The Path Forward for Legal AI

AI applications for law are perhaps the most widely debated and deeply funded AI applications aside from software engineering. Law and coding share fundamental similarities that make the relevance to LLMs clear: both have a corpus of associated text and precedent upon which new work draws, and both have relatively clear definitions of success and fidelity. Like AI applications for coding, applications for law have yet to coalesce around a clearly dominant form; frontier labs, legal AI application builders, and existing law firms alike are pursuing a mix of model development, wrappers and harnesses, and agentic infrastructure to best support the work of human lawyers.

Such strategies have frequently made headlines in 2026. For example, Kirkland & Ellis, a multinational white-shoe law firm, announced in May 2026 that it plans to put $500 million of revenue toward building a custom AI platform over three to four years. Latham & Watkins, a top-ranked US law firm, purchased its own GPUs to bring AI compute in-house in September 2026. Harvey, the highest-valued and most well-capitalized legal AI startup, released its own AI model in August 2026 and described its transition from an application layer to a “full stack AI company” the same month, generating skepticism about whether the frontier-model-wrapper business model is sustainable.

Like AI for software engineering, legal AI is outgrowing the subsidization-intensive model of effectively reselling frontier-lab tokens. Both buyers and sellers of legal AI are now considering the importance of data sovereignty, the inclusion of jurisdiction-specific data, and the orchestration of agents as critical drivers of adoption and trust. Legal AI companies have three paths forward to preserve their business models: offering greater data security, more accurate and complete case data, or better agentic infrastructure than either frontier models or legacy players can provide.

Understanding the strategies of AI model developers, legal AI application companies, and law firms in developing and implementing tools for lawyers requires examination of a continually expanding ecosystem. Increasingly, the lines are blurring between buyers and sellers of model tokens, compute, and legal tools for AI as players reposition themselves in light of increased demand, improved models, and a growing emphasis on security.

As of September 2026, this landscape consists of traditional law firms (like Morgan & Morgan or Latham & Watkins), frontier labs (like OpenAI and Anthropic), application layers (like Harvey and Legora), and existing legaltech providers (like Thomson Reuters and RELX). See the Appendix for complete detail on the players in this space and their forays into legal AI products and enablements.

Legal application-layer AI companies are subject to meaningful cash flow disruption based on changing model inference costs and customer usage patterns. At the same time, such tools can change legally admissible billable hours for their users, potentially upending the business model of the law firms that use them.

Seat-based vs Consumption-based Pricing

Seat-based pricing is the most common model for legal AI applications. This contrasts with the ongoing shift in both foundation model and application layer pricing to usage-based; both OpenAI and Anthropic now bill enterprise customers for model consumption on top of a base subscription. This means that per-seat contracts, often locked in for multiple years, provide a predictable revenue stream against unpredictable costs. This dynamic has caused Harvey's margins to decline from 50% to -50% as customer usage increased this year.

Legal AI customers are aware of this shift; the Thomson Reuters Institute Future of Professionals 2026 survey (conducted in March and April 2026) found that 71% of in-house legal professionals expect outside firms to alter how they charge as AI use grows. At the same time, the race to gain and retain customers between legal AI firms is slowing this shift. As Harvey President Gabe Pereyra put it in September 2026:

The End of Billable Hours?

Economic disruption for the firms using legal AI tools is of equal concern to the economics of the companies building it. Virtually 100% of legal firm income is from billable hours, the hours worked per day by lawyers on individual cases, and legal support remains unaffordable for many American individuals and small businesses. As of 2022, low-income Americans had access to either inadequate help or no help at all on 86% of the civil legal problems they faced. As of June 2025, at least one side was unrepresented in 75% of national civil cases.

The extent to which legal AI tools could reduce billable hours and per-case costs is debated; a firm might spend less time on each individual case but take on more cases at once. Wall Street banks are already pressing large law firms for lower fees on the grounds that technology accelerates the work, though 62% say their pricing has not changed because of AI. At the same time, one September 2026 report found that 31% of surveyed firms found AI had improved profitability per matter, and that 30% saw increased matter throughput. Even if some firms are lowering rates or in-house counsel is better-equipped, overall spend on outside counsel did not fall in 2026.

Source: BTI Consulting

Some proponents of AI for legal work have spoken out against billable hours. Crosby's website, for example, argues that billable-hour pricing misaligns a firm's interests with its clients' and describes the fees it produces as grotesque; Crosby’s pricing is set per document at a fixed rate in contrast with billed hours. John Morgan, the founder of Morgan & Morgan, said similarly, “Firms which bill by the hour to draft and review agreements or read through thousands of pages of documents are the practices AI will replace – almost entirely.”

