Most law firms use artificial intelligence the way they once adopted e-discovery platforms: as a new tool placed on top of a structure built for something else. An associate runs a deposition transcript through a summarization tool, a paralegal sends a contract set through a clause extractor, and the output enters the same review chain, the same staffing model, and the same billing entries that existed before. This post considers a different arrangement. It examines what would change in how a firm is organized, supervised, and paid if a litigation practice were designed from the start on the assumption that software can produce credible first drafts of most written work product.
The answer has less to do with technology than the phrase "AI-native" suggests. The tools will change quickly and unpredictably. A firm designed around them has to settle, deliberately, who verifies machine output and to what standard, where client data may go, how lawyers are trained when the work that used to train them is automated, and how clients pay for work that takes fewer hours. Each of those questions is already governed by professional obligations that do not change because the drafting method has.
An AI-native firm organizes work around review rather than production
Traditional litigation workflows are built around production. A junior lawyer drafts, a senior lawyer revises, and a partner signs, with hours concentrated at the bottom of the pyramid, where the drafting happens. When a model can produce a serviceable first draft of a meet-and-confer letter, a deposition outline, or a chronology in minutes, the scarce resource shifts from drafting capacity to the judgment needed to distinguish a sound draft from a plausible but wrong one. A firm designed around that shift would treat review as the primary unit of work, staff it explicitly, and measure it.
Consider an illustrative matter, Meridian Analytics, Inc. v. Halvorsen Data Partners, LLC, a hypothetical trade-secret and breach-of-contract case in federal court. Halvorsen has produced roughly 300,000 documents, and Meridian's team needs a chronology of the departure of two engineers and the transfer of source code. In a conventional workflow, contract reviewers code documents for weeks and an associate builds the chronology from their tags. In an AI-native workflow, a model proposes a chronology with document citations on the first day, and the real work becomes checking each entry against the underlying document, finding what the model missed, and deciding which gaps matter. The effort has moved rather than disappeared, and it has moved to a task that demands more experience, not less.
That relocation carries a design consequence firms tend to underestimate. An informal review step will erode under deadline pressure, because a polished draft invites skimming in a way that a rough one does not. An AI-native firm would therefore define verification for each type of output, such as citation checks for briefs, source-document tie-outs for chronologies, and sampling protocols for document classification, and would keep a record of who performed each step. That record serves the supervising lawyer, who can confirm that the check was actually done, and it serves the firm if a filing is later questioned, because it shows what inquiry was made and by whom.
Existing professional duties already set the verification standard
None of this requires new ethics rules, because the duties governing AI-assisted work are the ones that already govern delegated work. ABA Model Rule 1.1, read with Comment 8, requires competence that includes understanding the benefits and risks of relevant technology, and most states have adopted some version of that comment [1][2]. ABA Model Rules 5.1 and 5.3 make lawyers responsible for supervising other lawyers and nonlawyer assistance, and ABA Formal Opinion 512, issued in 2024, applies those supervisory duties to generative AI tools [2]. Because the opinion also addresses confidentiality, client communication, and fees, it is a reasonable starting framework for internal policy.
In litigation, the court-facing duties are sharper. By signing a filing, an attorney certifies under Fed. R. Civ. P. 11(b) that its legal contentions are warranted by existing law or a nonfrivolous argument for changing it, and that its factual contentions have or will likely have evidentiary support. Rule 26(g) imposes a parallel certification, made after reasonable inquiry, on discovery requests, responses, and objections [3]. Courts have sanctioned lawyers for filing briefs that cited cases a generative tool invented [4][5], and some courts have issued standing orders requiring disclosure or certification of AI use [citation needed]. Those orders vary by court and by individual judge, so a firm practicing in many districts needs a reliable way to track them.
An AI-native firm would build its systems on the premise that "the model drafted it" is never a defense. Every citation in a brief is pulled and read by a person before filing. Each factual assertion in the Meridian chronology traces to a Bates number that someone has opened. Privilege calls feeding a log under Rule 26(b)(5)(A), which requires the withholding party to describe withheld documents in a manner that enables others to assess the claim, are reviewed by a lawyer prepared to defend them [3]. The standard is the one a careful partner applies to a first-year associate's draft, applied without exception.
Confidentiality obligations shape the firm's technical architecture
Verification governs what leaves the firm, while confidentiality governs what enters the tools in the first place. Model Rule 1.6(c) requires reasonable efforts to prevent unauthorized disclosure of or access to client information, and that duty follows every document uploaded to a vendor's system [2]. Litigation adds a further layer, since material produced under a protective order entered under Rule 26(c) may be restricted to outside counsel or to specified uses, and many standard protective orders predate generative tools entirely. If Meridian's counsel uploaded Halvorsen's "Attorneys' Eyes Only" documents to a tool whose vendor retains inputs or trains on them, the firm could face a problem under the order independent of any ethics question.
For that reason, an AI-native firm would treat data governance as an architectural decision rather than a per-matter improvisation. It would know, for each tool, where data is processed and stored, whether inputs are retained or used for training, who at the vendor can access them, and how data is deleted when an engagement ends. It would also negotiate protective orders and ESI protocols with AI processing in mind, and would seek a Fed. R. Evid. 502(d) order early so that an inadvertent production of privileged material does not turn into a waiver dispute [6]. Client outside-counsel guidelines increasingly address AI use as well [7], and some require notice or consent before client data is processed by generative tools.
