Generative artificial intelligence (AI) is no longer a novelty in legal services. Instead, it is becoming infrastructure that pervades every aspect of the industry. By early 2026, 74% of lawyers, tax advisors, and other professionals reported using AI several times a week.[1] In corporate law departments, generative AI use more than doubled in twelve months to 52%.[2] Still, client-side demand is outpacing most firms’ delivery of AI services. Nearly 80% of corporate clients say AI-enabled quality is “very important or essential” when selecting advisors, while only 6% believe they receive it.[3] 32% of disappointed clients have reconsidered or plan to reconsider law firm relationships based on the use of AI, and among those clients, one-third estimate that more than $1 million in annual work is at risk.[4] With more AI use expected, legal providers that have paired AI with governance, transparency, and measurable outcomes are outpacing those that have merely purchased AI without a broader integration plan. Foley continues to position itself at the frontier of this technological revolution.
To evaluate the success of a firm’s use of AI in its work product, one must ask whether 1) it was designed as a repeatable process, 2) the AI involved was authorized to act and by whom, 3) the method would survive a skeptical adversary’s demands for proof and validation, and 4) a person will be accountable for the result. Design, authorization, proof, and accountability underpin Foley’s approach, while optimism and an expansive mindset enable Foley to transform AI proficiency into client value.
Key Takeaways
- Design: Workflows are superior to lists of tools. No matter the task at hand, it is generally better to ask about specific legal tasks rather than becoming mired in specific technologies. Ask instead about specific legal tasks end to end, including the inputs, the grounding documents and materials, and the areas where human intervention is indispensable.
- Authorization: You should insist on an owner for agentic work. AI may execute the steps, but someone should be able to say which attorney authorized the assignment, reviewed the output, and is accountable for it. Before letting an agent act on systems you do not fully control, you should define clear authority.
- Proof: The use of AI in areas such as eDiscovery must be defensible and not just efficient. AI-assisted review is inherently probabilistic, so you must pair it with validation, including sampling protocols, escalation procedures, and audit trails tracking the attorney’s decisions to confirm the AI’s calls were correct. For transactional, regulatory, and other practice areas, law firms must build the same validation protocols.
- Accountability: You should verify outputs and then trust in your team. Ask what AI output the team verified against original authority, how the team assembled the source materials, and what the team decided not to run through AI. Knowing when to keep AI out of the loop is an essential form of discernment.
Design: Turning AI Systems into Repeatable Workflows
Most law firms have access to substantially the same AI tools. What separates firms is whether an attorney can walk a client through how they achieved a specific legal task and reproduce that result on the next matter. A defensible workflow also documents the source materials provided to the AI, the outputs it generated, the attorney revisions made, and the authority supporting the final work product. A firm’s broad adoption can provide a critical springboard for deeper work on specific legal processes and workflows.
Foley’s attorneys have integrated AI into pre-existing workstreams. For instance, transactional teams use AI across a deal’s lifecycle so that diligence findings arrive early enough to shape negotiations, and so that they can benchmark drafted language against filing databases rather than a single lawyer’s recollection of market terms. Similarly, litigators use Relativity aiR, Relativity’s built-in AI software for discovery, to compress discovery review from weeks or months to days, and Harvey and CoCounsel to identify and organize key authorities so that attorney time goes to the strategy rather than only research. The objective is not automation for its own sake; it is faster, more consistent work that leaves attorneys more time for strategic advice and gives clients decisions made earlier on a more complete record.
Through its Practice Innovation program, Foley leverages AI across various workstreams (e.g., diligence, legal research, and contract drafting and review) by partnering its legal technology teams with practice groups to deliver vetted AI solutions and optimize client service. Foley’s AI Blueprint methodology starts with a legal task, maps where AI adds value, and specifies the inputs, source documents, and quality control checkpoints before a workflow earns “validated” status.
