Why AI Has Not Yet Transformed Legal Work
Law firms have spent the past three years experimenting with tools that can summarise documents, draft clauses and answer legal research questions in seconds. Adoption is no longer confined to a handful of innovation teams. According to the American Bar Association’s latest survey, reported use of AI among lawyers rose from 11 percent to 30 percent in a single year. Thomson Reuters found that use of generative AI across the legal profession increased from 14 percent in 2024 to 26 percent in 2025.
Yet the everyday practice of law has not been transformed at the same speed. Many lawyers use public AI tools informally while their firms remain uncertain about approved workflows. Others have access to expensive legal platforms but employ them only for low-risk tasks such as summarising correspondence or producing an initial draft. Thomson Reuters reported in 2025 that 89 percent of law-firm professionals believed AI could be applied to legal work, while almost half still worked in organisations without a formal generative-AI policy.
The gap between interest and institutional use is often explained as conservatism. That is only part of the story. Legal services have structural characteristics that make AI unusually difficult to deploy: lawyers cannot delegate responsibility to a model, confidential information cannot be placed casually into third-party systems, legal authorities change across jurisdictions and a plausible answer can be more dangerous than an obviously poor one.
There is also a commercial complication. If a tool allows a task previously billed at ten hours to be completed in two, the firm must decide whether to pass the saving to the client, change its pricing model or use the additional capacity elsewhere. AI may improve the economics of legal work while undermining the method by which much of that work has traditionally been sold.
A legal answer must be defensible, not merely useful
Generative AI is well suited to producing language that resembles professional analysis. It can organise facts, suggest arguments and create a coherent-looking response from a large body of material.
Legal practice demands something stricter. The answer must be based on valid authority, applicable in the correct jurisdiction and current on the date it is used. It may also need to survive scrutiny from a client, opposing counsel, a regulator or a judge.
This is where apparently fluent systems remain vulnerable.
A Stanford study of specialist AI legal-research products found that retrieval-based tools still produced incorrect or unsupported information. Depending on the product and test, the systems hallucinated in more than one in six responses, despite being marketed as more reliable than general-purpose chatbots. Earlier research found substantially higher error rates when general chatbots answered legal questions.
The practical danger is not simply that AI sometimes produces nonsense. It often produces an answer that sounds legally credible.
A junior lawyer reviewing thousands of documents can make mistakes too, but the firm understands how that person was trained, who supervises the work and how responsibility is allocated. An AI system may provide little visibility into why it selected one authority and ignored another. Its output can also change after a model update, making it difficult to reproduce an earlier result.
Legal AI therefore requires a verification layer that reduces some of the time initially saved. Cases must be opened and checked. Quotations must be compared with the source. Legislation must be confirmed as current. The lawyer must determine whether a case remains good law and whether the model has confused similar legal concepts from different jurisdictions.
For repetitive and bounded work, that process can still be faster than starting from nothing. For a novel dispute, an unusual regulatory question or advice carrying serious consequences, the human review may remain the most valuable part of the task.
Responsibility cannot be automated away
A client may buy software, but it retains a lawyer because someone must exercise professional judgement and accept responsibility for the advice.
The American Bar Association’s Formal Opinion 512 makes clear that using generative AI does not dilute a lawyer’s existing duties. Lawyers must understand the capabilities and limitations of the technology, protect client information, review outputs for accuracy, communicate appropriately with clients and charge reasonable fees.
Courts are taking a similar position. Updated guidance for judges in England and Wales states that the accuracy of information produced by AI must be checked and that judicial office holders remain personally responsible for material issued in their name. It also warns against entering private information into public AI tools.
That principle extends beyond litigation. If an AI system overlooks a change-of-control clause during an acquisition, fails to identify a regulatory filing or generates defective employment advice, the client is unlikely to accept that the model was responsible. The liability will usually remain with the professional or firm that delivered the work.
This creates an asymmetry. The benefit of using AI may be a reduction in time. The downside may be a negligence claim, loss of privilege, regulatory action or reputational damage. Firms will adopt cautiously when a modest efficiency gain exposes them to a disproportionately serious failure.