Others in the industry concur. Raghu Ramanathan, the President of Legal Professionals at Thomson Reuters, wrote in summary of the firm’s 2026 report, “The hourly billing model that has defined the industry for more than a century is now being challenged; clients are expecting more concrete results from their firms, and those that cannot deliver quickly will be at risk of being replaced by those that can.” In a formal opinion published by the American Bar Association on AI-use by lawyers, the Standing Committee on Ethics and Professional Responsibility stated that lawyers billing hourly must bill their actual time, quoting Formal Opinion 93-379 that a lawyer who becomes especially efficient still cannot charge for hours not spent. The opinion also states that where a tool lets a lawyer finish far faster, charging the unchanged flat fee may be considered unreasonable.

Given the challenges to economic sustainability for legal AI companies, and the significant hurdle of required benefit to law firms, companies building a sustainable role as tools for lawyers have several paths to differentiation and meaningful value-add to their customers.

One way that legal AI offerings may defend their value to lawyers is via the ability to supply otherwise hard-to-access data which, while public and neither confidential nor proprietary, is neither part of model training data nor easily retrieved by existing models. Access to such data has implications for tool output fidelity, and indirectly on the risks of using such tools. The scarcity of such data as of September 2026 defines the moat and the potential upside for firms willing to pursue access to such data.

Model Fidelity

Practicing law is among the most consequential applications of LLMs, especially concerning hallucinations. Studies in hallucinations of LLMs on legal applications show consistent shortfalls, which can be improved with more complete, human-audited, and Shepardized data.

One study of hallucinations in LLM legal work executed over 800K queries across four models. The study authors defined a hallucination as an output departing from the facts of the world, the frequency of which declined monotonically as the level of court rose: lowest for the Supreme Court, highest for District Courts, across every task and every model (among other determinants of hallucination rate, including factors like geography).

The authors found the highest risk of hallucinations was present for litigants in lower or less prominent courts, those seeking complex information, those whose questions rest on mistaken premises, and those unable to judge how much to trust an answer.

The study also identified a structural trade-off for prompts containing a false statement, which forces a model to choose between fidelity to the prompt and fidelity to the truth, meaning that minimizing one class of hallucination raises another. These authors wrote of the models, “So long as they suffer from gaps in their background legal knowledge… they will be unable to function as reliable sources of legal counsel and advice, no matter how strong their in-context reasoning abilities become.”

Source: Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models

Another study, which evaluated tools based on the completeness of the data corpus available to them, found that access to Westlaw or LexisNexis meaningfully improved hallucination rates. The study compared Lexis+ AI (pre-Protege), Thomson Reuters’ Ask Practical Law AI through CoCounsel, and Westlaw AI-Assisted Research, against GPT-4 as a comparison. The authors treated a response as hallucinated when it was either untrue or claimed that an authority backs a proposition it does not (i.e., the definition rests on two separately coded variables, correctness and groundedness, with groundedness scored only for answers already judged correct). Measured hallucination rates across the full query set were 0.43 for GPT-4, 0.33 for Westlaw, 0.17 for Lexis+ AI, and 0.17 for Ask Practical Law AI.

The study noted that while improved, these hallucination rates are still higher than model providers advertise. The report noted that LexisNexis claims “Unlike other vendors, however, Lexis+ AI delivers 100% hallucination-free linked legal citations connected to source documents, grounding those responses in authoritative resources that can be relied upon with confidence.” Similarly, a Thomson Reuters official said in 2023, “We avoid [hallucinations] by relying on the trusted content within Westlaw and building in checks and balances that ensure our answers are grounded in good law” and that RAG “dramatically reduces hallucinations to nearly zero.”

Source: Stanford

Sanctions Resulting from AI Mistakes

When selecting such a tool, the consequences of making the wrong choice can be meaningful. As of September 2026, hallucinated material has surfaced in at least 1,395 US state and federal cases, according to a researcher who maintains a tracker site. These include invented citations, false quotations, misrepresented authority, and outdated law (this record counts only the findings of courts, making it likely an undercount).

As the adoption of AI tools for law increases, the number of impacted cases rises in tandem: within a three-day stretch in September 2026, a New Mexico defense lawyer was rebuked over invented police witness testimony in a murder case; Oklahoma prosecutors disclosed that an order signed by a judge there had cited fake cases; and in Los Angeles a lawyer for State Farm was fined over erroneous citations spread across seven filings.

