The Mendes Hershman Student Writing Contest is a highly regarded legal writing competition that encourages and rewards law students for their outstanding writing on business law topics. Papers are judged on research and analysis, choice of topic, writing style, originality, and contribution to the literature available on the topic. The distinguished former Business Law Section Chair Mendes Hershman (1974–1975) lends his name to this legacy. Read an edited version of this year’s third-place winning essay, submitted by Benjamin A. Glinton of Fordham University School of Law, Class of 2027.
Introduction: From Substitution Claims to Practice-Level Consequences
Few developments have generated more speculation about the legal profession’s future than the rapid advancement of generative artificial intelligence (“AI”). Since large language model (“LLM”) tools became publicly available in late 2022, commentators and legal industry observers have debated with increasing urgency whether AI will replace lawyers. Headlines have declared that AI can pass the bar exam, draft contracts indistinguishable from those produced by experienced attorneys, and complete in minutes due diligence reviews that previously required weeks of associate labor.[1] This discourse is not without foundation, but framing the debate around replacement is analytically unhelpful. It blurs the more consequential question: not whether AI will replace lawyers, but how AI is already reshaping the structure of legal work and what follows from that transformation.[2]
This paper addresses that question directly, focusing on transactional legal practice (mergers and acquisitions (“M&A”), private equity, capital markets, and corporate advisory) work that constitutes the core revenue engine of large law firms.[3] Transactional practice is a particularly useful lens because it combines two categories of work that map cleanly onto AI’s strengths and limitations: document-intensive execution tasks on one hand, and judgment-intensive advisory functions on the other.[4] AI is genuinely transforming the first category. It is structurally limited in the second. Understanding the boundary between them, and what it means for how legal work is organized, priced, and practiced, is the central analytical task this paper undertakes.
This paper’s thesis proceeds in two related parts. First, AI materially improves efficiency in document-intensive transactional workflows by enabling faster and more consistent performance of pattern-recognition tasks such as contract review, due diligence, precedent mining, and first-pass drafting.[5] Second, the probabilistic architecture of generative AI systems creates durable structural constraints that currently prevent reliable performance on tasks that define legal judgment: reasoning about novel issues, making normative trade-offs between competing legal and policy values, and navigating the social and political context embedded in complex dealmaking.[6] Because these limitations correspond directly to the core functions of corporate lawyering, AI does not eliminate the need for lawyers but instead reconfigures where legal judgment is exercised and where professional risk resides.[7]
The “AI replacing lawyers” narrative fails because legal practice is not a single, undifferentiated activity. It is a composite of tasks with different cognitive profiles; some are highly susceptible to automation, and some are structurally resistant to it.[8] Asking whether AI will replace lawyers is, therefore, like asking whether spreadsheets replaced accountants: the question misses the point.[9] The more useful questions are which parts of legal work AI is changing, how those changes are propagating through the organization and economics of law firms, and what new demands they place on the lawyers who remain.
This paper follows in seven parts. Section II maps current AI deployment in transactional practice. Section III analyzes the structural limits of AI in legal judgment. Section IV examines real-world consequences inside BigLaw and corporate legal departments. Section V applies the paper’s framework to risk allocation in M&A transactions. Section VI analyzes the professional responsibility framework governing AI use. Section VII examines the long-term structural consequences of AI adoption for BigLaw as an industry. Section VIII details the trajectory of AI toward more effective normative legal reasoning. Together, these sections develop a coherent account of AI as a force that reshapes legal practice from within, concentrating professional judgment more precisely where it genuinely matters.
How AI Is Actually Used in Transactional Practice
Understanding AI’s long-term consequences for the legal profession requires starting from a precise empirical baseline: what AI is actually doing inside law firms and corporate legal departments today. The popular discourse shifts between utopian predictions of full automation and dismissive claims that legal work is too nuanced for machines; neither framing is useful. AI is being deployed in transactional practice in a defined set of high-volume, pattern-intensive workflows where its comparative advantages are clear, while human lawyers retain responsibility for the judgment-intensive work that gives transactions their legal and commercial significance.
Due Diligence and Document Review
The most widespread application of AI in transactional practice is due diligence, the process by which a buyer’s legal team reviews the target company’s contracts, corporate records, and regulatory filings to assess legal risk before closing. In large transactions, due diligence can involve tens of thousands of documents across multiple data rooms, historically consuming enormous amounts of junior associate time while creating significant opportunities for human error due to fatigue and inconsistency.[10]
AI-powered contract review platforms can now classify documents by type; extract specific contractual provisions; identify clauses with particular legal significance (change-of-control triggers, assignment restrictions, and termination rights); and flag deviations from market-standard terms—all at speeds far exceeding human reviewers.[11] Platforms such as Kira Systems, Luminance, and Relativity have become standard infrastructure at many large firms precisely because they operationalize AI’s core comparative advantage: pattern recognition across large datasets.[12] Where a human reviewer must read each contract sequentially, an AI system can simultaneously process hundreds of agreements, extract the same provision across all of them with consistent accuracy, and present results in a structured format for attorney review.
Independent studies have found that machine-learning-assisted review achieves accuracy rates comparable to, and in some conditions exceeding, those of experienced human reviewers conducting exhaustive manual review.[13] This performance profile makes AI-assisted due diligence not merely faster but qualitatively superior in consistency, reducing the variance that results when large document sets are distributed across multiple junior reviewers working under time pressure.
The practical implications for deal execution are significant. Due diligence timelines that previously stretched across several weeks can now be compressed substantially, enabling faster transaction closings and reducing the labor costs associated with large-scale review. Critically, however, AI-assisted due diligence does not eliminate the need for attorney judgment; it redirects it. The AI surfaces issues; the lawyer must assess their materiality, understand their transactional context, and determine the appropriate response. The analytical work of evaluating what a flagged clause means for the deal remains squarely within the domain of human professional responsibility.
Contract Analytics and Negotiation Support
Beyond initial due diligence, AI is increasingly deployed in ongoing contract analytics: systematic analysis of a firm’s or client’s existing contract portfolio to identify risk concentrations, track negotiation outcomes, and benchmark terms against market standards.[14] In M&A practice, this function supports both the diligence process and post-closing integration work, in which acquirers must assess how target-company contracts interact with their existing agreements.
