The Legal AI Landscape

The legal industry has crossed an inflection point. Generative artificial intelligence, a curiosity in late 2022 and a pilot project through 2024, became infrastructure in 2026: survey data published this spring shows the share of legal professionals using generative AI at work more than doubling in a single year, from roughly 31 percent to 69 percent, with more than four in ten now using legal-specific AI tools. Legal employment, far from collapsing, reached a ten-year high in January 2026. The question facing the profession is no longer whether AI will be used, but who will control it, who will profit from it, and who will bear its risks.

Three weeks between May 12 and June 4, 2026 reorganized the competitive landscape. Anthropic launched Claude for Legal, placing a frontier model lab in direct contact with legal buyers for the first time at scale. Kirkland & Ellis, the highest-grossing law firm in the world, committed half a billion dollars to building proprietary AI on its own institutional knowledge, partnering with Palantir rather than a legal tech vendor. OpenAI launched a dedicated legal vertical under Ironclad co-founder Jason Boehmig. And Palantir itself entered legal services. Four of the most powerful technology companies in the world now compete for the legal market — one they had previously served only as wholesale suppliers of intelligence.

This guide makes four arguments. First, understanding legal AI requires understanding the technology stack beneath it: foundation models are general-purpose engines, and nearly all of the legal-specific value is created in the layers above them. Second, the model layer is commoditizing, which systematically pushes differentiation — and pricing power — upward toward applications, workflow, and above all proprietary knowledge. The defining strategic insight of 2025, that data is the moat rather than the model, became the defining capital allocation of 2026. Third, the profession’s institutions have not kept pace with its individuals: a majority of firms still provide no AI training and many have no formal policy, even as courts sanction lawyers in growing numbers for filing machine-fabricated citations. Fourth, AI is not eliminating legal jobs so much as reshuffling them.

It is written for readers entering the industry from any seat — associates, partners, firm leadership, technologists, in-house counsel, and the recruiters and vendors who serve them — and it closes with the open questions no firm has yet answered, because fluency in those questions is what currently distinguishes sophisticated participants in this market.

A large language model, or LLM, is a neural network trained on an enormous corpus of text to perform a deceptively simple task: predict the next word (more precisely, the next “token”) in a sequence. Trained at sufficient scale — across most of the public internet, books, code, and licensed material — this prediction objective yields something unexpected: a system that can draft, summarize, translate, reason through multi-step problems, and converse fluently in the idiom of virtually any profession, including law. The model has not memorized a database of answers; it has internalized statistical patterns of language and, with them, a substantial amount of the world’s expressed knowledge and reasoning.

The term foundation model, coined by Stanford researchers in 2021, captures the economic role these systems play: they are general-purpose foundations on which many different applications are built, the way electricity or cloud computing underpins industries without belonging to any one of them. Anthropic’s Claude, OpenAI’s GPT series, and Google’s Gemini are foundation models; the chat interfaces most people encounter are merely the most visible products built on top of them. A handful of “frontier labs” train these models at costs reaching billions of dollars per generation, which is why almost no one else — including the wealthiest law firms — attempts to train their own from scratch.

Two properties of LLMs matter disproportionately for legal work. The first is hallucination. Because the model generates text by plausibility rather than by lookup, it can produce statements — including case citations, quotations, and holdings — that are fluent, confident, and entirely fabricated. This is not a bug to be patched but a structural characteristic to be engineered around, and it is the proximate cause of the sanctions epidemic discussed below. The second is grounding: a model’s reliability improves dramatically when it is connected to verified sources and required to retrieve before it generates. The technique, called retrieval-augmented generation (RAG), is the architectural heart of nearly every serious legal AI product — the application fetches authoritative material (a citator database, a document management system, a precedent bank) and instructs the model to answer from that material, with citations a human can check.

Around these basics has grown a vocabulary worth knowing. A context window is the amount of material a model can consider at once; modern windows accommodate hundreds of pages, which is what makes contract review and deal-document analysis practical. Fine-tuning adapts a model’s weights to a domain, though most legal applications today rely instead on prompting, retrieval, and tools. An agent is a model given the ability to take actions — search a database, run a checklist, draft into a template, route a document for signature — and to chain those actions toward a goal with limited human intervention. The industry’s shift from “chatbots” to “agents” across 2025 and 2026 is the shift from AI that answers questions to AI that performs work, and it is why questions of supervision and liability have become urgent.

