AI and legal tech: what the tools do, which rules apply and how to choose

The short version

• Artificial intelligence (AI) in legal tech supports research, drafting, document review and managing client work; the lawyer stays responsible for the result.
• In a peer-reviewed study of versions available in spring 2024, three AI legal research tools from two major legal publishers hallucinated between 17% and 33% of the time.
• Professional rules and guidance apply duties lawyers already have; California’s SB 574 adds statutory ones, such as verifying citations and disclosing AI use to courts, from 1 January 2027.
• The Upper Tribunal says putting confidential client documents into a public AI tool breaches confidentiality and waives privilege, and the Solicitors Regulation Authority (SRA) warns that privilege may be lost for good.
• Under the EU AI Act, providers and users of legal AI tools have AI-literacy duties and chatbots carry transparency duties; AI that helps a court or an alternative dispute resolution (ADR) body decide a case with legal effects can be high-risk from 2 December 2027.^{1,4,5,6,7,11,20,30,31}

What is legal tech, and where does AI fit?

AI and legal tech now overlap: many tools lawyers already use, such as research platforms, document review systems and practice management tools, include some AI (State Bar of California guidance, p. 1). The Law Society of England and Wales says “legal tech” usually means technology made for solicitors and other legal professionals. It prefers “lawtech”, a wider term that also covers everyday tools and access-to-justice tools such as chatbots.^1,7

The Law Society’s practice note also separates the practice of law, such as research and drafting, where professional duties apply most directly, from the business of law, such as billing and administration, which are not regulated legal activities.¹

Not all legal AI is generative. Formal Opinion 512 of the American Bar Association (ABA) calls e-discovery, where technology-assisted review sorts the electronic documents in a case, a well-known use (p. 1). Generative AI, which creates new content such as text, is non-deterministic: the same prompt can give different outputs, so testing it once does not guarantee it will behave the same way again.^1,2

Agentic AI can perform tasks or workflows, such as client intake, without human prompting. The State Bar of California’s 2026 guidance says lawyers must not let such systems make substantive legal determinations, give legal advice or file pleadings without meaningful lawyer supervision and review (p. 2). That keeps a human in the loop.⁷

The Law Society warns that a vendor’s own “legal-specific” label does not guarantee that a tool is appropriate or safe for legal practice. Our guide to AI in law covers how courts use AI and how the law regulates it.¹

How are AI tools used in legal work?

Lawyers use legal AI tools for research, drafting, reviewing and summarising documents, e-discovery and client intake, among other tasks. The first five items below follow the Law Society’s practice note of 21 October 2026, with its headings in brackets.^1,9

• Legal research: finding case law, legislation and commentary (“Legal research assistance”).
• Drafting: first drafts and routine documents such as standard agreements (“Legal document creation and automation”).
• Contract review and due diligence: picking out non-standard terms and risks in contracts and in large sets of documents (“Contract review”; “Due diligence and risk assessment”). We cover AI in contract drafting to prevent disputes separately.
• E-discovery: sorting and prioritising large volumes of documents, a use the Law Society calls widespread (“E-discovery and e-disclosure”).
• Summaries and transcription: case files, correspondence and meetings, reviewed by a person (“Document summarisation and transcription”).
• Client intake and chatbots: The Florida Bar’s advisory Ethics Opinion 24-1 says a lawyer’s chatbot must tell people it is an AI program and not a lawyer (pp. 1, 7).
• Arbitration and online dispute resolution: AAA’s opt-in AI arbitrator drafts awards in two-party, documents-only construction cases, and a human arbitrator reviews, may rewrite and signs each award (from 3 November 2025; R-8). Our separate guide explains how institutions, ombudsman schemes and online dispute resolution use AI.^1,9,27,28

How many lawyers use AI tools?