Contract analytics platforms enable legal teams to answer questions that would previously have required substantial manual effort: How many of the target’s customer contracts contain most-favored-nation pricing provisions? What percentage includes automatic renewal clauses that could complicate post-closing restructuring? Are there patterns in how certain counterparties have historically negotiated indemnification caps? AI makes these analyses tractable at scale, surfacing information that informs negotiation strategy and risk assessment in ways that were practically unavailable before.
The negotiation support function of AI is more limited but growing. Some platforms now offer clause-level benchmarking against negotiated market precedents, allowing attorneys to assess whether a proposed contractual provision is favorable, unfavorable, or within the range of typical outcomes for comparable transactions.[15] This capability enhances the informational foundation for negotiation, though the strategic decisions it informs remain with counsel.
Generative Drafting and Document Production
The application of generative AI to legal drafting represents a qualitatively different, and more contested, deployment than document review or analytics. LLMs can produce first-draft contract provisions, memoranda, term sheets, and other transactional documents from natural language prompts, drawing on patterns learned from the legal texts in their training data.[16] Law firms including Allen & Overy, Latham & Watkins, and Davis Polk have disclosed pilots or deployments of generative AI drafting tools, and legal technology vendors have introduced purpose-built drafting platforms tailored to transactional practice.[17]
The efficiency gains from AI-assisted drafting can be substantial for routine document production. A first draft of a standard nondisclosure agreement, board consent, or officer’s certificate that might previously have required an hour of junior associate time can be generated in seconds, freeing attorney time for review and customization rather than initial composition. For more complex instruments such as credit agreements, merger agreements, and purchase price adjustment mechanisms, AI can produce structurally coherent drafts that serve as useful starting points, though they require careful attorney review and substantial modification.
The critical limitation of generative drafting is that speed in producing text does not equate to accuracy in legal meaning. LLMs generate text by predicting statistically likely sequences of words based on training data; they do not reason about whether the provisions they generate are legally effective, commercially appropriate, or internally consistent in the context of a specific transaction.[18] Attorneys who use AI-generated drafts without rigorous review risk incorporating provisions that are superficially plausible but legally deficient—a risk that courts and bar authorities are beginning to address directly. The well-publicized cases in which attorneys submitted AI-generated briefs containing fabricated citations illustrate the professional responsibility stakes of overreliance on generative tools.[19]
The Structural Limits of AI in Legal Judgment
Mapping AI’s capabilities in transactional practice reveals a pattern: AI performs consistently well on tasks that can be decomposed into pattern recognition over well-defined inputs, and it performs unreliably on tasks that require genuine legal judgment. This is not a limitation that incremental scaling of existing architectures will straightforwardly overcome; it is a durable feature of how current probabilistic systems work, though one whose boundary may shift as architectures evolve. Understanding why this boundary exists is essential to understanding its implications for the practice of law.
The Architecture of Probabilistic Systems
LLMs generate outputs by predicting statistically likely continuations of input text based on patterns learned from training data.[20] This architecture is highly effective for tasks where the correct output is, in fact, the statistically likely one—identifying the type of clause that typically follows the phrase representations and warranties of the seller, for example, or recognizing that a given provision deviates from market-standard language. It is structurally problematic for tasks that require reasoning beyond pattern frequency.
This characterization, while technically accurate, requires qualification.[21] Current frontier models exhibit emergent capabilities that substantially exceed simple next-token prediction.[22] Chain-of-thought reasoning enables models to decompose complex legal problems into intermediate analytical steps before reaching conclusions, producing outputs that reflect structured analysis rather than mere pattern completion.[23] In-context learning allows models to absorb particularized information within a single interaction (a client’s risk profile, a set of deal terms, a jurisdictional preference) and adapt their outputs accordingly.[24] Analogical transfer enables the application of reasoning patterns from one legal domain to structurally similar problems in another, a capacity that closely mirrors one of the core cognitive operations of legal practice.[25] These capabilities are genuine and commercially significant; they account for the efficiency gains that AI-assisted legal work is already producing across transactional practice, and any framework that does not account for them will fail to describe the technology that lawyers are actually using.[26]
The structural limitation, however, persists. Understanding why requires distinguishing between the sophistication of the output and the nature of the process that produces it.[27] When a frontier model works through a material adverse change (“MAC”) clause analysis using chain-of-thought reasoning, it generates text that resembles the intermediate steps a human reasoner would produce because that is what structured legal reasoning looks like in the model’s training data.[28] The model is constructing a reasoning trace, rather than maintaining an internal representation of the commercial reality it is analyzing.[29] This distinction matters because these emergent capabilities are most reliable precisely where the training distribution is richest: standard contractual provisions, well-established regulatory frameworks, and frequently negotiated deal terms.[30] They degrade in the conditions where legal judgment is most valuable: when the problem is genuinely novel, when the answer cannot be derived from the distribution of prior outcomes, and when the stakes require an accountable professional who can defend the reasoning behind a decision.[31] Moreover, the gap between human few-shot generalization (the capacity to reason reliably from a small number of examples about genuinely unprecedented problems) and LLM in-context learning is not obviously closable through incremental architectural improvement alone.[32]
Legal judgment regularly involves reasoning about situations for which there is no adequate historical precedent; making normative choices between competing legal values when text and precedent underdetermine the outcome; and applying professional experience to assess the credibility, motives, and likely behavior of counterparties in ways that cannot be reduced to textual pattern matching.[33] These are not marginal or exotic features of legal work—they are at its core. The transactional lawyer’s value to clients lies precisely in the ability to navigate situations where the answer is not already encoded in existing patterns because if it were, sophisticated clients could find it themselves.