Additionally, Foley’s original chatbot (i.e., FoleyChat) received the Best Use of Artificial Intelligence Award at Legalweek 2026.[5] This proprietary technology further broadens the use of AI at Foley across workstreams and practice groups. FoleyChat is housed within the firm and allows attorneys to build proprietary agents for recurring legal tasks and supply the key information required. For example, Foley’s intellectual property (IP) attorneys use FoleyChat to stress-test proposed claim amendments and arguments against patent application materials, prior art, and examiner data, helping attorneys identify strengths, weaknesses, and likely counterarguments to gain a strategic advantage and achieve better results. In this manner, a team can work from materials it chose and organized rather than from the assumptions of a general-purpose model.
Authorization: Moving Beyond Simple Prompts to Agentic AI
The debate in the legal industry over agentic AI has largely focused on what agents can do, but the more important questions are where they are authorized to act and who is responsible if they act improperly. Clients are part of that analysis because many are already deploying their own agents into systems that interact with customers, vendors, and third parties.
Legal technology is shifting from standalone prompt-based tools to agentic systems that can organize and carry out multi-step tasks with less human intervention. Harvey is Foley’s primary AI assistant, but Foley maintains an extensive suite of AI software, available whenever it is a better fit for the task at hand. The practical difference between ordinary prompting and navigating agentic AI is not the length of the instruction; it is instead whether the lawyer must manually connect each step (e.g., reviewing the complaint, identifying pleading defects, researching authority, and drafting) or whether the system can sequence those steps on its own.
AI is also changing the legal services clients will consume. By reducing the time and cost of preliminary, recurring, and large-scale analyses, agentic workflows can make legal guidance a more regular part of business planning and strategy, including in areas where the cost or time required previously made that guidance impractical. For example, in the assessment of potential patent infringement, more frequent freedom-to-operate reviews can identify potential risks before a product is developed, launched, or commercialized, allowing the client to evaluate strategy with counsel to make informed decisions about design, licensing, investment, and market entry. The attorney must still define the assignment, supervise the workflow, and evaluate how the results affect the client’s business objectives.
The increased autonomy of agentic systems requires more structure. Complex assignments work best when divided into defined stages, which each have a clear input, limited scope, and a human checkpoint before the next step begins. Asking an agent to complete a complex task in a single instruction generally produces weaker results than a staged process, and the attorney must still refine and correct the output before it is used.
The traditional frameworks for assigning liability between humans and AI will also likely need revision. A recent Amazon lawsuit has brought these issues into the limelight Amazon.com Servs. LLC v. Perplexity AI, Inc. (N.D. Cal. Mar. 10, 2026). In this case, Amazon alleged that Perplexity’s Comet browser accessed and placed orders through customers’ Amazon accounts while disguising itself as a human shopper rather than identifying itself as an automated AI agent. Amazon claimed that Perplexity violated Amazon’s terms of service and federal and state computer-access laws. The district court granted a preliminary injunction blocking Perplexity’s AI shopping agents from accessing Amazon’s site and held that Amazon’s authorization, not the user’s permission, controlled whether Comet could access its website. On August 4, 2026, the Ninth Circuit vacated the preliminary injunction and remanded, holding that Amazon was unlikely to succeed because Perplexity did not itself “access” Amazon’s computers within the meaning of the federal and state computer-access laws Amazon.com Servs. LLC v. Perplexity AI, Inc. (9th Cir. Aug. 4, 2026). The Court held instead that the user accessed Amazon’s computers and merely utilized Perplexity’s AI shopping agent as a tool. Rather than establishing a settled rule, the case presents an open and actively contested question, which is whether organizations deploying agents must map an agent’s authorization to the systems it will touch as well as identify who will answer for any gap of authority between human and machine.
Agentic AI should expand what a legal team can execute, not obscure who is in control. The more that a system can sequence work, act across tools, or touch third-party environments, the more the team must deliberately monitor the assignment. Each agentic workflow should identify the human owner, the systems the agent may access, the checkpoints before an AI output is used, and the person responsible if an AI’s actions exceed the authority granted to it.
