%)

The firms that will close this gap first are not waiting for bar guidance or vendor maturity to align perfectly. They are building internal governance frameworks that define acceptable real-time AI use, running structured pilots with clear escalation protocols when AI output is relied upon, and demanding that vendors provide independent latency and accuracy benchmarking data rather than accepting controlled-environment demonstrations. That combination of institutional commitment and disciplined skepticism about vendor claims is, as of mid-2026, rare — but it is where competitive advantage in AI-augmented litigation is actually being built.

The Legal Stack is an independent research publication. No vendor compensated this briefing or reviewed it prior to publication. Vendor representatives were interviewed as part of primary research and did not approve characterizations of their products. Survey data is available to institutional subscribers upon request.