Novelty and the Limits of Training Data
One structural limitation of AI in legal practice is its dependence on training data that necessarily reflects past legal practice. An LLM trained on existing contracts, cases, and legal memoranda learns the patterns of how legal issues have been analyzed and resolved historically. It cannot reason reliably about genuinely novel legal questions—new regulatory regimes, unprecedented transaction structures, or emerging markets where established legal frameworks do not clearly apply—because it has no training signal for such situations.[34]
This limitation is acutely consequential in transactional practice, where deal structures regularly test the boundaries of existing legal frameworks. The emergence of cryptocurrency-denominated acquisition consideration, the development of new regulatory approval requirements in merger control law, and the ongoing evolution of Delaware corporate law doctrine all present issues where the answer is not derivable from prior patterns. The lawyer who can reason from first principles about how established doctrine should apply to new circumstances provides value that AI systems structurally cannot replicate.[35]
Normative Reasoning and Value Trade-Offs
Legal judgment also requires normative reasoning—the capacity to make principled choices between competing legal and policy values when text and precedent leave the outcome genuinely open.[36] In transactional practice, these trade-offs are pervasive: How should an indemnification cap be set when litigation risk is uncertain? How aggressive should a contract interpretation be when it would benefit the client but expose the counterparty to unexpected liability? What level of disclosure is appropriate when regulatory requirements are ambiguous?
These questions do not have statistically predictable answers. They require the application of professional judgment informed by an understanding of legal doctrine, business context, client risk tolerance, and professional ethics. AI systems can surface relevant precedents and identify the range of market outcomes for comparable provisions, but they cannot make the normative judgment about which outcome is appropriate for a specific client in a specific transaction.[37] That judgment is irreducibly human. The reason is that legal judgment is a normative act, requiring an accountable professional capable of being held responsible for its consequences.
This position is not without challengers, and the objection deserves serious engagement. A sophisticated critic might argue that normative judgment is itself reducible to patterns—that what lawyers call “value trade-offs” are, at a sufficiently granular level of analysis, structured preferences that can be learned from observed human decisions. Under this view, if a model trains on enough judicial opinions, regulatory guidance, negotiated deal terms, and legislative history, it might not merely describe what norms have been applied but replicate their application with sufficient fidelity to constitute functional normative reasoning. The argument gains traction from the demonstrated ability of LLMs to produce outputs that, on their face, closely resemble reasoned legal analysis—outputs that courts, practitioners, and scholars sometimes struggle to distinguish from human work product.[38]
This objection should not be dismissed. But it conflates mimicry with judgment. The distinction matters because normative legal reasoning is not merely pattern reproduction—it is the capacity to make a defensible choice between competing values in a context where the choice has consequences for real parties and where the reasoning behind it can be interrogated, revised, and held accountable. When a lawyer advises a client on whether a particular contractual allocation of risk is commercially reasonable, the answer is not derivable from prior examples alone. It requires a judgment about what the client’s interests actually are, what the counterparty’s tolerance for risk exposure is likely to be, what a court in a particular jurisdiction would likely find reasonable, and whether the deal’s commercial logic supports deviation from market standard. Each of these inputs is contextually specific, and the decision that integrates them involves a form of practical legal reasoning that is irreducibly first-person: the lawyer is deciding what should be done in this one, drawing on prior outcomes without being determined by them.[39]
There is also a deeper structural problem. AI systems are trained on outputs—on the artifacts of human judgment—but not on the deliberative process that produced them. They learn what decisions were made, not how the competing considerations were actually weighed, what was rejected, or why the result was regarded as justified. When a model reproduces the form of normative reasoning, it is reconstructing the surface of a decision-making process whose internal logic it has never observed. This is not a limitation that improved training data straightforwardly resolves; it reflects the fact that the normative dimension of legal judgment is embedded in a practice—a socially accountable form of reasoning—rather than in a dataset.[40] Accordingly, even a well-calibrated model that consistently produces plausible normative outputs should be understood as approximating the products of legal judgment, not exercising it.
Social, Political, and Relational Context
Complex transactions are not purely legal events—they are social and political processes in which relationships, reputations, personalities, and power dynamics shape outcomes in ways that legal text does not capture.[41] The decision to push hard for a particular contractual concession involves assessing the counterparty’s risk tolerance, the transaction’s strategic importance to both parties, the relationship between the principals, and the likely reaction of the other side’s counsel. These assessments require the kind of situated judgment that comes from professional experience and social intelligence.
AI systems have no genuine access to this social and relational context. They can process text; they cannot read a room, assess the credibility of a counterparty’s walk-away threat, or calibrate advice to the political dynamics of a corporate board. These capacities are central to what senior transactional lawyers actually do and are structurally unavailable to probabilistic systems operating on text inputs.[42]
Practice-Level Consequences Inside BigLaw and Corporate Legal Departments
The deployment of AI in transactional practice is not merely a technology story—it is producing concrete, measurable changes in how law firms are organized, how legal work is priced, how risk is allocated between firms and clients, and how lawyers are trained. These consequences are already visible inside the industry and will accelerate as AI capabilities continue to develop.
Workflow Transformation and Staffing Implications
The most immediate organizational consequence of AI deployment is the compression of workflows that previously required large teams of junior associates. When AI can perform a first-pass review of ten thousand contracts in hours rather than weeks, the staffing calculus for large transactions changes fundamentally. Firms that have deployed AI-assisted due diligence tools report significant reductions in the associate hours required for document review, with some indicating reductions of 50 percent or more on review-intensive matters.[43]
This compression has implications for staffing at every level. The junior associate pipeline, historically the entry point into BigLaw and the training ground for developing core analytical skills, is under structural pressure. If AI performs the document review work that previously occupied first- and second-year associates, the number of associates needed for that work decreases, and the developmental pathway through which associates built foundational skills is disrupted.[44] The long-term implication is that firms will need to redesign associate development programs to substitute structured training for the incidental learning that previously occurred through high-volume document review.
