Legal Services — AI Deployment

Legal services moved from broadly AI-skeptic to AI-adopting across 2024–2026, a shift attributed by industry reporting to client pressure, associate cost structures, and competitive dynamics among Am Law 100 firms. Adoption spans dedicated legal-AI vendors, incumbent legal-research platforms, firm-built in-house tools, and general-purpose models used directly by lawyers and in-house teams.

Vendors and products

Deployment in the sector falls across several vendor categories:

Hyperscaler vertical packages

Google launched Gemini Enterprise for Legal on August 25, 2026, a plug-in combining integrations with legal software and data platforms with a set of agents Google said can handle specialized legal and administrative functions without significant human oversight. The platform connects to Thomson Reuters and to the legal AI developers Harvey and Legora; Reuters corrected an earlier version of its connector list to remove LexisNexis and add Legora. Weil Gotshal, Cleary Gottlieb, Freshfields and Williams & Connolly are among the firms collaborating with Google on the product, and Google Cloud product director Amrita Mehrok said it lets firms choose among the models embedded in the platform. Google announced tools for financial services the same day, with other industry-specific offerings to follow (Source: reuters.com).

The launch places a hyperscaler alongside the specialist vendors listed above rather than beneath them: the agents are Google's, while Harvey, Legora and Thomson Reuters appear as connected products inside the platform.

Vendor-built domain models

Through mid-2026 the sector's AI products were built almost entirely on general-purpose frontier models reached through an API. Thomson Reuters departed from that pattern with Thomson, a legal- and tax-domain model it trained itself from an undisclosed open-source foundation on Westlaw, Practical Law, Checkpoint and Reuters content, publishing first benchmark results on July 31, 2026 (Thomson Reuters Built Its Own AI Model That Now Ranks Among the World's Best). The company reported the model competitive with frontier models on legal benchmarks while trailing them on general reasoning and coding, under inference settings it acknowledged were not matched across the compared systems.

The change is bounded. Thomson became the default model behind Tabular Analysis, the high-volume structured document-review capability of CoCounsel Legal, on August 24, 2026; the rest of CoCounsel Legal continues to run on Anthropic's Claude Agent SDK, and administrators retain the ability to select other models in Tabular Analysis. Thomson Reuters said it spent about $40 million over two years on the project, with the final training run costing roughly $450,000, and plans to extend the model across its legal and tax portfolio over the following year (Source: siliconangle.com).

Anthropic legal package

On May 12, 2026, Anthropic announced 12 Claude legal-practice plug-ins, including "commercial counsel" (vendor-agreement review), "employment counsel", "litigation associate", and "law student" (a bar-exam study aid). The release added integrations with Thomson Reuters Westlaw and CoCounsel (described as "fiduciary grade"), Harvey (a rival legal-AI product), Box, Everlaw, and DocuSign. The plug-ins run inside Claude Cowork or embedded into firm systems. The announcement built on Anthropic's January 2026 Claude Cowork legal plug-in launch, which had triggered a selloff in Thomson Reuters and CoCounsel (Sources: reuters.com; bloomberg.com).

Mark Pike, Anthropic Associate General Counsel, said "We're seeing an incredible uptick in adoption of AI in the legal industry," and noted that a recent webinar drew more than 20,000 registrations. Thomson Reuters CTO Joel Hron characterized the CoCounsel integration as complementary rather than substitutionary, stating it "doesn't replace the CoCounsel Legal platform or provide standalone access to that underlying content and workflow system" (Sources: reuters.com; bloomberg.com).

The arrangement positions Claude as a reasoning engine that connects to industry-specific data and workflow tools via MCP rather than competing directly with them. A parallel pattern appeared in financial services with the claude-for-financial-services open-source release on May 14 (see Financial Services — AI Deployment).

OpenAI legal package

OpenAI introduced Astra for Law on September 17, 2026, combining GPT-6 Astra with a legal search index and instructions for legal analysis and writing. The index covers U.S. case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, with sources added daily, and draws on the case-law collection of the Free Law Project's CourtListener, which OpenAI says covers more than 99.9% of published U.S. precedential case law. OpenAI describes the index as complementing licensed content from providers such as Thomson Reuters rather than replacing it. On 200 U.S. legal research questions from the private validation set of Vals AI's Legal Research Bench, at the highest reasoning effort, OpenAI reports that Astra for Law passed the benchmark's overall correctness check on 54.0% of questions against 38.7% for GPT-6 Astra using web search alone (Source: openai.com).