The essential takeaway for a newcomer: a raw model is not a legal product. The model supplies general intelligence; everything that makes it safe and useful for legal work — grounding in verified law, workflow design, security and privilege protection, audit trails, and the embedded judgment of the lawyers who configured it — is built in the layers above. Those layers are where the legal AI industry actually lives, and where the competition described in this guide is being fought.

It is useful to picture the market as a five-layer stack. At the base sits compute infrastructure — the chips, data centers, and cloud platforms on which everything else runs. Above it sit the foundation models themselves, produced by the frontier labs. The third layer is orchestration and middleware: the APIs through which applications call models; retrieval pipelines that ground answers in real sources; agent frameworks; security, permissioning, and audit machinery; and, increasingly, the Model Context Protocol (MCP), an open standard Anthropic released in late 2024 that has become the de facto plumbing by which AI systems connect to external software.

The fourth layer is the application layer: the legal-specific products a lawyer actually touches — Harvey and Legora; Thomson Reuters’ CoCounsel and LexisNexis’ Lexis+ AI; contract lifecycle platforms such as Ironclad and Spellbook; e-discovery and litigation analytics tools; and, since May 2026, the labs’ own legal offerings. The fifth layer — easy to overlook and increasingly understood as the most valuable — is firm knowledge and human judgment: the precedent banks, playbooks, negotiation histories, and supervisory review that convert a capable system into defensible legal work product. The lawyer’s signature on the filing lives at this layer, and no amount of capability below it removes the obligation it represents.

The strategic physics of the stack can be stated in one sentence: as the model layer commoditizes, value migrates upward. Frontier models from rival labs have converged toward comparable performance on most legal tasks, and applications increasingly treat them as interchangeable inputs — Harvey, originally an OpenAI-exclusive product, now offers Claude and GPT models side by side in a selector. When the engine is a commodity, the durable assets are the things that cannot be swapped: the workflow a firm has embedded, the data network effects a vendor has accumulated, and the institutional knowledge a firm alone possesses. Every major strategic move of 2026, from Kirkland’s build program to the labs’ vertical launches, is an attempt to claim territory in the upper layers.

Until this year the market observed a tidy division of labor. The frontier labs sold intelligence wholesale: OpenAI, Anthropic, and Google supplied models through APIs, and legal technology companies — flush with venture capital — built the legal products on top. OpenAI seeded Harvey and even built a custom case-law model with it. Anthropic’s models powered numerous legal tools. The labs were arms suppliers, deliberately neutral; the vendors competed with one another; and the incumbents, Thomson Reuters and LexisNexis, raced to embed AI in the research platforms firms already licensed. A law firm’s relationship with a frontier lab was indirect, mediated entirely by the application layer.

That peace ended in stages and then suddenly. In February 2026, Anthropic released a single legal plugin for Claude — and shares of RELX (LexisNexis’ parent), Thomson Reuters, and Wolters Kluwer fell on the news, a market verdict on what direct lab participation could mean for incumbent economics. On May 12, the full Claude for Legal launch followed: more than twenty MCP connectors into the software legal teams already run, twelve practice-area plugins, and an open-source library of more than ninety adaptable workflow agents. Microsoft shipped a Legal Agent in the same season. On June 1, OpenAI launched a dedicated legal vertical and hired Jason Boehmig, the lawyer-founder who built contract platform Ironclad to a $3.2 billion valuation, to lead it. Three days later Palantir — not a model lab but the premier builder of enterprise data “ontologies” — entered legal services through an exclusive platform partnership with Kirkland & Ellis.