Billing Models and Client Expectations
AI-driven efficiency creates direct tension with the billable-hour model that dominates BigLaw economics. When AI reduces the time required to complete a task, hourly billing produces lower revenue per matter unless rates are increased proportionally—a dynamic that sophisticated clients are already exploiting. Large institutional clients are increasingly demanding that efficiency gains from AI be passed through in the form of lower fees, either through reduced hourly billing or through alternative fee arrangements that price matters based on value rather than time.[45]
This pressure is reshaping the competitive landscape. Clients who are well-informed about AI capabilities are asking firms to justify staffing decisions and to account for AI use in billing. Some clients have begun requiring firms to disclose when AI tools are used on their matters and to explain how efficiency gains are being shared. The long-term trajectory is toward value-based pricing models in which the price of legal services reflects the complexity and risk of the judgment involved rather than the volume of hours spent, with AI-assisted tasks priced at rates that reflect their reduced labor input.[46]
Malpractice Risk and Quality Control
AI deployment creates new vectors of malpractice risk with which the profession is only beginning to grapple. The risks are less likely to involve overly wrong outputs, which experienced attorneys will catch, than subtly wrong ones: AI provisions that are legally plausible but commercially inappropriate, analyses that are accurate with respect to historical precedents but miss recently decided cases, or drafting that is internally consistent but fails to achieve the client’s actual objective.[47]
The supervisory challenge is compounded by the speed and volume at which AI operates. When AI reviews thousands of contracts and produces a summary report, the reviewing attorney is, as a practical matter, trusting the AI’s extraction and classification decisions for the vast majority of documents, with detailed human review reserved for flagged items. The quality of that process depends entirely on the sophistication of the attorney’s understanding of the AI’s capabilities and limitations, and on the robustness of the quality control protocols the firm has established.[48]
The harder malpractice question—and the one the profession has not yet adequately confronted—arises not when AI produces overtly fabricated outputs, but when it generates provisions that are subtly defective in ways that a competent attorney exercising reasonable care might not detect. Consider a scenario in which an AI system drafts a MAC clause for an acquisition agreement, the supervising attorney reviews the clause with appropriate diligence and approves it, and the provision later fails in litigation because of an interaction between the AI-generated carve-out language and a jurisdiction-specific interpretive doctrine that the attorney did not identify and that the AI’s training data did not adequately represent. In this scenario, the attorney has arguably satisfied the supervisory obligations established by Model Rules 5.1 and 5.3 and the competence standard of Rule 1.1—the review was conducted, the provision appeared legally sound, and the error was not reasonably detectable through the exercise of ordinary professional care. Yet the client has suffered a material loss attributable to a defect in the AI-generated work product. Where malpractice liability falls in such cases—whether with the supervising attorney, the firm, the AI vendor, or some allocation among them—is a question that existing professional responsibility frameworks are not well equipped to answer, and one that will become increasingly pressing as AI-generated provisions become more prevalent in executed agreements.
Disrupted Training Pipelines
The compression of junior associate work raises a structural challenge for the profession that extends beyond near-term staffing concerns. The traditional BigLaw training model relied on large volumes of supervised work—document review, due diligence, drafting under supervision—as the mechanism through which associates developed legal judgment over time.[49] The incidental learning embedded in that work—understanding how contracts are structured, how provisions interact, how deal teams function under pressure—is difficult to replicate in the abstract.
If AI performs an increasing share of this foundational work, firms face the challenge of developing senior lawyers who have sufficient foundational knowledge to supervise AI effectively, exercise independent judgment, and advise clients in complex situations, even though they have not acquired that knowledge through the traditional developmental pathway.[50] The profession has not yet developed consensus models for addressing this challenge, though some firms are experimenting with structured training programs, simulation-based learning, and deliberate mentorship frameworks designed to substitute for the experiential learning that automation is displacing.
Case Study: Risk Allocation in M&A Transactions
The theoretical framework developed in Sections III and IV finds concrete illustration in the design of risk-allocation mechanisms in M&A—a domain where AI’s capabilities and limitations are both acutely visible and commercially consequential.
The Structure of M&A Risk Allocation
The fundamental legal work of a corporate acquisition involves allocating risk between buyer and seller across multiple dimensions: the risk that the target’s business performs worse than represented, the risk that undisclosed liabilities emerge after closing, the risk that regulatory approvals are not obtained, and the risk of litigation arising from the transaction itself.[51] This allocation is achieved primarily through three contractual mechanisms: representations and warranties, indemnification provisions, and MAC clauses—each of which requires layers of legal judgment that cannot be automated away.
AI’s Contribution to Risk-Allocation Work
AI can contribute meaningfully to the information-gathering and preliminary drafting phases of risk-allocation design. In diligence, AI can systematically extract representations and warranties from the target’s existing contracts, identify inconsistencies between those representations and the acquisition agreement’s representations, and flag provisions in third-party contracts that could be triggered by the transaction.[52] In drafting, AI can generate first-pass provisions that incorporate standard market terms and identify the range of negotiated outcomes in comparable transactions.
The Irreducible Judgment in MAC Clauses
The limitations of AI become most visible in the design of MAC clauses—provisions that determine whether a buyer can terminate an acquisition agreement if the target suffers a material adverse change between signing and closing. The legal definition of what constitutes a “material adverse change” is among the most contested and fact-specific questions in Delaware corporate law, with judicial outcomes turning on nuanced assessments of business context, industry conditions, and the parties’ reasonable expectations at signing.[53]
Designing an effective MAC clause for a specific transaction requires the attorney to make a series of normative judgments: How much downside risk should the buyer bear for industry-wide disruptions? How should the clause address known risks that are disclosed at signing but have uncertain magnitude? What carve-outs are appropriate given the target’s specific business and the competitive dynamics of the deal? These judgments cannot be derived from historical patterns because they depend on the parties’ risk tolerances, the strategic importance of the transaction, and the attorney’s assessment of how a Delaware court would evaluate the clause in future litigation.[54]
An AI system can identify how MAC clauses have been structured in comparable transactions and what judicial decisions have said about specific formulations. However, it cannot make the normative judgment about which formulation is right for this client in this deal—and the consequences of getting that judgment wrong can be measured in billions of dollars.
Purchase Price Adjustment Mechanisms and Earn-Outs
Similar dynamics appear in the design of purchase price adjustment mechanisms and earn-out provisions. AI can assist with the mechanical aspects of drafting: generating accounting methodology definitions, structuring the calculation waterfall, and identifying market precedents for comparable provisions.[55] But the design judgment—how to balance the buyer’s interest in protecting against deterioration with the seller’s interest in clean certainty, or how to structure an earn-out that incentivizes management without creating perverse incentives or litigation risk—requires the kind of experience-based professional judgment that is irreducibly human.
The practical significance of this framework for M&A practice is that AI enables the legal team to work from a richer informational foundation and to produce higher-quality first drafts more efficiently, while the judgment required to customize those drafts to the specific transaction remains the exclusive province of experienced counsel. The result is not that AI replaces the M&A attorney but that it raises the minimum informational baseline from which the attorney’s judgment operates.