The challenge is especially acute where the quality of the output cannot be judged easily by the person using it. An experienced competition lawyer may identify a weak antitrust analysis. A trainee or non-specialist may accept it because the structure and vocabulary appear convincing.
The safest systems are therefore not those that remove lawyers from the process. They are those that make it easier for qualified lawyers to find, test and document the underlying evidence.
Confidentiality changes the technology decision
Lawyers work with information that clients would not place into an ordinary consumer application: transaction plans, litigation strategy, personal records, internal investigations, trade secrets and evidence of possible misconduct.
Whether that information can be processed by an AI platform depends on the tool’s contractual terms, security architecture, retention practices, training policies and location of data processing. The firm must also consider privilege, professional secrecy and the data-protection rules applicable to the client and matter.
The ABA has warned that lawyers need informed consent in some circumstances before placing information relating to a representation into self-learning generative-AI software. Its guidance requires lawyers to consider whether submitted information could be stored, disclosed or used to train the system.
The Solicitors Regulation Authority similarly advises firms to assess client confidentiality, supplier responsibility and who the client can turn to when technology causes harm. Its guidance does not prohibit the use of AI, but it leaves regulated firms responsible for making sure that their arrangements remain compliant.
This makes procurement more involved than purchasing a conventional productivity tool. Firms need enterprise security, access controls, audit records and clear restrictions on how client data is used. They may need to prevent certain practice areas from entering information altogether or build separate environments for particularly sensitive matters.
Public AI products can be useful for generic drafting and non-confidential research. The closer a task moves towards real client material, the more the firm needs a governed system integrated with its document-management and identity infrastructure.
That work is expensive. It also favours large firms and corporate legal departments with information-security teams, procurement specialists and substantial technology budgets. Smaller practices may benefit greatly from automation but have fewer resources to evaluate vendors properly.
Legal data is fragmented and difficult to standardise
AI systems perform best when they can draw on large amounts of reliable, accessible and consistently organised information. Legal data rarely meets all four conditions.
Court decisions may be spread across official databases, commercial services and scanned records. Contracts use different terminology even when they address the same commercial issue. Internal legal advice may be stored in emails, document-management systems and personal folders. Matter descriptions and billing codes are often inconsistent.
Even where a firm possesses decades of valuable work product, it may not be ready for AI. Documents can contain outdated advice, duplicated drafts, negotiated compromises or conclusions that were correct only because of facts not recorded in the final memorandum.
Before a system can use this material safely, the firm may need to classify documents, remove duplicates, establish access permissions and identify which versions are authoritative. It must also prevent one client’s confidential material from appearing in work for another.
This is not an AI problem in the narrow sense. It is an information-governance problem that AI makes impossible to ignore.
A model connected to a poorly organised knowledge base can retrieve the wrong answer more quickly. A firm that has never maintained reliable precedent banks will not solve the issue simply by adding a conversational interface.
The billable hour creates a commercial contradiction
AI is often marketed to lawyers as a productivity tool. Clients, however, may interpret productivity as a reason to pay less.
Under an hourly model, efficiency reduces the number of billable units generated by a task. A partner may support AI strategically while remaining uncertain about its effect on revenue, associate utilisation and leverage. The tension is strongest in practices where junior lawyers perform large volumes of review, research and drafting that are then marked up and billed.
Ethics rules make it difficult to avoid the issue. ABA guidance states that lawyers must charge fees that are reasonable in relation to the time actually spent. A firm cannot complete a task rapidly with AI and invoice the client as though a lawyer had performed ten hours of manual work. Nor can it generally charge the client for the lawyer’s basic learning of a technology used across the practice.
This does not mean AI necessarily reduces profitability. A firm can use fixed fees, subscriptions, retainers or value-based pricing. It can handle more matters with the same number of people, respond faster or offer services that would previously have been uneconomic.
But each of those options requires a business-model decision. The technology does not make the decision for the firm.