In the American Bar Association’s 2024 Legal Technology Survey, 30.2% of lawyers said their offices used AI-based tools, rising to 47.8% at firms of 500 or more lawyers. Use was lower at smaller firms: 29.5% at firms of 10 to 49 lawyers, 24.1% at 2 to 9 and 17.7% of solo practitioners. These figures come from the survey’s online research volume (512 respondents, October to December 2024).³

Asked for the most important benefit, 54.4% chose saving time or increasing efficiency. Accuracy was the concern named most often (74.7%), ahead of reliability (56.3%) and data privacy and security (47.2%).³

In the 2025 Queen Mary University of London survey, 64% had used AI for factual and legal research in the past five years (printed pp. 27, 37). Our guide to AI in arbitration covers how arbitrators and institutions use AI. We leave out adoption surveys published by legal technology vendors (how we research).¹⁶

How accurate are AI legal research tools?

AI legal research tools hallucinate, giving false or unsupported answers, less often than a general-purpose chatbot, but they are less reliable than their makers claimed. In a preregistered, peer-reviewed study, three AI legal research tools from two major legal publishers hallucinated between 17% and 33% of the time (Journal of Empirical Legal Studies, 2025); it tested versions available in spring 2024.⁴

Across 202 queries registered in advance (pp. 9, 16), the most accurate tool gave correct, properly grounded answers to 65% and the other two to 41% and 19% (p. 10). The authors found providers’ claims to avoid or eliminate hallucinations “overstated” (p. 1).⁴

Courts treat unchecked output as the lawyer’s failure: in Ayinde, [2025] EWHC 1383 (Admin), the High Court of England and Wales held that lawyers must check AI research against authoritative sources ([7], [8]), and in Mata v. Avianca (22 June 2023) a New York federal court found nothing “inherently improper” in using a reliable AI tool but fined the lawyers and their firm US$5,000 for submitting non-existent opinions.^17,18

What do professional rules say about lawyers using AI tools?

Professional rules mostly apply the duties lawyers already have to AI tools. ABA Formal Opinion 512 (29 July 2024) applies the duties of competence, confidentiality, client communication, candour, supervision and reasonable fees to generative AI.^2,5,6

Under Opinion 512, a lawyer needs “a reasonable understanding of the capabilities and limitations” of a tool, not expertise (p. 2). How much checking is enough depends on the tool and the task (p. 4), but output must be reviewed before it goes to a court (p. 10).²

California’s SB 574, approved on 30 September 2026, takes effect on 1 January 2027. It adds section 6068.1 to the Business and Professions Code. An attorney “shall not delegate the practice of law to generative artificial intelligence” ((a)(2)) and must take reasonable steps to verify outputs, including all case and statutory citations ((a)(3)(B)).^5,6

The Florida Bar applies its standards for supervising non-lawyer assistants to generative AI, including oversight policies (Opinion 24-1, p. 4). The New York State Bar Association’s 2024 task force recommended prioritising education for judges, lawyers and regulators (printed p. 9).^9,10

In England and Wales, the SRA’s warning notice of 17 August 2026 names two concerns: hallucinated content in court and other documents, and risks to client confidentiality in paid and free AI tools. The Law Society notes that legal practice has no separate AI regulator; existing duties apply. The Council of Bars and Law Societies of Europe (CCBE) warns that bias, sycophancy and “automation complacency” can undermine a lawyer’s professional objectivity (p. 20).^1,11,13

Courts add their own filing rules, some requiring disclosure of AI use. Since 15 June 2026, Florida’s Rule 2.515(d)(2) makes whoever signs a filing represent that the legal authorities identified “exist and are accurately cited” (SC2026-0673, 28 May 2026). From 1 January 2027, California also bars citations the responsible attorney has not personally verified and requires disclosure of AI use to the court (Code of Civil Procedure s 128.7(b)(2)(A); s 6068.1(a)(3)(C)). Our guide to AI in law sets out court rules on AI in filings and the sanctions cases.^5,6,15

Does putting client information into an AI tool breach confidentiality or privilege?