Professional Responsibility Framework
The integration of AI into legal practice does not occur in a regulatory vacuum. Lawyers are subject to professional responsibility obligations that govern how technology can be used in the delivery of legal services, and the emergence of AI has begun to generate specific guidance—from bar associations, courts, and legal ethics authorities—about how those obligations apply.
Competence and the Duty to Understand AI Tools
Model Rule 1.1 requires that lawyers provide competent representation, including the “legal knowledge, skill, thoroughness and preparation reasonably necessary for the representation.”[56] Comment 8 to Rule 1.1 extends this obligation to technology, providing that competent representation requires lawyers to keep abreast of “changes in the law and its practice, including the benefits and risks associated with relevant technology.”[57]
Applied to AI, the competence obligation requires that attorneys who use AI tools in client representations understand those tools’ capabilities and limitations sufficiently to evaluate their outputs and identify their errors. This is a higher bar than simply knowing how to operate the tool—it requires understanding the architecture of probabilistic systems well enough to recognize the categories of error they are prone to produce, the types of tasks for which they are reliable, and the circumstances that require independent verification of AI-generated work.[58]
The ABA’s Formal Opinion 512 (2024) addresses AI use directly, confirming that the competence obligation applies to AI tools and that lawyers must exercise independent judgment in reviewing AI-generated work product.[59] The opinion specifically cautions against overreliance on AI outputs in contexts where independent verification is feasible, noting that the failure to review AI-generated work with appropriate care can constitute a breach of the competence obligation regardless of whether the AI-generated output ultimately proves accurate.
Supervisory Obligations and Delegation to AI
Model Rules 5.1 and 5.3 establish supervisory obligations for partners and supervising attorneys with respect to other lawyers and nonlawyer assistants, respectively.[60] These rules have been interpreted to extend to AI tools when those tools are used to perform functions that would otherwise be performed by junior lawyers or legal assistants. The supervisory obligation requires that the supervising attorney establish adequate protocols for reviewing AI-generated work, maintain meaningful oversight of AI-assisted workflows, and take responsibility for the accuracy and quality of AI-generated outputs included in client deliverables.
This framework has important practical implications. A partner who delegates due diligence to an AI platform and accepts its output without adequate review is not insulated from professional responsibility liability by the delegation itself—the obligation to supervise the work, and the responsibility for its quality, remains with the supervising attorney.[61] Firms are therefore required to develop institutional protocols for AI oversight, including quality control procedures, documentation requirements, and clear allocation of supervisory responsibility for AI-assisted workflows.
Confidentiality and Data Security Obligations
Rule 1.6’s confidentiality obligation creates significant compliance challenges for AI use, particularly when legal work involves the use of third-party AI platforms that process client documents on external servers.[62] Uploading a client’s confidential contracts to an AI platform without adequate consent or data security protections can constitute a breach of confidentiality, regardless of whether the information is actually disclosed to unauthorized parties.
Firms have responded to this challenge by negotiating enterprise agreements with AI vendors that include data isolation provisions, prohibit use of client data for model training, and establish contractual liability for data breaches. The adequacy of these protections is a matter of ongoing legal and technical debate, and the professional responsibility requirements in this area are still developing.[63] What is clear is that confidentiality compliance must be integrated into AI deployment decisions from the outset, not treated as an afterthought to technical implementation.
Candor to Tribunals and the Fabrication Problem
The widely publicized cases in which attorneys submitted AI-generated briefs containing fabricated case citations have brought Rule 3.3’s candor obligation into sharp focus.[64] LLMs are prone to producing plausible-sounding but factually false citations—a phenomenon known as “hallucination”—because their architecture optimizes for statistical plausibility rather than factual accuracy. An attorney who submits AI-generated work product to a court without verifying citations and legal authority violates the candor obligation and exposes the client, the firm, and the attorney to sanctions.
While the hallucination problem is most evident in litigation contexts, it also has transactional relevance. AI-generated deal documents that cite regulatory provisions inaccurately, mischaracterize case law on MAC clauses, or misdescribe statutory requirements create risks beyond bar discipline—risks include potential malpractice liability and deal-level consequences if defective provisions are included in executed agreements.
Long-Term Structural Consequences for BigLaw
The cumulative effect of AI integration in transactional practice will, over time, produce structural changes in the BigLaw market that extend well beyond individual workflow modifications. These changes will affect the competitive dynamics of the legal industry, the economic model of large law firms, and the professional stratification of the lawyer workforce.
Market Consolidation and Competitive Differentiation
AI adoption is likely to accelerate market consolidation among large law firms, for reasons that reflect the economics of technology investment. Developing and deploying enterprise AI systems requires substantial capital investment in platform licenses, data infrastructure, training, and quality control systems.[65] Firms with larger revenue bases can absorb these investments more easily and capture efficiency gains at scale, thereby creating a competitive advantage over smaller firms that cannot match their technology investments.
The competitive implications cut in multiple directions. On the one hand, AI enables large firms to serve clients at a lower cost per transaction, making them more competitive on price-sensitive matters. On the other hand, AI makes sophisticated legal capabilities more accessible to smaller firms and alternative legal service providers, potentially enabling competition with BigLaw in categories of work that previously required large-firm scale. The net effect is likely to be increased polarization: very large firms with substantial AI capability will compete for complex, judgment-intensive work at the high end of the market, while AI-enabled alternative providers capture routine work at the bottom, squeezing mid-market firms that cannot differentiate effectively on either dimension.[66]
Business Model Disruption and the Billable Hour
The structural tension between AI-driven efficiency and hourly billing is more than a negotiating dynamic between firms and clients; it is a fundamental threat to the revenue model on which BigLaw has operated for decades.[67] When AI dramatically reduces the time required to perform legal work, the billable hour either produces dramatically lower revenue or requires dramatic rate increases to maintain revenue levels. Neither outcome is sustainable in the long term in a competitive market where sophisticated clients can observe and challenge billing practices.
The likely resolution is an accelerated transition toward alternative fee arrangements—fixed fees, success fees, subscription models, and value-based pricing—that decouple legal service pricing from time input.[68] This transition rewards firms that perform work efficiently and accurately, rather than those that generate high volumes of billable hours. In the long run, it favors AI-enabled practices that deliver judgment efficiently over labor-intensive practices that derive revenue from volume. The firms best positioned for this transition are those that can articulate and price the value of the judgment they provide, rather than depending on the volume of hours they bill.