The resulting structure is what strategists call co-opetition, and it is genuinely awkward. Harvey buys its intelligence from OpenAI and Anthropic, both of which now sell legal products against it; meanwhile Claude connects to Harvey through MCP, so the products interoperate even as their makers compete. The vendors’ defense is threefold: go model-agnostic so no lab controls them (Harvey’s pivot); go deeper into workflow than a general-purpose assistant can easily follow (the agent-builder race); and accumulate data and distribution the labs lack (the incumbents’ citation-grounded research moats — notably, Thomson Reuters chose to connect its data to Claude rather than cede the channel). The labs’ advantage is price and ubiquity: Claude for Legal’s plugins are open source and ride on an ordinary Claude subscription, collapsing the cost floor that had kept dedicated legal AI out of reach for solo practitioners, small firms, and lean in-house teams. Whether buyers ultimately prefer a polished closed product, an open layer they can shape, or both for different work is among the genuinely unresolved questions of the moment.

Every industry is adopting AI; few adopt it under the constraints law imposes. A lawyer personally vouches for every filing — by rule, by professional oath, and by sanction exposure. Client information is protected by confidentiality duties and evidentiary privilege, which means the question “where does this data go when it enters the tool?” is not an IT preference but an ethical mandate, and courts have already signaled that strategy typed into a consumer chatbot may be discoverable. The cost of error is asymmetric: a hallucinated citation is not a typo but a misrepresentation to a tribunal. And the profession is regulated: duties of competence now explicitly encompass technological competence; AI tools must be supervised the way nonlawyer assistants are; and a lawyer cannot bill for hours the machine saved as though a human had worked them.

These constraints produce what might be called a verification economy. AI compresses the cost of producing legal text toward zero while leaving the cost of verifying it — and the consequences of failing to — fully intact. The economic and professional premium therefore shifts from production to judgment: knowing what to ask, recognizing what is wrong, and standing behind what survives review. Nearly every theme in this guide, from the sanctions wave to the redesign of associate training, is downstream of that single shift.

The modern era opened in November 2022 with ChatGPT, which put a fluent general-purpose assistant in front of every managing partner and, more consequentially, every client. Harvey had been founded earlier that year with OpenAI’s backing, and in February 2023 Allen & Overy gave the industry its wake-up call by deploying it to more than 3,500 lawyers across forty-plus offices — the first firmwide generative-AI rollout in BigLaw. GPT-4’s arrival in March 2023 made serious legal work feasible, and Casetext’s CoCounsel proved it within days. The same spring delivered the cautionary counterweight: in Mata v. Avianca, a federal judge sanctioned lawyers who had filed a brief built on AI-invented cases — the profession’s first hard lesson in the technology’s signature failure mode. Thomson Reuters’ $650 million acquisition of Casetext that June announced that the incumbents would buy their way into the new era, and by late 2023 Lexis+ AI and Westlaw’s AI-assisted research had carried generative AI into the platforms firms already paid for.

2024 was the year the industry stopped trying to build its own models and started buying applications; 2025 was the year it understood why. As frontier models converged in capability, sophisticated buyers concluded that the durable asset was not the model but the data — a firm’s own work product, precedents, and institutional knowledge. Harvey crossed $100 million in annual recurring revenue that August with fifty of the AmLaw 100 as clients; in November, Clio completed its $1 billion acquisition of vLex, the largest deal in legal tech history, carrying AI research and drafting to the hundreds of thousands of professionals on its practice-management platform.

2026 converted thesis into capital. Adoption data published in the spring showed individual use doubling year over year even as institutional governance lagged badly. Harvey raised $200 million at an $11 billion valuation in March. Then came the three weeks chronicled above: Claude for Legal on May 12; Kirkland’s $500 million proprietary-build commitment reported May 27; OpenAI’s legal vertical on June 1; the Kirkland–Palantir platform on June 4. Five years after ChatGPT, the strategic question had inverted — from “which AI tool should we buy?” to “how do we weave AI into everything we do, and how do we get our own knowledge into whichever system we use?”

The numbers describe an industry that adopted faster than it organized. Sixty-nine percent of legal professionals now report using general-purpose generative AI for work, up from thirty-one percent a year earlier; forty-two percent use legal-specific tools, double the prior year. At the firm level, roughly half of law firms have implemented general-purpose AI, rising toward six in ten among firms with twenty or more lawyers. Solo and small practices, unburdened by committee decision-making, report some of the highest adoption of all. Reported productivity gains are real but unevenly converted: large majorities describe meaningful weekly time savings, while most small firms concede they have not changed their pricing to reflect them — efficiency captured, value uncaptured.

