You ask a junior associate to draft a services agreement. A day later, a polished-looking document lands in your inbox. You later learn much of it was written not by the associate, but by a generative artificial intelligence (AI) tool.

This is no longer hypothetical. Generative AI is now used to draft and redline contracts, write and polish memoranda, and review volumes of documents in data rooms, among other things. Already in use at many firms, these tools offer real efficiency – but they can also make error-filled work look deceptively complete.

For supervising lawyers, the question is not whether junior lawyers are permitted to use generative AI – in many firms, they already do. The pressing question is how to supervise a junior lawyer who is using it.

The Workflow Has Changed, Not Our Ethical Obligations

Our ethical obligations under the Rules of Professional Conduct have not changed in the age of AI. The existing duties of competence, confidentiality, communication, and reasonable fees already frame the analysis.

However, a couple of rules warrant a close look when evaluating a lawyer’s supervisory duties. For supervising lawyers, Supreme Court Rule (SCR) 20:5.1 and SCR 20:5.3 are the natural starting points. Under SCR 20:5.1(b), a supervising lawyer must “make reasonable efforts to ensure that … [a subordinate] lawyer conforms to the Rules of Professional Conduct.” SCR 20:5.3(b) imposes a parallel duty when supervising nonlawyer assistants.

The State Bar of Wisconsin has yet to issue a formal ethics opinion on SCR 20:5.1 or SCR 20:5.3 in the context of generative AI – or on generative AI more broadly. But we can learn from other state bars and a growing number of misconduct cases. The Professional Ethics Committee for the State Bar of Texas put it bluntly:

Though this should likely go without saying, a lawyer should always verify the accuracy of any responses received from a generative AI tool. … So, the Committee will say it again: lawyers are responsible for the work product they submit regardless of who (or what) does the original research and drafting.[1]

It should indeed go without saying – yet it bears repeating, given how many lawyers have been caught filing briefs with hallucinated citations.
In re Marullo involved a supervising attorney who signed and filed briefs, prepared by an associate and interns, that contained multiple citation errors. He did not check the briefs for accuracy and was unaware that the associate had used an AI tool for the underlying research. The Massachusetts Board of Bar Overseers publicly reprimanded him under several rules, including Rules 5.1 and 5.3, faulting him both for not checking the citations and for having no policies and procedures in place to ensure that anyone did.[2]

However, a written AI policy is not enough on its own, nor is assuming a lawyer used an approved platform, or that the platform did not hallucinate. In
Hill v. Workday, another case involving hallucinations, the Northern District of California sanctioned a supervising attorney even though the junior lawyer claimed to use Westlaw’s CoCounsel, the firm had trained its staff on AI misuse, and counsel was aware of Thomson Reuters’ disclaimer about the tool’s accuracy.[3]

Put directly, a supervising lawyer may delegate work to a subordinate, and that subordinate may use an AI tool to assist with the work.

But the fact remains – the supervising lawyer remains answerable for the result.

Supervise Both the Process and the Output

A significant danger of AI output is that it looks authoritative and, to the uncritical eye, correct – likely one reason why so many lawyers have filed briefs with hallucinations.

The danger is arguably more hidden for transactional lawyers than for litigators. A judge or opposing counsel is likely to flag a hallucinated case. Errors enter transactional documents, many of which contain no citations, more quietly. A contractual provision may be legally binding but suboptimal. A corporate charter provision may run afoul of a state governing statute. No opposing counsel or court is standing by to catch these mistakes – at least not when the document is signed or adopted.

This brings us to the central question: What supervisory efforts are “reasonable” when an attorney receives a transactional document a junior lawyer or assistant produced with AI?

Errors are best caught early, well before the document ever reaches the supervisor. The following are practical guidelines:

Adopt a firm AI policy: The foundational step is to adopt a written policy on generative AI use and to review and update it regularly. The contents of such a policy are beyond this article’s scope, but the policy should be mapped against the Rules of Professional Conduct, including duties of competence, confidentiality, reasonable fees, and supervision under SCR 20:5.1 and 20:5.3.[4] It should also address firm training and verification procedures, discussed below.

Provide training on AI use: Any firm employee using a generative AI tool should understand what the tool does well, how the tool can fail (hallucinated cases, outdated law, invented facts, phantom cross-references, etc.), and verification requirements and procedures. Training should also cover best practices for using each generative AI tool.