Professional Stratification and the Reconfiguration of the Associate Model
The most consequential long-term structural effect of AI adoption may be the reconfiguration of the associate workforce and the professional hierarchy within large law firms. The traditional BigLaw pyramid, in which large classes of junior associates performing high-volume work supported smaller classes of senior associates and partners, depended on economics that are no longer stable when AI performs much of the foundational document work.[69]
The emerging model is more stratified. At the top, senior lawyers with sophisticated judgment capabilities and client relationships will remain central—their work is the least susceptible to automation and the most valuable to clients. At the bottom, a smaller number of junior lawyers will handle AI-assisted work that still requires attorney oversight; but the volume of this work, relative to the number of people needed to do it, will be lower than in the past. In the middle, there is genuine uncertainty about what the mid-career lawyer’s role looks like in an AI-integrated firm—whether it contracts as routine work is automated away, or expands as senior lawyers delegate judgment-level work downward to associates who are freed from document review.[70]
This stratification has implications for access to the profession, diversity, and health of the legal workforce pipeline. If the volume of entry-level legal work decreases substantially, law schools will graduate more students than the market can absorb, and the profession will become more selective and less accessible to lawyers from less privileged backgrounds who depend on the volume-based model to develop skills and build careers. These distributional effects deserve serious attention from the profession’s regulatory and educational institutions.
The Redefinition of Legal Value
Ultimately, the structural consequence of AI integration is a redefinition of the purpose lawyers serve. The traditional conception of legal value—expertise operationalized through time-intensive document work—is giving way to a conception in which legal value resides primarily in judgment, relationships, and the ability to navigate complexity that resists systematization.[71] This is less a diminishment of the legal profession than a clarification of what has always been most valuable about it.
The lawyers best positioned for the AI-integrated future are those who combine deep understanding of AI’s capabilities and limitations with the capacity to exercise independent judgment in the spaces where those limitations bind. The profession’s challenge is to develop training, credentialing, and business models that cultivate and reward these capacities—rather than continuing to structure legal education and law firm economics around the volume-based model that AI is rendering obsolete.
The Trajectory of AI Toward More Effective Normative Legal Reasoning
The structural limitations identified in Section III are real, but they should not be mistaken for permanent boundaries.[72] The trajectory of AI development suggests several pathways through which probabilistic systems may become materially more effective at tasks that approximate normative legal reasoning, even if the philosophical question of whether machines can truly exercise judgment remains open.[73] Understanding these trajectories is important for the profession’s planning horizon because the practical consequences of AI that is functionally effective at normative-seeming tasks may be indistinguishable from those of AI that genuinely exercises judgment—and the legal industry must prepare for the former regardless of its position on the latter.[74]
Retrieval-Augmented Reasoning and Dynamic Contextual Integration
One of the most consequential limitations of current generative AI systems—their dependence on static training data that necessarily reflects past practice—is already being addressed through retrieval-augmented generation architectures that allow models to access and integrate external information sources in real time.[75] In the legal context, this development has immediate implications for the novelty problem identified in Section III.B: a system that can retrieve and reason over current judicial opinions, recently enacted regulations, and live deal databases is less constrained by the historical boundaries of its training corpus than one operating solely on pretrained patterns.[76]
The more significant development is the emergence of agentic AI architectures—systems that decompose complex problems into subtasks, invoke specialized tools, and iteratively refine their outputs through structured reasoning chains.[77] Applied to normative legal reasoning, an agentic system might not merely predict a statistically likely contract provision but instead identify the relevant jurisdictional precedents, assess the range of commercially negotiated outcomes in comparable transactions, model the risk exposure under alternative formulations, and synthesize these inputs into a recommendation that reflects identifiable reasoning rather than opaque pattern matching.[78] This is something other than normative judgment in the philosophical sense; the system is producing outputs rather than making value choices. Yet those outputs may be functionally indistinguishable from reasoned legal advice for an increasingly wide range of practical contexts.[79]
Preference Alignment and Client-Specific Calibration
A second pathway toward more effective normative output involves the development of preference-aligned AI systems that can be calibrated to specific clients’ risk tolerances, commercial objectives, and strategic priorities.[80] The normative reasoning limitation described in Section III.C is, in part, a function of the fact that current systems lack access to the particularized context that informs human judgment—they do not know what the client values, what the counterparty is likely to tolerate, or what the deal’s strategic logic requires.[81] But there is no structural reason why that context cannot be supplied to an AI system through properly designed input frameworks, institutional knowledge bases, and iterative feedback mechanisms that train the system on a firm’s or client’s historical preferences.[82]
Law firms that build proprietary datasets of negotiated outcomes—linking specific deal contexts to the normative choices that attorneys made—can create training signals that capture not merely what provisions were used but the reasoning and contextual factors that drove their selection.[83] A system trained on such data, and further refined through reinforcement learning from attorney feedback, could develop a functional capacity to recommend risk-allocation structures that are responsive to the same contextual inputs that inform human judgment, even if its internal process differs fundamentally from human deliberation.[84] The competitive implications are significant: firms that invest early in building these feedback-integrated AI systems will develop proprietary analytical capabilities that are difficult for competitors to replicate, precisely because the value of the system is not in the model architecture but in the firm-specific training data and preference alignment.[85]
Multimodal Integration and the Social Context Gap