Against that enthusiasm stands the governance gap, the defining institutional fact of the moment: a majority of firms provide no AI training at all, and more than four in ten have no formal AI policy and no plan to create one. The most common tools in actual daily use are consumer assistants — ChatGPT, Copilot, Gemini, Claude — adopted bottom-up by individuals, frequently outside any firm-sanctioned environment. The gap is not abstract. It is the direct mechanism behind the sanctions cases: lawyers using powerful tools without training, policy, or verification discipline. For firm leadership the implication is uncomfortable but clarifying: the technology has already been adopted; what remains unbuilt is the institution around it.

Every legal organization now faces a version of the same capital-allocation question, and the market conveniently ran three natural experiments in the spring of 2026 — one for each answer.

The honest synthesis is that these are postures, not tribes, and most organizations will blend them: a frontier assistant for general work, one or two applications for specialized workflows, and a growing investment — modest or Kirkland-scale — in structuring proprietary knowledge so that any of the above can use it. The decision that cannot be outsourced is the last one: whose system one’s institutional knowledge lives in, and on what ownership terms.

Two architectures now contend for the legal buyer. The closed environment — classic Harvey, the incumbents’ platforms, and in its own way Kirkland’s proprietary build — keeps documents, custom agents, and embedded knowledge inside one vendor’s walls. Its virtues are security, consistency, and accountability; its cost is lock-in, because the agents and configurations a firm accumulates do not travel. The open posture — emblematized by MCP and Claude for Legal’s connector strategy — positions AI as a layer across whatever the firm already runs, maximizing flexibility and exit options while shifting more integration and governance burden onto the buyer. Kirkland’s variant deserves its own label: closed, but closed in the firm’s favor, with the firm rather than a vendor owning the walls.

Beneath architecture sits pricing, where two venerable models face the same solvent. Legal software has traditionally been sold per seat — every licensed lawyer, a recurring fee — which made sense when software was a tool a human operated. Agents break the logic: when one agent performs the work of five seats, per-seat pricing taxes the customer’s efficiency, and the market is drifting toward platform, consumption, and outcome-based models. The billable hour is the same model wearing a different costume — pay per unit of human effort — and faces the same pressure from the same direction: in-house departments are already reallocating outside-counsel spend toward AI tooling and internal capacity, and increasingly decline to pay hourly rates for work a machine demonstrably compresses. Firms are responding with caution rather than revolution; the hour is not dead, but value-based arrangements are gaining ground matter by matter.

The distributional question — who keeps the surplus when a twenty-hour task becomes two — has four claimants: clients through lower fees, firms through margin, partners through profit, and technology providers through subscriptions. Early evidence suggests the answer is being negotiated engagement by engagement, with sophisticated clients using outside-counsel guidelines to claim their share and elite firms wagering that proprietary capability will let them defend premium pricing on judgment even as production costs fall. What is already clear is that passivity is a position too: firms that neither adjust pricing nor build capability are, in effect, electing to let others decide.

The defining risk story remains the hallucinated citation. What began with Mata v. Avianca in 2023 has become an industrial phenomenon: a public database now tracks on the order of 1,600 court filings worldwide containing AI-fabricated material, and judicial patience has run out. In February 2026 the Ninth Circuit fined two attorneys and suspended them from its bar for six months over fabricated citations and quotations, framing the order explicitly as a warning. The Fifth, Sixth, and Tenth Circuits followed with sanctions of their own. An Oregon federal judge issued the largest penalty in American legal history for the conduct — $110,000 against two lawyers who filed twenty-three invented citations — and the Alabama Supreme Court dismissed a family’s appeal outright after their counsel cited cases that did not exist, a reminder that clients, not only lawyers, absorb the damage. The remarkable feature of the pattern is its persistence in the face of publicity: the failure is not ignorance of the risk but the absence of verification discipline, which is an institutional failure as much as an individual one.