Putting client information into a public AI tool can breach confidentiality and waive privilege, which lets a client refuse to disclose confidential communications with its lawyers. A UK tribunal said closed tools that keep data out of the public domain can be used for tasks such as summarising without that risk, but the SRA warns that paid tools may also lack the contractual and technical safeguards needed.^11,20

In UK and R (Munir) v Secretary of State for the Home Department [2026] UKUT 81 (IAC), the Upper Tribunal said that putting confidential client documents into a public AI tool breaches confidentiality and waives privilege (para 21; headnote 4). The SRA cites that passage.^11,20

In United States v. Heppner (memorandum of 17 February 2026), a federal court in New York held, on what it said appeared to be a question of first impression (p. 2), that a defendant’s written exchanges with a publicly available AI platform were protected by neither attorney-client privilege nor the work product doctrine, which covers material prepared for litigation. The platform is not a lawyer and its privacy policy let the provider use and disclose the exchanges (pp. 5 to 7). The court left open use directed by counsel (p. 7).¹⁹

Opinion 512 requires a client’s informed consent before information about the representation goes into a self-learning tool, one that learns from the data it takes in, and general wording in an engagement letter is not enough (p. 7). The Law Society’s view is narrower: use inside a firm’s own secure systems, where data does not train the model, “is not likely, of itself, to require informed consent”. Use that could breach confidentiality, such as training AI on client data, is likely to need specific consent. From 1 January 2027, California lawyers must not put confidential or other nonpublic information into a generative AI system unless access is restricted to the lawyer and people the lawyer authorises who are bound to keep it confidential (s 6068.1(a)(3)(A)).^1,2,5,6

How does the EU AI Act treat legal tech providers and deployers?

On our reading, the EU AI Act reaches most legal AI tools through AI-literacy duties and, for chatbots and generated content, transparency duties. A research or drafting tool that a law firm uses for clients is not high-risk, the Act’s most regulated category, under point 8(a) of Annex III to Regulation (EU) 2024/1689. That point covers AI used by or for a judicial authority to research and interpret facts and the law and to apply the law to facts, “or to be used in a similar way in alternative dispute resolution”.^29,30

The Act places duties on providers, who develop a system or have it developed and place it on the market or put it into service under their own name, and on deployers, who use a system under their authority (art. 3(3), (4)). A legal tech vendor is usually the provider and a firm using its tool the deployer. Both are covered from outside the EU if the output is used in the EU (art. 2(1)(c)).³⁰

From 2 August 2026, providers must design chatbots so that people know they are dealing with AI (art. 50(1); art. 113). Providers of generative systems must mark outputs as artificially generated in a machine-readable way (art. 50(2)). Regulation (EU) 2026/1744, which amended the Act, left that date unchanged; generative systems placed on the market before 2 August 2026 have until 2 December 2026 to meet the marking duty in art. 50(2) (art. 111(4), added by art. 1 point (39)(b)).^30,31

The AI-literacy duty in article 4 has applied since 2 February 2025. Regulation (EU) 2026/1744 replaced it (art. 1 point (5)): providers and deployers must “take measures to support the development of AI literacy” of their staff, and the duty “does not require” them “to guarantee any specific level” of literacy.^30,31

From 2 December 2027, the EU AI Act can treat AI that helps a court or an ADR body decide a case with legal effects as high-risk (Annex III point 8(a); Regulation (EU) 2026/1744, art. 1 point (40)(b)). Recital 61 limits the ADR part to outcomes with legal effects and excludes purely ancillary administrative activities, such as anonymising decisions. The European Commission’s draft guidelines of 19 May 2026, not yet adopted, put search tools generally outside point 8(a). Tools that interpret sources or infer legal reasoning are likely inside it (paras 421, 422).^30,31,32

How should a firm or institution evaluate an AI tool before buying it?