The social and relational limitations identified in Section III.D—the inability of AI systems to read a room, assess credibility, or navigate political dynamics—represent perhaps the most durable boundary between AI capability and professional legal judgment.[86] Yet, even here, developments in multimodal AI systems suggest that the boundary will shift, if not dissolve.[87] Systems that can process not only text but also audio tone, facial expression, communication patterns, and behavioral data from negotiation histories are beginning to develop rudimentary capacities for social inference that, while far from human-level social intelligence, are meaningfully more sophisticated than text-only analysis.[88]
The more realistic near-term application is not AI that replaces the senior partner’s social intelligence at the negotiating table, but AI that augments it by providing structured analysis of counterparty behavior patterns, historical negotiation tendencies, and probabilistic assessments of walk-away thresholds based on comparable transaction data.[89] A system that can tell a negotiator, based on analysis of the counterparty’s responses across multiple deal points, that the counterparty is likely to concede on indemnification caps but remain firm on the MAC formulation provides useful strategic intelligence—even though it is not exercising social judgment in the way that a human lawyer does.[90]
The Functional Equivalence Problem and Its Professional Implications
The convergence of these developmental trajectories creates what might be called the functional equivalence problem: as AI systems produce outputs that increasingly resemble the products of normative legal judgment, the profession must confront the question of whether functional equivalence is sufficient for professional purposes, or whether the exercise of genuine human judgment remains a necessary condition for the delivery of legal services.[91] This question is not purely philosophical—it has direct implications for professional responsibility, malpractice liability, and the definition of legal practice itself.[92]
If an AI system can produce a MAC clause formulation that an experienced Delaware practitioner would assess as commercially reasonable, contextually appropriate, and legally sound—not because the system exercised judgment but because its training data, preference alignment, and retrieval capabilities converged on the same output that judgment would have produced—the practical question is whether the legal system should treat that output differently from human work product.[93] The answer, at least under current professional responsibility frameworks, is almost certainly yes: the obligations of competence, supervision, and accountability that attach to legal judgment presuppose a human professional who can be held responsible for the reasoning behind the decision, not merely for the decision itself.[94] But the pressure that functionally equivalent AI outputs will place on this framework should not be underestimated, particularly as clients become aware that AI-generated work product may be comparable in quality at dramatically lower cost.[95]
The profession’s response to this trajectory will define the next generation of legal practice. Firms that treat AI as merely an efficiency tool for document production will find themselves unprepared for systems that can contribute meaningfully to the analytical and advisory dimensions of transactional work.[96] Firms that invest in developing human-AI collaborative workflows—in which AI provides structured analytical inputs and the attorney provides the normative judgment, accountability, and client-facing interpretation that the system cannot—will be positioned to deliver higher-quality legal services more efficiently than either human or machine could achieve alone.[97] The competitive advantage in the AI-integrated legal market will accrue not to the firms with the most advanced AI systems, nor to the firms with the most experienced attorneys, but to the firms that most effectively integrate the two.
Conclusion
This paper has argued that AI’s impact on transactional legal practice is best understood through the lens of reconfiguration rather than replacement.[98] AI materially improves efficiency in document-intensive workflows, enabling faster and more consistent performance of the pattern-recognition tasks that constitute a large portion of transactional legal work by volume. It is structurally limited in the tasks that constitute legal practice by value: reasoning about novelty, making normative judgments, and navigating the social and political context of complex deals.
The consequences of this reconfiguration are already visible: workflow compression, disrupted training pipelines, billing model pressure, new malpractice vectors, and emerging professional responsibility obligations. They will become more pronounced over time as AI capabilities continue to develop and as firms, clients, and bar authorities respond to the challenges and opportunities that AI deployment creates.
The central analytical contribution of this paper is the identification of the durable boundary between AI’s capabilities and the domain of legal judgment. That boundary is not a function of limitations that incremental improvements to existing architectures will straightforwardly resolve—it reflects deep features of current probabilistic systems that constrain their reliable performance of the normative, novel, and socially embedded tasks that define professional legal work. Understanding this boundary is the necessary foundation for thinking clearly about how law firms should deploy AI, how lawyers should develop the skills needed in an AI-integrated practice, and how the profession should adapt its regulatory and educational frameworks to serve the legal system that is emerging.
That said, this boundary should not be understood as permanently fixed. The most consequential near-term development is the shift from isolated generative tools toward integrated legal intelligence platforms that combine document analysis, probabilistic risk assessment, and structured reasoning within a unified workflow. As these systems mature, the execution layer that AI currently dominates will expand into tasks that today require senior associate and junior partner involvement, and the boundary between AI-appropriate and judgment-intensive work will continue to move. This trajectory carries immediate implications: law firms must redesign associate development around the assumption that AI will absorb an expanding share of execution work and identify how lawyers can build the contextual knowledge and practical intuition that high-volume tasks currently provide. Professional responsibility frameworks will similarly need to evolve as AI outputs become harder to verify through conventional human review. Most fundamentally, as next-generation systems close the contextual gap, the profession will need to articulate with greater precision what remains irreplaceable—and invest in developing it rather than assuming its permanence.
AI will not replace lawyers. But it is changing what lawyers do, how they develop, how they are compensated, and what it means to practice law at the highest levels of the profession. These changes deserve serious, sustained analytical attention—the careful, structural analysis that the legal profession is particularly well positioned to provide.