Corporate legal departments face the inverse problem. They may expect outside counsel to work faster but struggle to determine whether efficiency savings are being passed through. This is likely to increase pressure for budgets, fixed fees and greater transparency over which work was completed by people and which was assisted by software.
AI adoption will accelerate when firms can connect efficiency with a pricing model that rewards the outcome rather than the number of hours consumed.
Training may become the hidden constraint
Law firms traditionally train junior lawyers by giving them research, document review and first-draft work. These tasks are time-consuming, but they also teach lawyers how legal arguments are built and where errors tend to appear.
AI is capable of completing precisely these entry-level tasks.
That creates a long-term question: how does a trainee learn to assess a draft that the trainee would previously have been required to produce? A lawyer cannot supervise an AI-generated contract effectively without understanding why individual clauses matter. Nor can a junior develop judgement solely by editing polished language whose reasoning remains partly hidden.
Firms will need to redesign training rather than simply remove routine work. Junior lawyers may spend less time searching manually, but more time comparing authorities, testing AI outputs and discussing why one argument is stronger than another. They will need earlier exposure to clients, commercial context and the consequences of legal decisions.
That transition will not be automatic. More than half of respondents to a 2026 US legal-industry survey said their firms had provided no responsible-AI training and had no plan to do so.
Without structured training, firms risk producing two undesirable groups: lawyers who use AI without understanding its limitations and lawyers who avoid it because they were never taught how to evaluate it.
Regulation will vary according to the use
“Legal AI” covers systems with very different risk profiles.
Software that summarises a lawyer’s own notes is not equivalent to a system recommending a prison sentence or determining whether someone qualifies for a public benefit. A contract-search tool used internally by a regulated lawyer raises different questions from a public chatbot giving unsupervised advice to consumers.
The EU AI Act reflects that distinction. It classifies certain AI systems used in the administration of justice as high-risk because of their potential effect on the rule of law, individual rights and access to a fair trial. That does not mean every AI product used by a law firm falls into the high-risk category, but it demonstrates why legal applications cannot be regulated as a single group.
The boundary between software and legal service will also become more important. In May 2025, the Solicitors Regulation Authority authorised Garfield.Law as the first law firm in England and Wales designed to provide regulated legal services through an AI-led model. The regulator emphasised that the firm remained responsible for the service and that safeguards had been reviewed before authorisation.
This is a more plausible model for the near future than an autonomous “AI lawyer”. The technology may conduct much of the process, but an authorised entity remains accountable for governance, complaints, insurance and professional standards.
The useful unit is the workflow, not the lawyer
The legal profession is unlikely to be replaced by a single general-purpose system. Individual workflows will change at different speeds.
Document comparison, chronology building, due-diligence extraction and first-pass contract review are relatively bounded. The input can be defined, the output checked and errors escalated. These are strong candidates for automation.
Complex advocacy, negotiation, crisis advice and decisions involving conflicting client interests are harder to reduce to a repeatable process. AI may prepare the lawyer, identify inconsistencies or test possible arguments, but the value still lies in judgement, accountability and understanding what the client is not saying.
This is why firms should avoid beginning with a broad ambition to “adopt AI”. The more productive question is where work is repetitive, expensive and capable of being verified.
A firm implementing contract review might begin with one agreement type, one jurisdiction and a defined set of clauses. It can compare AI findings with experienced reviewers, record false positives and establish when the matter must be escalated. Only after the process is reliable should it be expanded.
The difficult work is not choosing a model. It is defining acceptable error rates, assigning responsibility, integrating the tool with existing systems and deciding how the resulting efficiency will be priced.
AI is already altering legal practice. What it has not done is remove the institutional architecture surrounding legal advice.
The law still requires someone to protect confidential information, verify authority, understand the client’s objectives and stand behind the final answer. Until AI products are deployed within systems that preserve those functions, the industry will continue to move unevenly: rapid experimentation at the edges, cautious adoption at the centre and human responsibility wherever the consequences become serious.