The consequences of incorrect material in case filings vary, and include fines (one such fine was over $100K), recommendations for professional discipline, and multi-year bans from practicing law. Courts do not have to establish that AI was involved before sanctioning a lawyer who cites bad authority; longstanding candor rules apply regardless of whether models created the fabricated citations.

More recently, courts have even considered sanctions for use of AI even if filed casework does not contain any hallucinations. In the case of a Florida lawyer’s filed casework, the 4th DCA. Judge Robert Gross wrote, “An AI supercomputer would struggle to find meaning in some of the prose used in this case,” and ordered the lawyer to explain within 10 days, without using AI, why she shouldn’t be sanctioned or referred to the Florida Bar.

Courts across the country are considering formal guidance on the use of AI in law. The Alabama State Bar, for example, issued guidance instructing lawyers to use enterprise or closed AI systems instead of free versions in the interest of client data security. Senator Chuck Grassley of Iowa asked judges across the country to put formal safeguards in place for AI use inside their chambers after multiple judges were found to have authored rulings using AI. As of September 2026, few courts ban AI use outright (the exception being Judge Sharon Johnson Coleman of the Northern District of Illinois

Data Scarcity

Given the implications of errors in legal casework, it is unsurprising that securing complete and accurate legal reference material is a priority for firms. LexisNexis and Westlaw are used by most lawyers, and integration with the data that these tools provide is critical for making AI useful for caselaw. As Paul Graham put it in an X post, “And if ever there was a territory you could defend against the model companies, law is it.” Accessing such data is not possible using the internet alone, because while legal rulings are public, not all proceedings are centrally organized. In the US, for example, the versions of rulings that lawyers must cite sit in printed volumes that LexisNexis and Westlaw control. Official reporter page numbers and pincites matter because lawyers must cite them.

Harvey integrates data from LexisNexis directly, via an “alliance” integrating statutes, case law, and citations directly into the platform, though LexisNexis has its own tools. Based on reporting from the two companies about the integration, however, it seems that this change allows lawyers to use LexisNexis AI tools and models through Harvey (not for other models to access the same database):

Legora, on the other hand, is taking a more ambitious path to own and offer access to this data itself, which CEO Max Junestrand has described as “the hardest thing in legal AI.” The platform is already distinguished from other AI tech offerings by its integration of official sources of jurisdiction-specific legal data, including US federal/state case law, continuous regulations, the SEC's EDGAR database, Italian legal record from Tinexta Visura, and UK company law guidance via FromCounsel. Legora is taking this distinction a step further by integrating this data into the platform directly themselves, eliminating the need for external databases or partnerships.

The company has said openly, “In the US, access to citable case law is constrained by the companies that own the official reporters and won't sell a digital feed. In Germany, the challenges are digitization and anonymization. But there is always a way, and we’re doing whatever it takes to bring together every data type in over 100 countries around the world. We are bringing all the world’s legal data into Legora.” Legora staff get citable case law into the system by cutting spines off books, removing publisher-written headnotes, scanning pages, and then checking the scans for accuracy. The platform’s repository of case law is updated every 24 hours, includes page numbers and pincites, and is manually verified through a dual-keying process. This process, while meaningfully more resource-intensive up-front, means Legora will not have long-term financial commitments to data providers.

It is worth noting that both integrations (Legora building the structured database for caselaw and Harvey selling access to it through LexisNexis) retain the scaffolding of case law, which is critical for helping AI to interface with it correctly. Legora’s head of legal research has described over 50 failure modes for legal AI tools, including AI citing a rule but missing an amendment modifying it, quoting a dissent as a holding, or treating a non-binding agency decision as equivalent to a Supreme Court ruling. The company describes using “proprietary technology” from 2026 acquisitions Wexler and Qura to build this scaffolding.

In tandem with the growth in AI application resources and capabilities, however, come rising concerns about internal data, model, and compute sovereignty. Firms considering ongoing AI-enabled security breaches are increasingly concerned about protecting legal commitments to clients and protecting firm IP.

Internal Data

Obvious among these considerations is the protection of legally confidential client information. While OpenAI has said specifically that eligible users of Astra for Law will have Zero Data Retention API usage and ChatGPT Enterprise usage will be excluded from human review, Anthropic, for example, requires 30-day storage of enterprise data (even if stored on customer cloud instances). According to one vendor in legal tech, ‘the buying question has shifted from "which model is smartest" to "who can prove our data never leaves.”’

Internal Models

As model usage and SOTA-model token prices increase, the economic justification of alternatives is clear, especially as the performance of those models approaches that of frontier models. For many AI-application layer companies, the choice is to build a model in-house as diversification from the token-resale business model. Abridge, a healthtech startup, announced in June 2026 that it is building a bespoke clinical foundation model trained on Nvidia’s open models. Decagon, an AI customer support startup, said 80% of its query volume now runs on models it owns. Ramp and Rogo are each similarly exploring training proprietary models for the first time.