Automating junior work removes the firm's training ground
The hardest organizational problem is not technical. Firms have long relied on a leverage model in which associates billed many hours on document review, research memos, and first drafts, and in doing so learned how cases are built. The work easiest to automate is the same work that taught junior lawyers to recognize a significant document, to notice when a case does not support the proposition for which it is cited, and to see why a discovery request is overbroad. A firm that automates that work without replacing the instruction it provided will soon have senior lawyers expected to verify output they were never trained to produce.
Training would therefore have to become deliberate where it was once incidental. One approach treats verification itself as structured training, with a senior lawyer reviewing both the result and the method the junior lawyer used to check it. In the Meridian matter, an associate who traces each entry in a model-generated chronology to its source, and writes up what the model missed, may learn the record faster than one who spent a month coding documents, provided someone reads the write-up and responds. Another approach reserves some assignments for unassisted drafting. Either requires partner time that never appears on a bill, which makes training a question of firm economics as much as of pedagogy.
Hourly billing and efficient tools pull in opposite directions
The economic tension follows directly from the shift in labor. Under the billable hour, a tool that turns a 40-hour task into a six-hour one reduces revenue on that task proportionally unless the firm finds more work or raises rates. Model Rule 1.5 requires reasonable fees, and Formal Opinion 512 states that a lawyer billing hourly may charge only for time actually spent, not for the time a task would have taken without the tool [2]. A firm cannot capture efficiency gains by billing as though they had not occurred.
An AI-native firm would accordingly move toward pricing tied to deliverables rather than hours. Fixed fees for discrete phases, such as initial disclosures under Rule 26(a)(1), responses to a set of interrogatories, or a first-level privilege log, become more workable once a firm can predict its own costs. These arrangements shift risk to the firm, however, and they depend on task-level cost data that most firms do not yet keep. They also raise a communication issue, because clients will want to know how their work was done, and Model Rule 1.4 may require explaining AI use where it bears on the representation [2].
Established firms can adopt the design one practice group at a time
Few established firms will rebuild around AI in a single step, and they do not need to. The design choices described above can be made incrementally, beginning where the risk is greatest. A litigation group preparing to use generative tools on live matters should have written answers to the following questions before the first upload.
Which categories of output require which verification steps, and who signs off on each?
Which tools are approved for confidential or protected material, and on what vendor terms?
How will protective orders, ESI protocols, and Rule 502(d) orders address AI processing?
How will junior lawyers learn the work when first-pass tasks are automated?
How will efficiency be reflected in fees and explained to clients?
None of these answers will be permanent. Tools will change and courts will keep refining their expectations, so the policy should be revisited on a fixed schedule rather than after a problem surfaces. The advantage is likely to go to firms that make verification, data governance, training, and pricing explicit, because those decisions outlast any particular tool. What remains unsettled is how the efficiency gains will ultimately be divided among firms, clients, and the lawyers whose training depended on the work now being automated.
References
Robert Ambrogi, Tech Competence, LawSites, https://www.lawnext.com/tech-competence (last visited Oct. 12, 2026). Note: this is a secondary source, a well-known running tracker by a lawyer-journalist. It also reprints the text of amended Comment 8. For a law-firm audience, consider adding the ABA's own text of Comment 8 or the rules of particular states. https://www.lawnext.com/tech-competence
ABA Comm. on Ethics & Pro. Resp., Formal Op. 512, at 2 (2024) (citing Model Rules of Pro. Conduct r. 1.1 & cmt. 8). Page 14 also quotes the full text of Comment 8. https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
Fed. R. Civ. P. 26(g)(1). The rule also covers Rule 26(a)(1) and (a)(3) disclosures, not only requests, responses and objections. https://www.law.cornell.edu/rules/frcp/rule_26
Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023). The court imposed sanctions under Rule 11 and, alternatively, its inherent power, including a $5,000 penalty. The reporter citation was confirmed from the E.D. Tex. standing order below, which cites it. Note: the filing in Mata was an "affirmation in opposition," not a brief in the strict sense. https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54/
Standing Order MC-11, Disclosure and Certification Requirements for Use of Generative Artificial Intelligence (E.D. Tex. Beaumont Div. July 20, 2026). The order requires disclosure, a certification that the party independently verified the output, and a certification that no confidential information was disclosed. Formal Op. 512, at 10, separately notes that "Some courts have responded by… https://www.txed.uscourts.gov/sites/default/files/judgeFiles/Standing%20Order%20MC-11%20(7.15.26).pdf
Fed. R. Evid. 502(d). The Advisory Committee's explanatory note says such an order may provide for return of documents without waiver "irrespective of the care taken by the disclosing party." Fed. R. Civ. P. 26(f)(3)(D) (same URL as the Rule 26 blocks) also calls for discovery plans to address whether to seek a Rule 502 order. https://www.law.cornell.edu/rules/fre/rule_502
FDIC, Outside Counsel Deskbook § 1.17 (rev. Nov. 2025). This is one client's binding guideline. It requires written notice before using GenAI and bars putting sensitive information into "open environment" tools. It supports the second clause only. Formal Op. 512, at 8, also contemplates outside counsel guidelines that require disclosure of GAI use. https://www.fdic.gov/doing-business-fdic/outside-counsel-deskbook.pdf