Access runs first through a Trusted Access program for selected firms in ChatGPT and Codex, with the model to reach the API as gpt-6-astra-law; eligible firms get Zero Data Retention on the API. OpenAI named Harvey and Legora as API customers that will be able to build on it, and Latham & Watkins as a collaborator on information permissions, ethical walls, client instructions and firm oversight. Sullivan & Cromwell, Ropes & Gray and Cooley described firm-specific tools built with OpenAI's forward-deployed engineers, and Wachtell, Lipton, Rosen & Katz is named as a research collaborator. The launch also added 26 partner-built legal plugins — among them iManage, Intapp, DeepJudge and Thomson Reuters HighQ, with a CoCounsel Legal connector previewed — and nine community plugins carrying 47 custom skills (Source: openai.com).

Where Anthropic's May 2026 package connects Claude to incumbent legal-data vendors, OpenAI's pairs its model with its own legal index built partly on open case-law data, while still integrating those vendors as plugins. Both Harvey and Legora now appear as customers or connected products of more than one frontier platform, including Google's Gemini Enterprise for Legal.

Billable-hour and pricing pressure

Coverage of legal AI has centered on its effect on the billable-hour model. A Fortune feature (September 3, 2025) profiled Harvey under the headline "Meet the $100m AI startup that wants to kill the billable hour." If AI automates associate-level research and drafting, the economic foundation of hourly billing is reduced; commentators note firms could move to value-based pricing, which some argue may increase margins.

Client demand is described as a driver of adoption. Law360 Pulse (October 22, 2025) reported under the headline "Clients Are Driving The AI Revolution In Law Firms" that in-house legal teams are imposing AI-use expectations on outside counsel.

Bar ethics guidance addresses the same question under the rules on reasonable fees. The State Bar of New Mexico's Ethics Advisory Committee, in Formal Opinion 2024-004 (September 24, 2024), states under Rule 16-105 NMRA that a lawyer using generative AI "must only use the time the lawyer actually spent on the given task to calculate their fee and not include the time saved by the use of the tool." A lawyer who incurs costs for using an AI tool may seek reimbursement from the client, but must inform the client in writing of the actual cost and of the intent to charge it "prior to the representation" (State Bar of New Mexico Formal Ethics Advisory Opinion 2024-004 — Using Generative Artificial Intelligence in the Practice of Law).

Vendor scale and funding

Sector-specific vendors have attracted funding at multiples well above conventional software norms, on the argument that value can be captured one layer above the frontier models by applying them to expensive, specialized professional work.

Harvey is the most heavily capitalized example. The Information reported on August 7, 2026 that the company was in talks to raise at least $500 million at a $15.5 billion valuation including the investment, a 40 percent premium to the $11 billion valuation set five months earlier, with Lightspeed Venture Partners expressing interest in leading. The talks had not concluded and terms could still change. The reporting places Harvey above $350 million in annualized revenue, of which roughly $300 million is annual recurring revenue, up from about $190 million ARR at the end of 2025 and near $100 million during 2025. A round on the terms discussed would take total funding past $1.7 billion, against more than $1.2 billion raised to date from investors including Sequoia Capital, Andreessen Horowitz, GIC, Coatue, Kleiner Perkins, Conviction and Elad Gil, and would value the company at roughly 44 times annualized revenue (Sources: techstartups.com; theinformation.com).

The reported valuation sequence runs $3 billion in February 2025, $5 billion in June 2025, $8 billion in December 2025 in an Andreessen Horowitz-led financing, $11 billion in March 2026 on a $200 million round co-led by Sequoia and Singapore's GIC, and the proposed $15.5 billion (Source: techstartups.com).

On deployed scale, Harvey states that more than 100,000 lawyers across 1,300 organizations use its software, with customers including Latham & Watkins, A&O Shearman and O'Melveny, and corporate legal teams at Comcast and Verizon; the company reports more than 25,000 custom AI agents created inside customer organizations. Competitors named in the same reporting include the legal-technology incumbents Clio and Relativity, which are building their own AI capabilities, and the Swedish startup Legora (Source: techstartups.com).

Risk assessment and professional skills

An emerging practice termed legal red teaming involves lawyers stress-testing AI systems for legal-risk surfaces before deployment, described by DLA Piper in "Legal red teaming: A systematic approach to assessing legal risk of generative AI models."