At the same time, the pivot to training and owning a company’s own model is a defense against being cut off by model developers looking to expand into the legal space (OpenAI, for example, said it was cutting off Cursor’s model access about a week after SpaceX closed its purchase of the startup, though this represents cutting off against a rival model developer).

While it is unlikely, due to attorney-client privilege, that client data will be used to train in-house models, there can still be benefits for these firms from training models on their own data. Multiple legal AI benchmarks exist, and at least one study has found that legal-specific AI models outperform lawyers on certain tasks.

In August 2026, Harvey announced the release of its own model, Tenet, developed from the open-weight Kimi K3 model from the Chinese developer Moonshot AI and post-trained jointly with Fireworks for long-horizon legal work. The company claims that Tenet ranks first on its own contracts benchmark and second on its legal benchmark overall. Harvey has clarified that it cannot draw on its clients’ confidential legal material to train models and that no customer data entered any part of its post-training; the company says that it buys the domain expertise it needs from Mercor instead.

This is Harvey’s second exploration of an owned model (Harvey tried creating its own model in 2024, but frontier models quickly surpassed it even in law, making the resources dedicated to training that model unnecessary when they could have been otherwise used on marketing and customer acquisition/support). In an interview about the launch, Harvey President Pereyra said “Until late last year, AI applications’ performance mostly depended on the models underneath them” and that training those base models was an extremely expensive effort usually left to the biggest AI labs. Any efforts to train custom models weren’t so meaningful in comparison.”

Legora has also begun to explore post-training of open models, saying “We post-train when we know it buys our customers better performance on a specialized task. Training is a tool we reach for when it helps, nothing more than that.” As of September 2026, Legora has not released its own model.

Internal Compute

Finally, there have been several breaches of retained AI model usage data (like individual chats), the risk of which may be intolerable for law firms, motivating companies to consider a combination of owned hardware and owned models to run on-premise instances of AI.

As noted above, Latham and Watkins reportedly acquired “several” servers over “recent years”, each holding multiple GPUs, though the exact scale of the buildout is unknown; the firm also has not said what it has spent or expects to spend on the approach. The company also reportedly rents space in a data center facility kept locked and entered only by the firm’s own staff. “Sometimes we may have information that is so sensitive, client information that we really want to protect, we don’t want to put it to any cloud vendor,” said Rene Mendoza, the chief information officer of the company. Independent analysts estimate that the hardware, plus the specialists needed to run it, will come to tens of millions of dollars annually.

The in-house team developing models for the firm is reportedly building from Nvidia Nemotron 3 open-weight models, and has over 900 technology specialists on staff as of September 2026. The decision to develop models in-house represents a distinction from large law firms that avoid AI capital expenditure because it reduces the profit pool distributed to equity partners each year. The firm is testing open-weight models it can download, customise and run on its own hardware, framed as an in-house alternative to cloud tools and as mattering most for the client information it treats as most sensitive.

The third moat that legal AI providers are increasingly apparently considering is the provision of agentic infrastructure for law firms, enabling automated AI assistance for a variety of workflows.

In September 2026, several weeks after the announcement of Tenet, Harvey shared that it is effectively walking away from SaaS to do custom implementation for customers, saying, “The transition we are starting to make as a company is from an application layer company to what we call like a full stack AI company… about a year and a half ago, first with agents, people realized that the infrastructure layer of agents… the infrastructure you need to build and serve that is actually pretty complicated.” The company rolled out agents in March 2026, and is now hiring members of implementation teams to expand the company’s footprint within the firms it works with.

Legora is considering a similar path, arguing in April 2026 that “the software model is beginning to shift from SaaS (Software as a Service) toward AaaS (Agent as a Service).” In June 2026, the company announced the release of LegoraAgent, which the company describes as “a generational leap from assistive AI to agentic execution,” claiming “the Legora Agent works with you and for you - while you're with clients, your kids, or even while you sleep.” The announcement of the release highlights automated due diligence, pleadings matrix construction, and review of regulatory releases as tasks the Agent can complete.

Whether legal casework, data and model security, or agent infrastructure becomes the dominant advantage for AI applications, the firms fighting for market share in this space will continue to contend with dynamics beyond their business models.