Around that core, a wider risk perimeter is taking shape. Ethics authorities, anchored by the ABA’s 2024 guidance and a growing body of state opinions, have settled the framework: competence includes understanding the tools, AI must be supervised like a nonlawyer assistant, client confidentiality governs what may be entered into which systems, and billing must reflect work actually performed. Privilege adds a sharper edge — courts have begun treating material typed into consumer chatbots as discoverable, with at least one ruling that a litigant’s use of a general-purpose bot waived protection over his case preparation. Malpractice carriers are pricing AI as a foreseeable-error category. Agentic systems raise a liability question no court has fully answered: when an autonomous agent errs, responsibility must be allocated among lawyer, firm, vendor, and lab — and the early presumption is that it lands on the lawyer who deployed it. Above all of this sits a thickening regulatory layer: Colorado’s comprehensive AI act takes effect June 30, 2026, state statutes took force across California, Texas, and Illinois in January, the EU AI Act phases in for multinational practice, and a federal-state preemption fight is live — a compliance matrix that is itself becoming a significant legal practice area.

The employment data refute the simple replacement narrative: legal employment touched a ten-year high of roughly 1.24 million jobs in January 2026, and nearly three-quarters of legal leaders surveyed planned to add permanent headcount in the first half of the year. What the data do show is reshuffling. AI literacy has become a screening criterion — for lawyers with two to three years’ experience, familiarity with AI-enabled research and drafting tools is now an expectation rather than a differentiator — while genuine expertise in AI governance, privacy, and cybersecurity remains scarce enough to command a premium. Demand is tilting toward experienced lateral associates over new graduates, toward legal operations professionals and technology-fluent paralegals, and, in a development without precedent, toward engineers and data scientists hired by law firms themselves: Kirkland’s build alone employs a technology team of roughly 180.

The deepest question is developmental. Associates have always built judgment through the research, drafting, and document review that AI now performs; firms that automate the work without redesigning the training have created a pipeline that leads nowhere. The most deliberate responses are already visible — Latham & Watkins ran an internal AI academy for more than four hundred first-year associates — and law schools are racing to add curriculum. For individuals at every level, the durable advice is a pairing: deep substantive grounding plus genuine tool fluency, including the verification discipline the sanctions cases prove is scarce. Either alone is increasingly common; the combination is what the market currently pays for.

The same facts mean different things depending on where one sits, and fluency in the other chairs’ concerns is itself a professional asset. What follows is the landscape refracted through the seats around the table — each summarized here and expanded in depth in the companion guide, The View from Each Chair.

Sophistication in this market currently consists less in answers than in holding the right questions. Whether to buy, build, orchestrate, or go direct is unresolved because the experiments are months old. Whether the billable hour and the per-seat license — twin tariffs on human effort — will be reformed or merely eroded is unresolved because pricing power is being renegotiated client by client. What constitutes a durable moat is contested between those who believe workflow depth defends and those who believe only proprietary knowledge does.

The leverage pyramid’s future, and with it the training of the next generation of partners, is the industry’s hardest design problem. Whether clients will begin selecting counsel on AI capability — early evidence suggests yes — will determine how much of this is competitive necessity versus discretionary efficiency. Who captures the surplus AI creates is a four-way negotiation among clients, firms, partners, and vendors that has barely begun. And looming over all of it is the categorical question: whether “legal AI” will persist as a software category at all, or dissolve into infrastructure — woven into everything, branded as nothing — with the firm’s own knowledge as the only layer that still bears a name.

The safest forecast is structural rather than specific. The model layer will keep commoditizing; the contest will keep moving up the stack; and the gap between institutional adoption and institutional governance will close, either through deliberate investment or through the involuntary education that sanctions, privilege waivers, and client audits provide. The organizations best positioned for the next phase share three traits that are entirely within reach: they have structured their own knowledge so that any system can use it; they have built verification and governance into daily practice rather than policy documents; and they have made a deliberate, revisable choice about where on the buy–build–orchestrate–direct spectrum they stand.

The technology will keep changing quarterly. The judgment, the knowledge, and the signature on the filing remain the profession’s own — and, properly understood, they are the point of leverage from which everything else in this landscape is moved.