A litigation partner reviewing a brief drafted with the help of a generative AI tool faces a question no software can answer for her: whether she is prepared to sign it. Under Fed. R. Civ. P. 11(b), her signature certifies that the legal contentions are warranted by existing law or by a nonfrivolous argument for changing it, formed after an inquiry reasonable under the circumstances [1]. The tool may have produced the draft in minutes, but the certification is hers, and so is the judgment about whether the argument serves the client. That allocation of responsibility is the most useful place to begin thinking about how firms should adopt AI. The firms likely to benefit are those that redesign their work around it. They will give machines the tasks whose output a lawyer can check, keep lawyers on the decisions for which they are accountable, and judge the result by the quality of the representation rather than the speed of its production.
The professional rules already assign judgment to the lawyer
The obligations that govern practice do not change when the drafting assistant is software. ABA Model Rule 1.1 requires competent representation, and Comment 8 to that rule extends competence to keeping abreast of the benefits and risks associated with relevant technology [2]. Ethics authorities have read ABA Model Rule 5.3, which governs a lawyer's responsibilities for nonlawyer assistance, to reach AI tools as well as human staff, so a supervising lawyer must take reasonable measures to ensure that the tool's work is compatible with the lawyer's own professional obligations [2]. The ABA's Formal Opinion 512 addresses generative AI under the duties of competence, confidentiality, communication, supervision, and reasonable fees [2]. State bars have issued their own guidance, which is not uniform, so a firm practicing in several jurisdictions needs to check each one.
Discovery practice points the same way. Rule 26(g) requires an attorney to sign every disclosure and every discovery response, certifying after reasonable inquiry that the disclosure is complete and correct and that the response is consistent with the rules [3]. A firm that uses a model to draft objections to interrogatories or to code documents for production does not shift that certification to a vendor. Confidentiality adds a separate constraint, because ABA Model Rule 1.6(c) requires reasonable efforts to prevent unauthorized disclosure of client information, which bears directly on what a lawyer may enter into a tool whose data-handling terms no one at the firm has reviewed [2]. Read together, these rules describe a division of labor before any firm designs one: the machine may produce, but the lawyer must verify, decide, and answer for the result.
Machines add the most value where their errors can be detected
If the lawyer must verify, the most productive uses of AI are those in which checking the output costs far less than doing the work from scratch. First-pass document review is the clearest example. Courts have accepted technology-assisted review for more than a decade, generally on the condition that the producing party can show its process was reasonable, often through statistical validation [3]. The output of that process can be tested: elusion sampling estimates how many responsive documents were missed, and any individual coding decision can be confirmed by a reviewer reading the document.
A hypothetical matter shows how this works in practice. In Meridian Analytics, Inc. v. Halvorsen Data Partners, LLC, an illustrative case, Meridian alleges that two former engineers took its source code to Halvorsen, and Halvorsen produces roughly 400,000 documents from 12 custodians. A model can rank those documents by likely relevance, flag every reference to the repository at issue, and draft a chronology of each engineer's communications in the weeks before departure. An associate can test each of those outputs against the documents themselves. Deposition digests, privilege-log first drafts, and routine responses to requests for admission share that quality, since a lawyer reads each against the record before relying on it.
The calculation changes when the output is hard to check. A model's account of the law on a contested question may be fluent and wrong, and the error stays invisible until someone reads the authorities it relies on. Such tasks are not off-limits, but their verification cost is high. A firm deciding whether a tool saves anything should count that cost honestly rather than measuring only how quickly the draft appeared.
Some work is the representation itself and cannot be delegated
A second category of work resists delegation because it consists of judgment rather than production. Suppose that in Meridian the Halvorsen production omits the chat messages of the two engineers. Whether Meridian's counsel should send a deficiency letter, propose narrowed search terms at a meet-and-confer, or file a motion to compel under Rule 37(a) depends on matters no document contains: the posture of the case, the court's likely view of proportionality under Rule 26(b)(1), the cost to the client, and what the gap may suggest about Halvorsen's preservation practices. A model can list those options, but it cannot weigh them against what Meridian actually wants from the litigation.
Privilege review illustrates the same limit at a finer grain. Attorney-client privilege and work-product protection are distinct doctrines with different elements and different waiver rules, and the hard calls, such as an email that copies in-house counsel on what is mostly business advice, turn on a primary-purpose analysis that varies by jurisdiction [4]. A tool can surface the email; deciding whether to withhold it, and defending that decision if challenged, belongs to a lawyer who understands both the doctrine and the risk of a waiver argument. The distinction explains why a privilege-log first draft belongs among the checkable tasks described above while the withholding decision behind each entry does not. The draft entry can be compared with the document, but the decision is a characterization for which the lawyer must answer.