Consumer Legal Advice

Consumers are not likely to comprise a meaningful share of the market for these tools given the requirements for giving legal advice to individuals. One dissenting line of scholarship responding to this argument holds that a model cannot commit unauthorized practice because it is not a person able to exercise judgment. A second dissenting line treats AI output as informational in the same way a printed guide to court procedure is. Existing law-adjacent platforms have encountered pushback for similar reasons. LegalZoom's automation of routine work such as trademark filing paperwork, for example, has drawn unauthorized-practice suits repeatedly over more than a decade. The nonprofit Upsolve started training volunteers to help defendants with legal forms, then pre-emptively sought an injunction on the theory that such help is First Amendment-protected speech (which was granted).

Whether there will be a role for these tools to give legal information to consumers explicitly, either as aides to attorneys or used on their own, remains to be seen. Creating an intermediate licence tier for legal service providers is among the fastest-growing categories of unauthorized-practice reform, with seven states having adopted it and ten more weighing it.

Ethics of Legal AI

The considerations around directly or indirectly giving legal advice based on AI output is still debated. Specifically, AI alignment critics have argued against the position that an AI acting for you should be loyal to you and nothing else, and takes the human professional as its model: a lawyer owes you confidentiality and loyalty but is also an officer of the court, with duties to the tribunal that override those owed to the client, and in some circumstances is obliged to turn on the client. Legal ethics codes treat instruction from the client as no excuse for a violation, meaning lawyers are not unambiguous agents of the people they act for. An American Bar Association opinion piece on the topic states that "Humans demand human accountability for life-altering decisions."

An Evolving Profession

Finally, the adoption of AI tools hinges fundamentally on demand from lawyers, which has been strong as of 2026. In the Thomson Reuters survey of legal professionals from April 2026, 77% of surveyed legal clients said it was “very important or essential” that the firms they hire deliver AI-derived quality improvements, though only 5% said most or all of their providers actually deliver it, and 36% of legal professionals who have used professional-grade AI tools said they “would categorically decline a job offer that did not offer them.” The report also found that 38% of legal professionals are under “financial pressure” to adopt AI quickly, and 22% expected to see “financial consequences from moving too slowly on AI within the next 12 months.”

At the same time, the impact of AI on industry personnel is debated. NALP data from August 2026 showed that the largest firms (above 500 lawyers) took 7.5% fewer first-year associates out of the Class of 2025 than out of the year before it, the first decline in big-law entry-level hiring since 2014. One report found most firms intended through 2027 to favor hiring experienced associates away from rivals over taking on entry-level graduates. 78% of law firm professionals say early-career lawyers depend on experienced mentors to build exactly the skills AI is absorbing.

Application Layer

Harvey

Harvey, a legal AI startup founded in 2022, is the most established application layer company in the space, with over 500 employees and a valuation of $15.6 billion as of September 2026. Since inception, Harvey’s business has been built around adapting frontier lab models to the specialized work of lawyers, providing a compliance-certified system for document review, legal research, drafting, and risk management. The company claims that 80% of Am Law 100 firms were using its product as of September 2026, and claims five of the Fortune 10 among its in-house customers. The company has used a human implementation engineer in every single deployment (with 180 such employees, most former lawyers) and has integrated with both in-house legal teams and external law firms.

In March 2026, updates to Harvey’s agents and improvements to the underlying models they drew upon generated a steep increase in customer usage; the company said token consumption has risen twenty-fold in 2026 compared to 2025. Harvey uses a per-seat pricing model with no token or output limits on usage, meaning its costs to service customers are variable even though its revenue is fixed; customer contract durations are not public but are thought to span multiple years. Improved models carrying increased costs per token also enabled increased token usage (as models can handle larger and more complex tasks).

In an interview, Harvey CEO Winston Weinberg said that, as of May 2026, the company was consuming 13 trillion tokens per month. Assuming that the company was using a blend of models, ranging from $1 to $10 per million input/output/cached tokens, implies roughly $13 to $130 million per month in inference costs, or $156 million to $1.56 trillion. According to Weinberg, Harvey's ARR was around $300 million at the time. One person familiar with the circumstances shared that Harvey’s gross margin had moved from roughly 50% at the start of 2026 to minus 50% by June 2026, suggesting a tripling in inference costs. These costs were likely unrecoverable with the company’s business model at the time; Harvey and Legora have reportedly engaged in aggressive discounting to compete in winning multi-year contracts with Big Law firms (some even reportedly being free).

This shift has pushed Harvey to explore multiple channels for not only decreasing costs but transforming the business entirely, providing a playbook for the various ways in which the industry is shifting. These shifts include creating its own models and pivoting to provide a full-stack agentic-infrastructure product for law firms scaling their own AI implementations. Harvey’s President shared that the company saw positive margins again in September, in part due to these changes.