See John O. McGinnis & Russell G. Pearce, The Great Disruption: How Machine Intelligence Will Transform the Role of Lawyers in the Delivery of Legal Services, 82 Fordham L. Rev. 3041, 3042 (2014). ↑
See Daniel Martin Katz et al., GPT-4 Passes the Bar Exam, 382 Phil. Transactions Royal Soc’y A 20230254 (2024). ↑
See Richard Susskind, Tomorrow’s Lawyers: An Introduction to Your Future 47 (2d ed. 2017). ↑
See id. at 48. ↑
See Ronald J. Gilson, Value Creation by Business Lawyers: Legal Skills and Asset Pricing, 94 Yale L.J. 239, 240 (1984). ↑
See Harry Surden, Artificial Intelligence and Law: An Overview, 35 Ga. St. U. L. Rev. 1305, 1308 (2019). ↑
See Maura R. Grossman & Gordon V. Cormack, Technology-Assisted Review in E-Discovery Can Be More Effective and More Efficient Than Exhaustive Manual Review, 17 Rich. J.L. & Tech. 11, 13 (2011). ↑
See Frank Pasquale, A Rule of Persons, Not Machines: The Limits of Legal Automation, 87 Geo. Wash. L. Rev. 1, 6 (2019). ↑
See Surden, supra note 6, at 1321. ↑
See Thomas H. Davenport & Julia Kirby, Only Humans Need Apply: Winners and Losers in the Age of Smart Machines 83 (2016). ↑
See id. at 85. ↑
See William D. Henderson, Legal Market Landscape Report 12 (Inst. for the Advancement of the Am. Legal Sys. 2018). ↑
See id. at 14. ↑
See Jan Hatzius et al., Goldman Sachs, The Potentially Large Effects of Artificial Intelligence on Economic Growth 4 (2023). ↑
See Pasquale, supra note 8, at 18. ↑
See ABA Comm. on Ethics & Pro. Resp., Formal Op. 512 (2024). ↑
See Gilson, supra note 5, at 243. ↑
See Model Rules of Pro. Conduct r. 1.1 (Am. Bar Ass’n 2020). ↑
See Susskind, supra note 3, at 112. ↑
See Grossman & Cormack, supra note 7, at 13. ↑
See Surden, supra note 6, at 1310. ↑
See Jason Wei et al., Emergent Abilities of Large Language Models, 2 Transactions on Mach. Learning Rsch. 1, 2 (2022). ↑
See Jason Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, 36 Advances in Neural Info. Processing Sys. 24824, 24825 (2022). ↑
See Tom Brown et al., Language Models Are Few-Shot Learners, 33 Advances in Neural Info. Processing Sys. 1877, 1878 (2020). ↑
See Surden, supra note 6, at 1315. ↑
See Hatzius et al., supra note 14, at 8. ↑
See Pasquale, supra note 8, at 29. ↑
See Wei et al., supra note 23, at 24,830. ↑
See Surden, supra note 6, at 1308, 1315. ↑
See Grossman & Cormack, supra note 7, at 35. ↑
See Pasquale, supra note 8, at 5, 29–30; see also McGinnis & Pearce, supra note 1, at 3049. ↑
See François Chollet, On the Measure of Intelligence 11, 20 (Nov. 25, 2019) (arXiv preprint). ↑
See Pasquale, supra note 8, at 22, 25, 29–30. ↑
See Surden, supra note 6, at 1316, 1322, 1324–25. ↑
See McGinnis & Pearce, supra note 1, at 3049. ↑
See Grossman & Cormack, supra note 7, at 45. ↑
See Pasquale, supra note 8, at 22. ↑
See Ryan Calo, Artificial Intelligence Policy: A Primer and Roadmap, 51 U.C. Davis L. Rev. 399, 405 (2017). ↑
See Pasquale, supra note 8, at 5, 29–30. ↑
See Surden, supra note 6, at 1323–24, 1326. ↑
See Gilson, supra note 5, at 255. ↑
See Surden, supra note 6, at 1325. ↑
See Hatzius et al., supra note 14, at 6. ↑
See Henderson, supra note 12, at 18. ↑
See Susskind, supra note 3, at 95. ↑
See Henderson, supra note 12, at 22. ↑
See ABA Comm. on Ethics & Pro. Resp., Formal Op. 512, at 4 (2024). ↑
See id. at 4, 10. ↑
See Davenport & Kirby, supra note 10, at 91. ↑
See Henderson, supra note 12, at 25. ↑
See Gilson, supra note 5, at 243. ↑
See Press Release, Litera, Litera Agrees to Acquire Leading Machine Learning Contract Analysis Company, Kira Systems (Aug. 10, 2021). ↑
See Akorn, Inc. v. Fresenius Kabi AG, No. 2018-0300, 2018 WL 4719347, at *1 (Del. Ch. Oct. 1, 2018). ↑
See Gilson, supra note 5, at 258. ↑
See McGinnis & Pearce, supra note 1, at 3052. ↑
Model Rules of Pro. Conduct r. 1.1 (Am. Bar Ass’n 2020). ↑
Id. r. 1.1 cmt. 8. ↑
See ABA Comm. on Ethics & Pro. Resp., Formal Op. 512, at 3 (2024). ↑
See id. at 5. ↑
Model Rules of Pro. Conduct r. 5.1, 5.3. ↑
See ABA Comm. on Ethics & Pro. Responsibility, Formal Op. 512, at 7. ↑
See Model Rules of Pro. Conduct r. 1.6. ↑
See Pasquale, supra note 8, at 30. ↑
See, e.g., Mata v. Avianca, Inc., 678 F. Supp. 3d 443, 448 (S.D.N.Y. 2023) (sanctioning attorneys for submitting AI-generated brief containing fabricated citations); see also Model Rules of Pro. Conduct r. 3.3. ↑
See Hatzius et al., supra note 14, at 9. ↑
See Henderson, supra note 12, at 30. ↑
See Susskind, supra note 3, at 103. ↑
See id. at 107. ↑
See Davenport & Kirby, supra note 10, at 98. ↑
See Henderson, supra note 12, at 34. ↑
See Gilson, supra note 5, at 262. ↑
See Dario Amodei, Machines of Loving Grace: How AI Could Transform the World for the Better (Oct. 2024). ↑
See Surden, supra note 6, at 1330. ↑
See ABA Comm. on Ethics & Pro. Resp., Formal Op. 512, at 4–5 (2024). ↑
See Hatzius et al., supra note 14, at 11. ↑
See Pasquale, supra note 8, at 35. ↑
See McGinnis & Pearce, supra note 1, at 3055. ↑
See Surden, supra note 6, at 1335. ↑
See ABA Comm. on Ethics & Pro. Resp., Formal Op. 512, at 8 (2024). ↑
See Grossman & Cormack, supra note 7, at 48. ↑
See Susskind, supra note 3, at 118. ↑
See Gilson, supra note 5, at 265. ↑
See Henderson, supra note 12, at 36. ↑
See Davenport & Kirby, supra note 10, at 105. ↑
See Pasquale, supra note 8, at 40. ↑
See McGinnis & Pearce, supra note 1, at 3060. ↑
See Calo, supra note 38, at 418. ↑
See Surden, supra note 6, at 1340. ↑
See Susskind, supra note 3, at 125. ↑
See Gilson, supra note 5, at 270. ↑
See ABA Comm. on Ethics & Pro. Resp., Formal Op. 512, at 10 (2024). ↑
See Henderson, supra note 12, at 40. ↑
See Pasquale, supra note 8, at 45. ↑
See ABA Comm. on Ethics & Pro. Resp., Formal Op. 512, at 4 (2024). ↑
See Hatzius et al., supra note 14, at 14. ↑
See Gilson, supra note 5, at 275. ↑
See Henderson, supra note 12, at 42. ↑
See Susskind, supra note 3, at 130. ↑
