Your firm undoubtedly has a GenAI policy. It probably tells lawyers to avoid tools that leak client data, admonishes them to avoid hallucinations, and encourages them to incorporate AI into their practice to satisfy client demands. It likely does not tell them which cognitive functions they should delegate to GenAI, which they should not, and where the line between the two is absolute. In the space between the restriction and the mandate, lawyers are making consequential decisions about GenAI delegation every day. Most firms have not addressed that space with meaningful governance.

GenAI Can Make Legal Work Worse

GenAI’s capacity to transform legal work for the better is real, but so is its capacity to degrade it. Most law firm leaders know that AI can hallucinate. Far fewer know that it can make expert legal judgment and work product actively worse.

The best evidence of this dynamic comes from a Harvard Business School experiment with Boston Consulting Group consultants.^[1] For professional tasks requiring breadth and option generation, GenAI delivered: output quality improved by 40%, and consultants worked faster. But for tasks requiring judgment and synthesis, something unexpected happened. Consultants using GenAI were 19% less likely to produce correct solutions than those working without it.

The same pattern appears in research evaluating GenAI use in legal analysis. A study of law students confirmed that AI dramatically improves performance on straightforward analysis while producing no measurable benefit for complex reasoning, while introducing recurring failures: jumping to conclusions, missing less obvious issues, generating confident prose that masks superficial analysis.^[2]

And a recent study focused on legal tasks showed that GenAI assistance on a synthesis task improved performance by nearly 60% and produced a surprising downstream benefit: participants who used AI for synthesis outperformed the control group on the subsequent independent reasoning task even after GenAI was removed.^[3] But when GenAI was introduced at the revision stage, the picture changed. It helped weaker performers, but it actively degraded the work of stronger ones. The best lawyers in the study produced worse revised work product when they used GenAI than when they worked without it.

A Use-Mode Governance Framework

Governing GenAI’s uneven performance requires asking a question most firms are not asking. Instead of asking whether GenAI is appropriate for a particular deliverable (a brief, a contract, or a board presentation), the governance question should be: what cognitive function is being delegated to GenAI at each step in the workflow? My proposed framework organizes common GenAI uses into seven recurring modes following the sequence in which lawyers actually use GenAI to produce legal work product. Governance controls are calibrated to the risk profile of each mode.

Table 1 (below) sets out the framework.

Table 1: Use-Mode Governance Framework

Modes 1 and 2: Retrieval and Organization. At the mechanical end of the cognitive spectrum are two distinct functions. In retrieval mode (Mode 1), a lawyer reviewing a merger agreement asks GenAI to identify every representation and warranty in the document. In organization mode (Mode 2), a litigator reviewing fifty depositions asks GenAI to construct a timeline from the testimony. The first locates material that already exists. The second arranges it into a usable structure. No new content is created in either case. The risk is verification failure. Both uses are low-risk and should be actively encouraged, subject to modest verification controls. Firms that unduly restrict these use modes are leaving value on the table.

Mode 3: Summarization. Summarization (Mode 3) introduces selection risk. In this mode, GenAI chooses what to emphasize, include, and omit. Consider a lawyer preparing a board presentation on the results of an internal investigation. GenAI can condense dozens of witness interviews into key points and themes in minutes. But a summary may focus on procedural detail while missing credibility issues that a lawyer would immediately recognize as material. The appropriate control is to mandate meaningful review by a lawyer with first-hand knowledge of the source material. A lawyer encountering the summary cold has no reliable way to evaluate what GenAI missed.

Mode 4: Candidate Generation. Mode 4 is exploratory. A lawyer drafting a brief might ask GenAI to generate a list of potential arguments, propose alternative framings, or identify supporting authority. This “candidate material” expands options and accelerates iteration. The work product is not filing-ready and must be treated as provisional. GenAI can suggest, but a lawyer must decide.