Legora

Founded in 2023, Legora is a Swedish legal AI company that sells software to large law firms and to in-house legal departments. Serving primarily European law firms, the company originated with Y Combinator before embedding inside business law firm Mannheimer Swartling for several months to study how legal work is done and iterate on the product. In April 2026, Legora announced a $50 million extension of its earlier Series D, leading the round to $600 million in equity at a $5.6 billion post-money valuation. The same month, the company announced that it had passed $100 million in annual recurring revenue, less than 18 months after beginning general availability (as of September 2026, the company is now at $200 million in ARR)

Similar to Harvey, Legora offers a generative AI-powered operating system for both law firms and in-house corporate legal departments. The product is designed to automate repetitive, data-heavy processes by connecting large language models (including those from frontier labs) with internal firm databases and public legal sources within a highly secure environment. The company cites customer retention of 95%, NRR above 300%, and DAU/MAU above 50%. Despite strong traction, the company has trailed Harvey in ARR and valuation, and has turned to similar measures to diversify its offerings, including creation of a bespoke jurisdiction-specific case data offering and beginning to post-train models.

Adjacent Application Layer Companies

While Harvey and Legora are the biggest players among firms offering AI-assisted legal work, a number of other companies offer AI-enabled efficiencies for law firms in case management and administration, some of which are new and AI-native, while others have adapted from pre-AI offerings. While these companies likely have lower ceilings with regard to the total volume of work they might supplement, they are also likely less exposed to the economics of growing token prices at fixed pricing. Understanding these offerings completes the picture of AI applications for law.

• Filevine is a legal case-management platform built around LOIS, the company’s Legal Operating Intelligence System, which offers AI-assisted contract management, document assembly, intake, and billing. Filevine frames its differentiation around data and scope, claiming that LOIS is built on a legal graph of more than 40 million legal matters and runs AI agents across a firm's entire caseload rather than on one task for one user. Because the AI is embedded in the case-management system, lawyers do not need to upload client data to an external tool. Filevine serves over 6.5K firms, including those at boutique firms, government agencies, and Fortune 500 companies, and reached $250 million in ARR in April 2026.

• Relativity is a platform for AI-assisted e-discovery, investigations, case strategy, second requests, early case intelligence, contract review, breach response, and FOIA requests. The company claims that 198 firms out of the AmLaw 200 use the platform and that its customers include over 300K legal professionals across 40 countries. Relativity frames the platform’s architectural differentiation as bringing the model to the data instead of copying sensitive material out to a separate AI tool; it says its Legal Hold module preserves data where it sits and holds preservation material in an Azure environment. The company bills based on committed data volume rather than seat count, offered as pay-as-you-go or discounted one- and three-year commitments.

• Intapp is a company-level law firm administration platform, distinct from AI that does legal work itself. The company says it serves more than 1.4K firms that each generate over $50K in annual recurring revenue, claims 97 of the AmLaw 100 and 17 of the 20 largest accounting firms among its customers. The platform’s time-capture AI reconstructs a professional's day from emails, meetings, and documents and drafts compliant entries without manual input, making the professional labor billing process easier.

• Darrow is a legal-AI startup that positions itself as less of a legal workflow tool and instead as a Legal Exposure Management tool, or an upstream detection system that reads signals of legal risk before a lawsuit exists. The company says that its core research asset is a structured knowledge base linking laws that impose obligations, the organizational failures that appear when obligations go unmet, and the enforcement that follows such failures. The company offers three product lines: Case Origination, Predictive Litigation Analytics, and Strategic Case Advancement.

• Spellbook is the builder of a legal-AI plugin that runs review, redlining, and drafting inside Microsoft Word. The company claims more than 5K commercial legal teams across over 80 countries as customers, pitching itself as a substitute for outside counsel in routine contract drafting and review processes.

• Ironclad is a contract lifecycle management (CLM) platform built for corporate legal, procurement, finance, and sales teams. The platform manages the life cycle of contracts, including intake, drafting from templates, internal approval routing, AI-assisted redlining, e-signature, and post-signature obligation tracking. Ironclad introduced AI tooling in March 2026 to answer questions and launch built-in agents, such as a Renewal Agent and a Cost Savings Agent that flags volume discounts and rebates before a renewal; the company claims that over 65% of its customers have adopted its AI features. The company says it has passed $200 million in annual recurring revenue, and names Rivian, AMD, and L'Oréal as customers.