Training should be practice-area specific. Junior litigators will likely know – even if they should be reminded – how to verify facts, cases, and statutes in court briefs and related documents. Junior transactional lawyers often have not been taught a systematic method for verifying a document’s deal terms, consistency with market practice, and compliance with governing statutes and regulations.

In the transactional clinic I direct at Marquette University Law School, students are trained on generative AI before they counsel their first client. The training covers permitted AI tools, confidentiality, and verification, and I provide prompting templates designed for the projects they are likely to undertake for their small-business clients.

Build verification into the workflow: Treat an AI-generated draft like a summer associate’s first attempt – useful, but presumptively unverified and likely to contain errors. Build verification into the workflow rather than bolting it on at the end:

  • Require the use of vetted, capable AI tools configured for legal work, confidentiality, and source-verification workflows.
  • Direct the junior lawyer to prompt the tool to flag its assumptions, note missing information, and cite sources for the facts and law it supplies. A draft that surfaces its own uncertainties is easier to check than errors hidden behind fluent prose.
  • Require a redline against the firm’s forms or prior precedent, which will facilitate verification.
  • Independently verify what an AI tool is most likely to get wrong – facts (names, dates, and deal terms), formulas, document cohesion, and alignment with governing statutes and regulations.
  • Require a verification record saved to the client file. (In my transactional clinic, students complete an “AI Use Log,” which I review each time an AI tool touches client work.)

Must You Check Every Line Yourself?

Does a supervising attorney personally need to re-verify the entire contents of a transactional document?

The answer is no, with qualifications. ABA Formal Opinion 512 is explicit that, while a lawyer should not rely on AI output without independent review, the lawyer need not necessarily review and verify every output.[5] The appropriate level of review depends on the tool, the task, and the lawyer’s experience with workflow.

For example, Opinion 512 describes that if a vetted AI tool in a tested workflow reviews and summarizes numerous, lengthy contracts, the lawyer could potentially manually review a subset of the contracts and compare them to the summaries created by the tool to confirm the process is reliable.

The Quieter Risk: Associates Who Never Learn to Draft

There is a second concern the ethics rules do not capture.

Junior lawyers build judgment by both doing the work and observing how senior lawyers change and improve it. When generative AI does the heavy lifting, that formative friction can disappear. The associate passes along facially competent work without personally acquiring the competence behind it.

Much can be said on this topic, but the best guidance I can offer is to pause and mentor rather than simply revise the document and send it to the client. Discuss with the junior lawyer the AI tool used, the verification that occurred, and how the work product could be improved.

To the extent that time constitutes substantive review and improvement of client work, it may be billable, subject – as always – to the reasonableness requirement of SCR 20:1.5.[6]

Conclusion

Generative AI can make transactional work faster and, in some respects, better. But speed is not soundness, and a polished draft is not a verified one.

Your ethical obligations, including under SCR 20:5.1 and 20:5.3, do not shrink because an AI tool wrote the first version. If anything, they call for a review and updating of firm policies and a change in traditional workflow.

And remember that the point of assigning the first draft to the junior associate was never just the draft. It was the lawyer they were becoming while writing it.

This article was originally published on the State Bar of Wisconsin’s Business Law Blog. Visit the State Bar sections or the
Business Law Section webpages to learn more about the benefits of section membership.

Endnotes

[1] Tex. Comm. on Prof’l Ethics, Op. 705 (2025). ↩

[2] See In re Marullo, Public Reprimand No. 2025-2 (Mass. Bd. of Bar Overseers Apr. 30, 2025).↩

[3] See Hill v. Workday, Inc., 2026 WL 1146289, at *1, 7 (N.D. Cal. Apr. 28, 2026); Hill v. Workday, Inc., No. 23-cv-06558-PHK, ECF No. 161, at 7 (N.D. Cal. Sept. 5, 2025), accessible on this website.​ ↩

[4] This list is meant to be illustrative, not comprehensive. Firms with a litigation practice will also want to address in their AI policy ethical obligations with respect to claims and contentions, and candor toward the tribunal.
See ABA Formal Opinion 512, 9-10 (2024). ↩

[5] See ABA Formal Opinion 512, at 4. ↩

[6] ABA Formal Opinion 512 provides guidance on what fees and expenses can be charged with respect to generative AI tools.
See id., at 11-13. ↩