Before buying an AI tool, check what happens to your data, then test the tool on your own work, then settle who supervises it and how it is billed. The questions below draw on the CCBE’s technical guide of 27 March 2026.^1,2,7,11,14

• Will our inputs train the model, and does the contract say so? Enterprise and API services, which connect software directly to a model, often exclude training by contract, while consumer services may reserve the right to use inputs (CCBE technical guide, p. 8).
• Is there a data processing agreement? In England and Wales an enterprise vendor is likely to be a processor, which handles data for the firm. UK GDPR article 28 lets a firm use only processors that give sufficient guarantees, set out in a written contract (Law Society; CCBE technical guide, p. 7).
• How accurate is it on our work? Test it before relying on it, and review the quality of its outputs regularly (Law Society). Treat claims of hallucination-free output with care (the 2025 study, p. 1). Opinion 512 says a lawyer who tested a tool on a hand-checked sample and found its summaries accurate need not necessarily review every document (p. 4).
• What is it built on, and how do we leave? Rival products may share one underlying model, so one fault can affect several tools. Check the exit terms and how data would be retrieved if the service ends (Law Society).
• Who supervises it? Managers must set clear policies (Opinion 512, p. 10; Florida Opinion 24-1, p. 4) and train staff (Opinion 512, p. 10).
• What can an agent do on its own? The State Bar of California’s guidance says no autonomous filing or communication with the court, and no sending client information out without safeguards and human review (Practical Guidance, pp. 6, 9).
• How will it be billed? Hourly bills reflect time actually spent, not hours that AI saved (Opinion 512, p. 12; Law Society). A tool that works like office overhead is not charged to the client without advance disclosure, but a per-use service bought for one client’s matter can ordinarily be billed at actual cost (p. 13).
• Do we become the provider? Under the EU AI Act, anyone who puts its name on a high-risk system, substantially modifies it or repurposes it so that it becomes high-risk is treated as its provider (art. 25(1)).^{1,2,4,7,9,14,30}

Have regulators approved AI-based legal services?

Regulators have approved some AI-based legal services, but we found only one, the SRA, that has authorised a firm working purely through AI. In May 2025 the SRA authorised the first purely AI-based law firm it has approved, for debt recovery through the small claims process; the SRA said the system “is not autonomous” and that named regulated solicitors stay accountable (6 May 2025).²¹

The SRA sought reassurance on quality checks, confidentiality and conflicts of interest, and the system cannot propose relevant case law. The approval is of the firm: the Law Society’s practice note says the SRA does not “authorise, certify or approve AI systems or other software”.^1,21

Two US reforms widen who may offer legal services, neither written for AI. Utah’s regulatory sandbox, a supervised trial of new legal services (Standing Order No. 15, from 14 August 2020), covers entities using “technology platforms, tools, or applications” to practise law (p. 10). It is closed to new applications and is due to end on 14 August 2027. Utah’s innovation office lists “technology-based services such as AI” among the providers it has authorised.^22,23,24

Since 1 January 2021, Arizona has licensed alternative business structures, in which non-lawyers may own or help run a firm (Order R-20-0034, Rules 31.1(c) and 33.1). We found no approval route written for AI in either state.^22,24,25,26

What is still undecided about AI and legal tech?

• Whether California adopts the State Bar’s proposed AI amendments to six Rules of Professional Conduct. A second round of comment on a revised draft closed on 6 August 2026.
• Whether exchanges with an AI tool made at a lawyer’s direction can be privileged. Heppner left the point open (p. 7).
• How long Opinion 512’s consent rule for self-learning tools holds as tools change. The ABA tied its view to tools as of its publication (note 34, p. 7).
• Where the Commission’s final guidelines put research and drafting tools sold to courts and ADR bodies.
• What follows Utah’s sandbox after 14 August 2027, and whether other regulators authorise firms that work purely through AI.^{2,8,19,21,23,32}