• Clio is a legal startup that offers case and matter management, time tracking, billing, client intake, and documents, mostly for individual lawyers and smaller firms. The company states that its software is used by more than 400K legal professionals across over 130 countries and holds over 100 bar-association approvals covering every US state. While caseload management is the company’s primary offering, it has added AI features and has moved into research, spending $1 billion to acquire vLex, the legal research company behind the Vincent AI assistant.

Frontier Labs

OpenAI

OpenAI offers a suite of solutions for legal teams, built around the Astra for Law, the legal-sector edition of GPT-6 Astra released in September 2026. OpenAI has described Astra for Law as designed to let firms and legal software vendors run research, prepare advice, and build their own applications for lawyers, pairing the specifically trained model with an index of American statutes, regulations, decided cases, and further legal source material.

OpenAI has not detailed specifically how the model differs from other versions of GPT-6, but said the model contains purpose-written guidance for legal reasoning and drafting. The company reports that at maximum reasoning effort, Astra for Law cleared the overall correctness test on 54.0% of questions (against 38.7% for standard GPT-6 Astra). The announcement page states no hallucination or citation-error rate.

OpenAI has not published any token- or seat-based pricing for the model, which was limited at launch to selected US law firms via the company’s Trusted Access Program. OpenAI says eligible firms will experience Zero Data Retention on the API, and that for those firms ChatGPT Enterprise usage will be excluded from human review by default.

The legal source material that Astra for Law draws from comes from CourtListener, a non-profit case database consisting of 230 million URLs covering five kinds of American legal documents: decided case law, administrative decisions, court rules, regulations and statutes. OpenAI puts the CourtListener collection's reach above 99.9% of case law that is both published and precedential. It is worth noting that CourtListener does not contain unpublished and non-precedential decisions, which are collectively a larger share of overall case law. In recent years, for example, 12% to 15% of federal courts of appeals’ merits decisions are published as precedential opinions, leaving the remaining 85–88% to be issued as unpublished, nonprecedential decisions.

Source: Administrative Office of the US Courts' annual Judicial Business

OpenAI has said that Astra for Law will connect to software suppliers including Relativity, Clio, Intapp, and Thomson Reuters, and that the model was developed by building legal applications with firms including Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins, and Wachtell Lipton. The launch of Astra for Law referenced partnership with Harvey and Legora, both of which offer access to OpenAI models; the announcement named Harvey and Legora as API customers that will be in a position to build the model into their own products and workflows (It has been speculated that the release of Astra for Law after Harvey’s September 2026 fundraising round but before Legora’s anticipated fall round was calculated given OpenAI’s investment in Harvey).

Anthropic

Anthropic has offered lawyer-oriented tooling for Claude since January 2026 in the form of Claude for Legal plugins, which enable integration with firm document management systems. Anthropic has not mentioned specific data resources available in conjunction with Claude for Legal. Instead, it includes tooling for commercial, corporate, IP, litigation, privacy, employment, regulatory, and AI governance, and academic law school, as well as over 90 specialized, named agents for multi-step tasks like vendor agreement reviews, data subject access request (DSAR) responses, termination reviews, and claim chart building. It also includes integrations to tools like DocuSign, Ironclad, iManage, Everlaw, CourtListener, and Thomson Reuters, and skills for drafting, redlining, contract review, and document comparison inside Microsoft Word and Office apps. As of September 2026, the plugins are available for open-source use (not limited to law firms) with paid Anthropic subscriptions.

Law Firms

Latham and Watkins

Latham and Watkins, the second-largest law firm in the United States, has begun purchasing Nvidia GPU servers to tailor and run AI models for legal work in-house. The combination of client confidentiality requirements and the ongoing shift to consumption-based AI pricing justified owning the hardware and fine-tuning open-weight models for the company; the Financial Times called this instance the first case made public of a large firm buying AI hardware outright and adapting models on it.

Kirkland and Ellis

Kirkland and Ellis, another prominent US law firm, said in May 2026 it will put $500 million of revenue toward building a custom AI platform over three to four years. Kirkland has said that the platform architecture it is building allows users to switch models, and the firm employs over 180 AI engineers and data scientists as part of the project, 50 of which dedicated AI engineers.

Though Kirkland and Ellis declined to specify whether it has purchased GPU hardware either outright or through a subsidiary, the company has posted a job description since the announcement seeking an “AI Infrastructure Director” to manage the company’s “on‑premise Graphics Processing Unit (GPU) clusters, Microsoft Azure AI and Machine Learning (ML) services, and shared AI platform components - with accountability for reliability, scalability, and lifecycle management.”