State bar ethics guidance describes the terms on which lawyers and firms may use the technology. The State Bar of New Mexico's Ethics Advisory Committee concluded in Formal Opinion 2024-004 that New Mexico lawyers may use generative AI, though they are not required to, and listed ten Rules of Professional Conduct that its use implicates. On confidentiality (Rule 16-106), lawyers must understand whether a tool shares data with third parties or trains on inputs, must not enter confidential information into tools lacking adequate protections, "should always anonymize client information," and in appropriate cases should obtain informed consent. On duties to former clients (Rule 16-109), a lawyer who cannot independently verify that a tool retaining prior client information creates no conflict "should not use the tool for their current representation." On supervision (Rules 16-501 and 16-503), lawyers "must develop appropriate policies and procedures on the ethical use of Generative AI and train all other lawyers and staff" before using it. On communication (Rule 16-104), lawyers may need to discuss AI use with a client where it plays a significant role in final work product, the client asks, or use requires disclosing confidential client data. The opinion states that it is advisory only and binds neither lawyers, the disciplinary board nor any tribunal (State Bar of New Mexico Formal Ethics Advisory Opinion 2024-004 — Using Generative Artificial Intelligence in the Practice of Law; State Bar of New Mexico).

Commentary has also addressed effects on lawyer skills. A Fortune piece (July 14, 2025), "AI means lawyers must double down on this one skill," argued that judgment is a capability AI augments but cannot replace, a position parallel to AI Deskilling findings in healthcare.

The application of existing transactional law to autonomous agents remains unsettled. "UETA and LLM Agents: A Deep Dive into Legal Error Handling" (liftlab, February 3, 2025) examined how the Uniform Electronic Transactions Act applies when LLM agents execute legally binding transactions, an area the piece describes as untested.

Hallucinated citations in court filings

AI-fabricated citations in court filings have become a documented professional-responsibility problem. A running compilation of court filings flagged for AI hallucinations had recorded just over 1,000 filings as of July 2026, a figure commentators set against roughly 65 million new US cases filed annually (Source: hardcoresoftware.learningbyshipping.com).

The problem has surfaced in administrative adjudication as well as in filings by parties. On September 2, 2026 the Supreme Court of India set aside a ₹425.27-crore customs penalty order in Vijay Ghanshyam Gadiya v. Union of India, after finding that the order — issued by the Additional Commissioner of Customs in Surat — relied on non-existent case law, fake citations and AI-hallucinated legal ratios (Source: sci.gov.in). The distinction from the US filings compilation matters: the fabricated authorities were in the decision of a government adjudicator imposing a penalty, not in an advocate's submission, which places the failure inside the state's own reasoning rather than in material it was asked to evaluate. See Sycophancy and Hallucination, AI Liability.

A judicial rule addressed to the same problem appeared in China five days later. Article 19 of the Supreme People's Court's Opinions on Lawfully Adjudicating AI-Related Dispute Cases, issued September 7, 2026, requires litigation participants who use AI to prepare court filings or case-search reports to verify the authenticity and accuracy of the law, judicial interpretations and cases cited before submission, to disclose the AI assistance when filing, and to bear responsibility for the accuracy of the content. The same article directs courts to dismiss claims, and to impose fines or detention, where a party fabricates case facts using evidence obtained by deleting or altering generated-content labels, by specific instruction input, selective presentation of results or adversarial interference (Supreme People's Court Opinions on Lawfully Adjudicating AI-Related Dispute Cases (Fa Fa [2026] No. 10); China — SPC Opinions on Lawfully Adjudicating AI-Related Dispute Cases (法发〔2026〕10号)). Article 18 sets a companion evidentiary standard for AI-generated content offered as proof of infringement, directing courts to weigh the design of the prompts and their effect on the output, similarity to the asserted work, consistency across repeated tests, and the model's training, algorithm design and output-filtering mechanisms.

A primary sanctions record appeared in the United States the following week. In State of New Mexico v. Oscar Renee Sandoval; In re Stephen D. Aarons, No. S-1-SC-40845, the Supreme Court of New Mexico held Santa Fe attorney Stephen D. Aarons in direct contempt for filing a brief in chief prepared with ChatGPT, after he admitted the brief contained testimony from four wholly fabricated witnesses — Officer Michelle Amarillo, Officer Sanchez, Manal Al-Jibury and Teresa Marquez — false testimony attributed to three real witnesses, and misrepresented authority in State v. Lopez, 2005-NMSC-018, and State v. Manus, 1979-NMSC-035. He further admitted he did not verify the factual claims or legal authority before signing and filing, did not tell his client of that failure, and did not tell his client the show-cause proceedings were under way; the Court concluded he "demonstrated a lack of remorse and a lack of concern for his client." The sanctions are a $5,000 payment to the State Bar of New Mexico Client Protection Fund within thirty days, referral to the Disciplinary Board, and a bar on appearing before the Court pending the Board's outcome; all briefing in the underlying murder appeal was stricken and the Law Office of the Public Defender appointed (In re Stephen D. Aarons — Dispositional Order of Direct Contempt (N.M. Sup. Ct., 2026)). Reuters reports that Aarons told the August 21, 2026 hearing he had fed a computer-generated transcript and other case materials to the model presuming it would generate "a bulletproof summary," that Justice C. Shannon Bacon told him the problem of lawyers relying on AI hallucinations "is an above-the-fold story every single day," and that the appeal was reassigned on September 2, 2026 to public defender Kim Chavez Cook (Source: reuters.com).