The authority verification obligation at this stage deserves special emphasis. GenAI will identify cases, summarize holdings, and weave them into an argument structure. The output will read fluently and cite real-looking cases. A lawyer cannot assume the model has accurately characterized the holdings or context.^[4] Any authority cited in an external filing must be independently read and verified. GenAI can help find the cases. A lawyer must read and apply them.

Mode 5: Editing and Rewriting. In Mode 5 (editing and rewriting for clarity), a lawyer asks GenAI to tighten a dense contract provision or restructure a wordy paragraph. The risk is unintended meaning change. An edit may read cleanly while subtly narrowing a representation, softening a covenant, or eliminating a carve-out. The revision risk is not hypothetical. The study referenced above found that stronger performers produced worse work product when GenAI revised their independently produced memos than when they worked without it.^[5] In this mode, a lawyer must confirm that the edit produced no shift in meaning and introduced no new factual assertions.

Mode 6: Critique and Stress-Testing. Mode 6 may be the most underutilized GenAI capability. Before filing a brief or presenting to regulators, a lawyer can ask GenAI to identify weaknesses in the argument, surface adverse authority, and generate the strongest counterarguments. GenAI functions as a “red team,” finding vulnerabilities before adversaries do. Unlike every other mode, the risk here runs in one direction. Lawyers who skip this step are missing one of GenAI’s core value propositions. Firm governance frameworks should require it in appropriate cases, not merely permit it.

Mode 7: Evaluation and Decision. The boundary becomes absolute when GenAI is asked to evaluate or decide (Mode 7). A lawyer advising a board on whether an event requires disclosure cannot delegate that determination to GenAI. A litigator assessing settlement value cannot outsource probability judgments. These are core expressions of professional responsibility. In this mode, GenAI may inform background analysis, but it may not substitute for a lawyer’s judgment in making the call. This is a categorical prohibition. Professional judgment cannot be delegated.

Conclusion

Law firm leaders who have moved from restriction to mandate without governing the space between have not finished the job. Their lawyers are making consequential decisions about GenAI use every day without the guidance they need and deserve.

The use-mode framework gives firm leadership a practical tool for filling that gap: identifying where GenAI enhances legal work, where it introduces serious risk, and where professional judgment is non-negotiable. Firms that govern at that level will capture GenAI’s value. Firms that do not will have policies that look serious but govern nothing important.

Reprinted with permission from Thomson Reuters Institute, “The GenAI governance gap: Why current law firm policies fall short,” thomsonreuters.com/en-us/posts/technology/genai-governance-gap/, with permission of Thomson Reuters. Copyright © 2026. For further information on Thomson Reuters Institute, please visit thomsonreuters.com/

Endnotes

*The views expressed in this article are solely those of the author in his individual capacity and do not represent the views, positions, or opinions of Foley & Lardner LLP, its partners or clients, or the University of Wisconsin Law School. ^

[1] Fabrizio Dell’Acqua et al., Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality, Harvard Bus. Sch. Working Paper No. 24-013 (2023) (forthcoming in Organization Science). ^

[2] Jonathan H. Choi & Daniel Schwarcz, AI Assistance in Legal Analysis: An Empirical Study, 73 J. Legal Educ. 384, 420 (2025). ^

[3] Nicholas Bednar, David R. Cleveland, Allan Erbsen & Daniel Schwarcz, Artificial Intelligence and Human Legal Reasoning (Apr. 5, 2026) (unpublished manuscript) https://ssrn.com/abstract=6525800 at 22-29. ^

[4] The risk is two-fold: fabricated authority (citations to cases that do not exist) and the subtler and more dangerous “misgrounding” (real cases cited for propositions they do not support, in the wrong jurisdiction, or under overruled law). A recent empirical study of two leading commercial legal research tools found hallucination rates above 17% even after retrieval-augmented generation was applied. Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. Empirical Legal Stud. 216 (2025). ^

[5] See Bednar et al., supra note 3. ^