Generative AI in Legal Practice: A 12-Month Field Report
Twelve months ago, the legal industry was absorbing a wave of vendor promises that bordered on the messianic. AI would draft contracts in seconds, eliminate associates, democratize legal services, and compress billing cycles. Twelve months later, the picture is considerably more textured. Generative AI has...
Where the Technology Has Actually Landed
Twelve months ago, the legal industry was absorbing a wave of vendor promises that bordered on the messianic. AI would draft contracts in seconds, eliminate associates, democratize legal services, and compress billing cycles. Twelve months later, the picture is considerably more textured. Generative AI has proven genuinely transformative in a narrow band of tasks, conspicuously unreliable in others, and has quietly reshaped workflows in ways that neither its champions nor its skeptics predicted with much accuracy.
The Use Cases That Have Proven Out
Document summarization and first-pass review is the clearest win. Firms using tools like Harvey, CoCounsel (formerly Casetext), and Lexis+ AI report consistent productivity gains when the technology is deployed to summarize lengthy contracts, deposition transcripts, or due diligence bundles. A 2024 survey by the Thomson Reuters Institute found that lawyers using AI for document review reported time savings averaging 30 to 40 percent on those specific tasks. Kirkland & Ellis, Allen & Overy (now A&O Shearman), and Linklaters have all publicly acknowledged deploying AI-assisted review tools at scale in their transactional practices, with the primary use case being the rapid distillation of large document sets rather than autonomous drafting.
Legal research as augmented search has also delivered. Tools like Lexis+ AI and Westlaw Precision with CoCounsel have moved beyond keyword retrieval toward something closer to conversational research. Practitioners at mid-size litigation firms describe using these tools to generate research memos on discrete questions of law — a first draft that a junior associate then verifies and refines. The critical word is verifies. Thomson Reuters' own user data, reported in their 2024 annual survey, indicated that attorneys found AI-assisted research most valuable precisely when they treated outputs as starting points rather than conclusions. That qualification matters enormously.
Contract drafting support — specifically clause-level drafting rather than full agreement generation — has found genuine traction. Standard commercial agreements: NDAs, SaaS subscription agreements, employment offer letters. Firms and in-house teams using tools like ContractPodAi, Ironclad, and Spellbook report that drafting boilerplate-heavy documents has accelerated materially. Microsoft's legal team reportedly used AI-assisted drafting to reduce the time spent on routine vendor agreements by roughly half. The qualifier again is routine: complex bespoke agreements remain heavily attorney-dependent.
The Disappointments
Autonomous brief-drafting has not arrived. The much-hyped promise that AI would generate litigation briefs ready for filing has collided with reality. The reasons are both technical and structural. Large language models hallucinate citations with enough frequency to create serious professional liability exposure. The case that crystallized this for the profession was Mata v. Avianca (S.D.N.Y. 2023), in which attorneys Steven Schwartz and Peter LoDuca submitted a brief containing entirely fabricated case citations generated by ChatGPT. Judge Castel imposed $5,000 sanctions on the attorneys and their firm, Levidow, Levidow & Oberman. The case became a Rorschach test for the profession — interpreted either as a cautionary tale about careless adoption or as an indictment of a specific misuse rather than the technology itself. Regardless of interpretation, it installed a deep institutional wariness around AI-generated citations that has not fully dissipated.
A 2024 study by Stanford's RegLab found hallucination rates in legal citation tasks ranging from 17 to 33 percent depending on the model and prompt design — rates that are simply incompatible with professional standards without systematic verification layers.
Transactional drafting for high-stakes deals has underdelivered. Partners at major M&A practices report that AI tools struggle with the contextual complexity of deal-specific provisions, jurisdiction-specific regulatory requirements, and the kind of commercial judgment embedded in heavily negotiated agreements. One senior partner at a Magic Circle firm, speaking to The American Lawyer in early 2024, described the technology as "excellent at producing plausible text and mediocre at producing correct text" in complex deal contexts. The distinction is uncomfortable because plausible-but-incorrect is arguably more dangerous than obviously wrong.
Client-facing AI tools have moved slowly. Despite considerable investment in AI-powered client portals and chatbots, uptake has been tepid. Concerns about unauthorized practice of law, data privacy, and liability have pushed most deployments into internal-only configurations. Rocket Lawyer and LegalZoom have made public claims about enhanced AI features, but independent assessments of their accuracy on jurisdictionally variable questions have been mixed.
How Workflows Have Actually Changed
The most honest answer is that workflows have changed in ways that resemble the early adoption of legal research databases more than they resemble the disruption narratives circulating a year ago.
Junior associates are not being replaced — but their work product has changed in character. First-year associates at firms that have deployed Harvey or CoCounsel describe spending less time on initial research compilation and more time on critical evaluation of AI-generated drafts. Several attorneys at AmLaw 100 firms told The American Lawyer they now begin most research tasks with an AI-generated memo and spend their time identifying what is wrong with it rather than building from scratch. Whether this is skill-atrophying efficiency or productive reallocation of cognitive effort is a live debate in legal education circles.
Billing models are being quietly renegotiated. The billable hour calculus is under genuine pressure. If a task that once took four hours now takes ninety minutes with AI assistance, value-based billing advocates argue the profession has an opportunity to reframe pricing. The more common short-term response has been less principled: some firms have reduced associate hours billed on AI-assisted tasks while absorbing the efficiency gain as margin improvement. The ABA's Formal Opinion 512 (2024) confirmed that attorneys may not charge for time saved by AI tools as if those tools had not been used, but enforcement is practically nonexistent.
Data governance has emerged as the dominant operational concern. Firms discovered quickly that using commercial AI tools with client data requires careful vendor assessment and, in many cases, bespoke enterprise agreements. The deployment of Microsoft Copilot across law firm environments created significant data residency and confidentiality questions. Several Am Law 200 firms paused rollouts in late 2023 to renegotiate terms. This is where much of the past year's internal legal tech work has actually concentrated — not deployment, but the governance architecture around deployment.
What Practitioners Say About Reliability
The most consistent finding across surveys, interviews, and published firm assessments is that reliability is task-specific to a degree that generic endorsements or condemnations of AI in law are almost meaningless.
A 2024 survey of 600 legal professionals by the International Legal Technology Association found that 71 percent rated AI tools as "reliable" or "very reliable" for summarization, while only 31 percent rated them similarly for legal analysis. The gap is instructive. Tasks where reliability is high share common features: they are bounded, verifiable, and tolerant of imprecision. Tasks where reliability is low involve legal judgment, jurisdictional specificity, and citation accuracy.
Practitioners who have found the technology most useful cluster around a consistent methodology: decompose complex tasks into discrete subtasks, apply AI to the subtasks most amenable to pattern-matching, and apply human judgment to assembly and quality control. It is less a revolution in legal work than a restructuring of where within legal work cognitive effort is concentrated.
The firms positioned to benefit most are not those that have adopted AI most enthusiastically, but those that have developed the clearest protocols for knowing exactly where it can be trusted — and exactly where it cannot.
This briefing draws on publicly reported data, industry surveys, and published practitioner commentary. Statistics and citations should be independently verified before professional reliance.