Proof: AI in eDiscovery as an Illustration of the Vanguard the Legal Profession Is Reaching
Whereas design and authorization relate to how the AI output is built, it is in areas like discovery that an adversary can actively challenge that output. This concept is not new to eDiscovery. Predictive coding and technology-assisted review were approved by courts more than a decade ago. Further, first-level attorney review was displaced long before generative AI arrived by technology-assisted review (TAR) and contract-attorney arrangements. As such, within the legal industry, attorneys have already become accustomed to defending a machine-assisted process to a skeptical judge, providing sampling statistics and audit trails.
However, a new development is the use of agentic sequencing with review, issue organization, and privilege triaging occurring as connected stages under human supervision. AI can now identify relevant material, surface key documents, organize issues, and synthesize large volumes of electronically stored information in hours or days rather than weeks or months. Notwithstanding, courts do not accept efficiency as an answer to a deficient production. Auditable work product is therefore critical.
Foley was among the earliest adopters of aiR, Relativity’s built-in AI software for discovery. The firm uses aiR at the outset of discovery to accelerate first-level review, surface key documents, and help clients assess facts in the context of overall case strategy. Faster review is reliable only with defined quality-control mechanisms. These mechanisms include sampling protocols to validate AI coding, escalation procedures for ambiguous documents, and audit trails documenting decisions to include or exclude certain records. Accordingly, aiR drives value by streamlining a repeatable and verifiable workflow that still requires attorney-led decision making.
By example, in one multilingual internal investigation, Foley used aiR to analyze Spanish-language documents, generate outputs in English, and provide issue identification and supporting citations.[6]Through deliberate prompt refinement and quality control procedures, less than 1% of aiR’s recommendations were ultimately overturned by Foley’s attorney reviewers. This example illustrates the benefit of using AI to efficiently gain insights from large datasets while the process remains measurable and reviewable. As AI use expands, other legal processes will benefit from the same verifiable workflows that have been implemented throughout eDiscovery.
Accountability: Confidentiality, Accuracy, and Knowing When AI Should Not Be in Control
Accountability is an essential element of the successful use of AI. A team may have a well-designed workflow, a properly authorized agent, and a tested review process, but the work can still fail if no one verifies the output or knows when AI should not be a part of the process. In addition, the AI tools can provide a deceptive facade of accuracy, which can lead to a false sense of confidence in the results. A polished answer can conceal weak or inapposite authorities, incomplete grounding, or untested assumptions. Accountability requires attention to confidentiality and privilege implications, accuracy grounded in sources that avoid hallucination risks, and the judgment to know when AI should not control the work.
Recent court decisions illustrate the developing legal framework surrounding confidentiality, attorney-client privilege, and the work-product doctrine in connection with the use of AI. In United States v. Heppner (S.D.N.Y. Feb. 17, 2026), the court concluded that communications with a publicly available AI platform (Claude) were not privileged or confidential because they involved a non-lawyer and a platform that retained user conversations. By contrast, Warner v. Gilbarco, Inc. (E.D. Mich. Feb. 10, 2026) recognized that AI-assisted litigation materials qualify for work-product protection when disclosure of work product is to an AI program rather than an adversary, noting specifically that generative AI programs are “tools, not persons.” Morgan v. V2X, Inc. (D. Colo. Mar. 30, 2026) reinforced that protection where outputs were not likely to reach an adversary, while nonetheless requiring disclosure of the AI platform used. Taken together, these cases focus less on whether AI was used and more on how, by whom, and with what safeguards. Security and governance, as opposed to mere access to a platform, will increasingly determine whether privilege and confidentiality have potentially been waived.
Accountability also requires monitoring for potential hallucination errors, which generally fall into three categories:
- Fabrication occurs when AI invents cases or citations that do not exist and is usually the easiest error to detect.
- Misattribution occurs when AI cites a real authority but references language or quotations the source never used, making the error harder to catch because the citation appears legitimate.
- Mischaracterization occurs when AI quotes authority accurately but wrongly claims it supports a legal proposition that it does not.