The order is nonprecedential by its own election under Rule 12-405(B) NMRA and is not New Mexico precedent on AI-generated filings; its significance is as a record of what a state supreme court imposed rather than as authority. The two failure modes it distinguishes — witnesses who do not exist, and words attributed to witnesses who do — are separate defects that accounts of the episode tend to collapse. See AI-Fabricated Citations for the pattern across jurisdictions.

The order grounds the contempt finding in Rule 12-312(D) NMRA, which makes noncompliance with the briefing requirements of Rule 12-318 NMRA grounds for sanctions, and in the Court's superintending authority under the New Mexico Constitution; it does not cite any Rule of Professional Conduct. Advisory ethics guidance on the conduct had been issued in New Mexico two years earlier. Formal Opinion 2024-004 of the State Bar of New Mexico's Ethics Advisory Committee, issued September 24, 2024, construes Rules 16-303 and 16-804(C) NMRA to require that a lawyer using generative AI to draft a court filing "independently verify the accuracy and sufficiency of all research, citations, and analysis conducted by the AI tool," without relying on the same tool to check its output, and treats a failure to correct the tool's errors as "a knowing misrepresentation of the law." The opinion is advisory, and its verification language is framed around research, citations and analysis rather than AI-generated accounts of the factual record such as the testimony at issue in the Aarons brief (State Bar of New Mexico Formal Ethics Advisory Opinion 2024-004 — Using Generative Artificial Intelligence in the Practice of Law). A federal court in the same state quoted the opinion in Dehghani v. Castro (D.N.M. May 9, 2025) when affirming Rule 11 sanctions, including a $1,500 fine and self-reports to the New Mexico and Texas state bar disciplinary boards, against an attorney who filed a contract-drafted brief citing six cases that did not exist and thirteen that did not support the propositions stated (Source: caselaw.findlaw.com).

Legal education

Law schools have begun formalizing AI policies. In July 2026 the University of Chicago Law School announced a new AI teaching strategy that bans phones and laptops in class for first-year students, pairing device restrictions with a structured approach to AI use in instruction (Source: law.uchicago.edu). Steven Sinofsky argued in a July 11, 2026 essay that the device ban repeats Harvard Law School's 1982 ban on portable computers in exams, characterizing it as a recurring institutional response to new technology rather than a durable policy, and using the hallucinated-filings compilation to argue the rate is small, "just over 1,000 cases" against "about 65,000,000 new cases filed every year" (Source: hardcoresoftware.learningbyshipping.com).

Relationships

• deployed-by: Am Law 100 firms, in-house teams, Thomson Reuters, Cleary Gottlieb, Anthropic legal customers.
• deployed-by: Thomson (Thomson Reuters legal-domain model) — powers Tabular Analysis in CoCounsel Legal.
• deployed-by: Astra — configured as Astra for Law (September 2026) for selected firms and legal-technology API customers.
• regulated-by: State bar associations, ABA (guidance), judiciary (discovery rules).
• regulated-by: State Bar of New Mexico — advisory ethics guidance on generative AI in Formal Opinion 2024-004 (State Bar of New Mexico Formal Ethics Advisory Opinion 2024-004 — Using Generative Artificial Intelligence in the Practice of Law).
• depends-on: Anthropic, OpenAI, Harvey.
• related: AI-Fabricated Citations — the cross-jurisdictional pattern of fabricated authority in filings and adjudications.
• related: In re Stephen D. Aarons — Dispositional Order of Direct Contempt (N.M. Sup. Ct., 2026) — the New Mexico Supreme Court contempt order.
• related: AI Deskilling, Anthropic Economic Index — March 2026: Learning Curves (legal among top Claude use categories), AI Liability, Enterprise AI Deployment Gap.