Negotiation, advocacy, and counseling sit firmly in this category as well. A settlement conference turns on reading the other side's priorities and on credibility that counsel has built over time, and an oral argument turns on answering the question the court actually asks. ABA Model Rule 1.4 requires a lawyer to explain a matter to the extent reasonably necessary for the client to make informed decisions [2]. When Meridian's general counsel must decide whether to accept a structured settlement, she will want that explanation from someone who understands her business and will stand behind the advice.
Faster production can lower the quality of representation
The danger in AI adoption lies less in the technology than in the incentive to skip the step that makes it safe. Courts have sanctioned lawyers who filed briefs citing nonexistent cases generated by AI tools, and some courts have adopted standing orders or local rules requiring disclosure or certification of AI use in filings [5][2]. In each of those episodes the failure was one of verification. Efficiency pressure tempts a team to skip that step, and the tool's fluency makes skipping it feel safe.
A subtler risk is the gradual flattening of legal work. Model-generated drafts tend toward the conventional argument and the familiar structure, which is adequate for routine filings and a poor substitute for the theory opposing counsel has not anticipated. A firm that measures success by hours saved will reward the conventional draft. One that measures the accuracy of its productions, its results on contested motions, and whether clients return will push its lawyers to use the time they recover on the analysis that wins cases.
Firms must rebuild training and billing around judgment
The work most suited to automation is also the work through which junior lawyers have traditionally learned. First-pass review taught associates what a responsive document looks like, and drafting discovery responses taught them where objections succeed and fail. If machines take over those tasks, firms must find other ways to build judgment. They might assign associates to verify model output against the record and explain each disagreement, bring them into meet-and-confers and depositions earlier, and have partners review their reasoning as well as their conclusions.
Billing will need the same attention. Formal Opinion 512 indicates that a lawyer billing hourly may charge only for time actually spent, not for time a tool saved [2], and ABA Model Rule 1.5 requires that every fee be reasonable. For work whose cost structure AI has changed, flat or phased fees may align the firm's interest with the client's more closely than the hour does. Whatever the model, a firm's written policy on AI should address at least the following matters.
Which tools are approved, and on what terms of data retention and use.
Which categories of output require verification, by whom, and against what source.
When client consent is needed before confidential information is entered into a tool.
How the firm will comply with court rules and standing orders on disclosure of AI use.
How time and fees are recorded for work performed with AI assistance.
A policy of this kind does little unless it is enforced through supervision. Partners who sign filings under Rule 11 and discovery responses under Rule 26(g) are the natural owners of the verification step, and the policy should make that ownership explicit rather than leaving it to habit. Assigning ownership this way also tracks the supervisory duty under Model Rule 5.3, since the lawyer who directs a tool's use on a matter is the one best placed to take reasonable measures to confirm its output.
The firms that gain the most will be those that treat AI as a reason to raise the standard of their work rather than to lower its cost alone, because clients can buy cheaper document handling from many sources but sound judgment from few. The harder question is generational. If the routine work that once trained young lawyers is now done by machines, the profession has to show that it can still produce partners capable of the judgment that the rules, and clients, require. Firms that design deliberate answers to that question now, through supervised verification, earlier exposure to strategy, and candid review of reasoning, will be the ones with lawyers worth trusting a decade from now.
References
Fed. R. Civ. P. 11(b)(2). https://www.law.cornell.edu/rules/frcp/rule_11
Model Rules of Prof'l Conduct r. 1.1 cmt. 8 (Am. Bar Ass'n), quoted in ABA Comm. on Ethics & Prof'l Responsibility, Formal Op. 512, at 14 (2024). https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
Fed. R. Civ. P. 26(g)(1); see Da Silva Moore v. Publicis Groupe, No. 11 Civ. 1279 (S.D.N.Y. Feb. 24, 2012). https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2011cv01279/375665/96
In re Grand Jury, No. 21-55085 (9th Cir. Sept. 13, 2021), amended and superseded, 23 F.4th 1088 (9th Cir. 2022) (amended citation to be verified). https://law.justia.com/cases/federal/appellate-courts/ca9/21-55085/21-55085-2021-09-13.html
Mata v. Avianca, Inc., No. 22-cv-1461, 678 F. Supp. 3d 443 (S.D.N.Y. June 22, 2023) (reporter citation to be verified). https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54/