Mischaracterization is the most dangerous of the three errors because it requires legal judgment to identify, as opposed to just cite-checking. The practical obligation is therefore broader than simply confirming that a citation exists. Conclusions must be checked against their sources. The legal team should read the underlying cases, verify statutes and regulations, and confirm docket developments against the court record. Each type of citation error can undermine legal filings and has resulted in sanctions against attorneys and firms who relied on AI-generated authorities without adequate verification.[7]
Foley’s AI implementation enables responsible AI use because it treats AI as an augmentative technology and not as a substitute for human judgment. Human oversight, structured AI governance, and rigorous quality-control procedures enable Foley to capitalize on the efficiencies AI offers while minimizing risks associated with privilege, confidentiality, and hallucinations. The firm’s use of secure, siloed AI systems substantially mitigates any risk that client information will be used to train public models or be disclosed to third parties. Data isolation, access controls, and audit trails make ambitious AI workflows defensible. The challenge is to balance that protection with the growing expectation that knowledge be accessible and actionable. Foley protects these AI systems that drive value by remaining highly transparent with its methodologies and implementing validation standards that allow clients to see and shape how AI knowledge is applied.
Key Questions to Ask Your Law Firm and Your Company
Clients should ask their law firms certain questions about the law firm’s use of AI. A company’s general counsel should be able to answer these same questions about the company’s own use of AI. Using AI responsibly is not something a law firm can simply claim in a pitch; the law firm must be able to demonstrate its AI use. When evaluating a law firm’s AI use, do not ask for a list of tools. Instead, ask to see one actual piece of work from start to finish and ask the following questions:
- Design. What information went into AI? What sources was it checked against? Did a human review it? What had to occur before the task was considered finished?
- Authorization. Which parts of the work did a person complete, and which did AI complete? Where did responsibility pass from AI to the lawyer? Did the AI ever act on its own outside systems that the law firm controls? If so, who is responsible when that happens?
- Proof. How did the team check AI’s work? How often? What triggered a second look? Did the team set those standards before the work began rather than constructing them after to explain a mistake? What record exists to show what happened if a question arises later?
- Accountability. How was client information handled? What was verified against the governing law or original source material? What categories of work has the firm decided AI should not touch at all?
The firms that lead will be the ones that can explain the entire process clearly — how the work is designed, who is doing each part, how the results are verified, and who answers if something goes wrong. Early adoption alone is not enough. Clients should expect nothing less than a full picture of how AI is improving process and delivering stronger results every day.
This article was prepared with the assistance of 2026 summer associate Nick McNulty.
[1]Thomson Reuters, Future of Professionals Report 2026, at 4, https://www.thomsonreuters.com/en/institute/future-of-professionals-2026/report (last accessed Sept. 30, 2026).
[2] ACC & Everlaw, Generative AI’s Growing Strategic Value for Corporate Law Departments (Oct. 14, 2025) (active GenAI use among U.S. in-house respondents more than doubled from 23% to 52%), https://www.acc.com/resource-library/generative-ais-growing-strategic-value-corporate-law-departments-survey-results (last accessed Sept. 30, 2026).
[3] Thomson Reuters, Future of Professionals Report 2026, at 4, https://www.thomsonreuters.com/en/institute/future-of-professionals-2026/report (last accessed Sept. 30, 2026).
[4] Id.
[5] Foley & Lardner, Foley Wins Best Use of AI at Legalweek Leaders in Tech Law Awards (Mar. 10, 2026), https://www.foley.com/news/2026/03/foley-wins-best-use-of-ai-at-legalweek-leaders-in-tech-law-awards/ (last accessed Sept. 30, 2026).
[6] Foley & Lardner, Breaking Language Barriers with Generative AI: How Foley & Lardner Conducted Multilingual Document Review with Relativity aiR for Review (Mar. 17, 2025),Olivia S. Singelmann & Byron J. McLain, https://www.foley.com/insights/publications/2025/03/breaking-language-barriers-generative-ai/ (last accessed Sept. 30, 2026).
[7] See, e.g., Malkeet Lnu v. Blanche, 177 F.4th 1014, 1031 (9th Cir. Jun. 3, 2026).