Morgan and Morgan

Morgan and Morgan, described as the largest personal injury practice in the United States, said in September 2026 that it has pledged at least $1 billion to “AI and technology” over the coming decade. This spending will be specifically dedicated to the firm’s MX2 technology platform, on which it had spent $300 million as of September 2026. Morgan and Morgan said it uses MX2 to pull details out of medical records, produce case paperwork, and prepare for trial, and that it intends to eventually open MX2 to other practices, including corporate and transactional law, on an invitation-only basis before the end of 2027. The firm did not say how it intends to price access to MX2.

Crosby

Crosby is a hybrid law firm that, while newer and smaller than other firms listed in this section, is worth noting in light of its unique approach. The company presents itself not as software sold to lawyers but as a registered agentic law firm with barred attorneys and malpractice cover that happens to run AI agents internally. The firm markets its product as finished work requiring no software onboarding, offering fixed rates per document as a replacement for billable hours.

Crosby argues that law is the largest AI-transformable market after software engineering. As of September 2026, the company’s own scope is deliberately narrow, limited to commercial contracts aimed at go-to-market teams whose deals stall in legal review, and has worked with companies like Cognition, Ramp, Clay, Cursor, Bilt, Rogo, and Granola, among others. The company claims that its agents learn from each document review which proposed edits are least likely to stall a negotiation; it says review times for Cursor fell by roughly half as it accumulated and encoded Cursor's preferences, recurring customer objections, and escalation thresholds.

Existing Legal Tech

In order to put the efforts to establish complete case law data repositories in context, it is critical to understand the role and providers of case law source material today. In the US, these are principally Thomson Reuters (the provider of Westlaw) and RELX (the provider of LexisNexis).

Westlaw and LexisNexis are searchable databases of primary law, consisting of centuries of federal and state court opinions, statutes, regulations, court rules, and administrative decisions. Attorney-editors add headnotes and classify legal points into topic systems like Westlaw's Key Numbers, making narrow issues easy to find. Citators used by the databases show whether a case is still considered “good law” or in current legal standing. The databases also offer proprietary secondary sources such as treatises and practice guides, hard-to-find materials like unpublished opinions, briefs, and dockets, and public records, news, and company filings used in investigations and due diligence.

Alternatives exist, but are either incomplete in case law coverage or indexing. CourtListener's PACER mirror, for example, covers federal filings and dockets, but is not a case law database. The Caselaw Access Project covers published cases through 2018, but does not include recent and unpublished opinions. Google Scholar has no reliable citator, preventing searchers from confirming whether a case is still good law. Bloomberg Law and vLex/Fastcase are the closest rivals, but their secondary-source libraries are smaller, and Fastcase's citator relies mainly on algorithmic flagging rather than editorial review.

Thomson Reuters Westlaw

In 2023, Thomson Reuters began its expansion into AI offerings, acquiring the startup Casetext for $650 million in August 2023. This acquisition came five months after Casetext launched CoCounsel as the first generative AI legal assistant built on then-state-of-the-art models. Thomson Reuters claims that one million professionals across 107 countries depend on CoCounsel.

In August 2026, Thomson Reuters relaunched CoCounsel Legal as what it calls a “fully agentic AI experience” equipped to handle document review, research memos, deposition preparation, and contract analysis. Within the tool, multiple models are available, including Thomson 1.0, a model the company said was trained on its own legal research material and rests on open-source technology. Thomson Reuters says the CoCounsel Legal system is built on Anthropic’s Claude Agent SDK, though it is not clear where inference is hosted.

RELX LexisNexis

In January 2025, LexisNexis announced the US general availability of Protégé, a personalized AI assistant with an agentic layer that, according to the company, can complete tasks using LexisNexis case law on its own once given a goal, without step-by-step direction, and can review its own output to find ways to improve it. LexisNexis says Protégé can draft complete transactional documents as well as litigation motions, briefs, and complaints, and checks this work itself before handing it to a lawyer. The tool works with models from both Anthropic and OpenAI.

Protégé replaced an earlier RELX AI Integration called Lexis+ AI in February 2026, calling the new offering Lexis+ with Protégé, which takes the form of a single prompt-box platform for over 300 workflows. LexisNexis has not published hard adoption figures, limiting comparison with tools like Harvey and Legora. It is worth noting that LexisNexis and Harvey formed an alliance in June 2025 to bring Protégé, US primary law, and Shepard's into Harvey, and Lexis's parent RELX is a Harvey investor. Protégé is most comparable to Thomson Reuters' CoCounsel, which follows the same incumbent model grounded in Westlaw case law and Practical Law guides. Many firms run two tools: a general platform like Harvey for broad workflows, plus CoCounsel or Protégé for cited research.