The Clock That Won't Wait: Why Law Firms That Only Buy Technology Have Already Lost

The institutional absorption of artificial intelligence in the practice of law: an analysis of the two-clocks paradigm drawn from Zack Shapiro's essay.

1. Introduction: Shapiro's diagnosis and why the moment matters

1.1. Framing: the gap between technological capability and legal institutions

The legal profession stands at an unprecedented crossroads. Generative artificial intelligence has reached a level of sophistication that, barely five years ago, would have been dismissed as science fiction: frontier models read entire case files, break complex problems into subtasks, work in parallel, draft documents, check citations, and return finished legal work product with no human intervention beyond the initial instruction (1). The institutions that have traditionally held a monopoly on legal advice — the large law firms — move, however, at a pace that the technology itself has rendered obsolete. This disconnect, which forms the core of Zack Shapiro's essay (1), is in our view the single most consequential fact for the economics of professional services this decade.

Shapiro's diagnosis is not a superficial observation. In the months before publishing his essay, he worked directly with two of the largest and oldest law firms in the United States, helping them integrate AI into the daily work of their practice groups (1). That first-hand experience allows him to argue that the public debate — focused almost exclusively on model capability — misses the real problem: an organization's capacity to absorb that capability and convert it into effective practice. In short, the bottleneck no longer lies in artificial intelligence itself, but in institutional fitness to assimilate it (1).

This report analyzes Shapiro's "two clocks" thesis, subjecting it to legal, economic, and historical scrutiny, testing his claims against the sources he himself cites and against regulatory developments that postdate the essay's publication — developments that, as explained in Section 11, have materially altered some of the regulatory deadlines in force at the time the essay was written.

1.2. The piece under analysis and its context

On July 8, 2026, Zack Shapiro published a lengthy essay on the platform X (formerly Twitter) titled "The Two Clocks" (1). Shapiro is a partner at Rains LLP, a two-lawyer firm known in legal-tech circles for its "Claude-native" approach — that is, for having rebuilt its professional practice around Anthropic's models. His profile is not that of an innovation theorist but of a practicing lawyer who has experienced first-hand both the limitations and the transformative potential of these tools (1).

The timing of the publication is significant. By mid-2026, the generative AI market is undergoing a phase of accelerated maturation: technical reports document notable productivity multiples in sectors such as software development (4), while large firms undertake strategic moves of considerable scale, such as Kirkland & Ellis's announced $500 million investment (8), and private-equity-backed "AI-native" operators begin winning talent and clients (11). Shapiro's analysis lands in the middle of that strategic ferment, proposing a roadmap that avoids both naive technological optimism and paralyzing institutional conservatism.

1.3. The central thesis: the problem is not intelligence, but absorption

Shapiro's thesis breaks down into three propositions. The first identifies two diverging clocks: the "fast clock," which measures technological progress and advances every few weeks through smarter models, wider context windows, and more capable agents; and the "slow clock," which measures the speed at which institutions can put that technology to use — a pace set by committees, approvals, pilots, policies, and trainings (1). For Shapiro, the distance between the two clocks is the single most important fact in business today (1).

The second proposition holds that the initial failure of law firms testing these tools does not stem from any weakness in the models. Shapiro documents a recurring pattern among partners at large firms: they upload a document and instruct the model to review it and flag issues; the response is competent but generic, confirming the lawyer's initial skepticism (1). When, instead, the instruction incorporates the client's context, the commercial objective, the dynamics with the counterparty, and the required confidence level, the result changes radically — despite using the same model and the same document (1). Shapiro's conclusion is that the model was never too weak; the institution simply had not learned how to instruct it.

The third proposition identifies the strategic bottleneck: if the problem is absorption rather than intelligence, then the most significant business opportunity does not lie in building the models — the labs' business — but in redesigning the workflows of the organizations that use them. Shapiro calls this the "absorption business," which he defines as moving capability from the fast clock to the slow clock without breaking the institution in the process (1).

1.4. Methodology and sources of this report

This report follows a documentary and critical-analysis methodology. Its primary source is Shapiro's essay (1); its secondary sources are those he himself cites in his notes, cross-checked and, where necessary, updated with information published after the essay — particularly on regulatory matters, where deadlines have shifted substantially, as detailed in Section 11.

The analysis is organized around three axes: a historical-economic axis, examining the electrification-of-factories analogy (2)(3) and the Coca-Cola-versus-General-Electric parable (16); an empirical-strategic axis, centered on cases such as Kirkland & Ellis (8)(9) and Sullivan & Cromwell (7), as well as the rise of new competitors (11)(12)(13); and a regulatory-ethical axis, addressing the implications of generative AI for the professional standard of care, civil liability, and the regulatory framework — European, Spanish, and American — as verified as of this report's closing date.

2. The two-clocks paradigm: theoretical foundations

2.1. Defining the fast clock: the exponential evolution of frontier models

The "fast clock" represents the speed at which advances in AI model capability occur — a speed Shapiro describes as approaching exponential in the domains relevant to knowledge work (1). A lawyer in 2016 would have called today's frontier-model capabilities science fiction; that same professional in 2026, faced with an equivalent demonstration, simply asks whether IT has approved the tool (1). That normalization of the extraordinary conceals the true scale of the institutional lag.

2.2. The empirical evidence: Anthropic and code automation

The software sector offers the clearest evidence of how the fast clock operates. Shapiro relies on the Anthropic Institute report "When AI Builds Itself," by Marina Favaro and Jack Clark (June 4, 2026), which found that more than 80% of the code merged into Anthropic's production repository in May 2026 was written by Claude, up from single-digit percentages before the launch of Claude Code in February 2025 (4). The same report cites an internal survey from March 2026 of roughly 130 researchers, whose median respondent estimated their productivity at around four times what it would be without AI — a figure the report itself cautions should be treated carefully, given the well-known tendency of such self-assessments to run optimistic (4). Engineering work, Shapiro notes, has shifted from direct production to orchestration (1).

2.3. Defining the slow clock: the institutional inertia of large law firms

The "slow clock" measures the pace at which institutions manage to integrate technology into daily operations: committees, approvals, pilots, policies, trainings, and, underlying all of it, the hope that nothing fundamental will need to change before the next compensation cycle (1). Shapiro precisely describes the reality inside a typical AmLaw 50 firm, where the average lawyer uses the most powerful technology ever built to clean up time entries, summarize documents no one intends to read, and draft emails scheduling the next meeting — while the capabilities that would actually matter, such as substantive delegation or briefing the model the way one would brief a trusted associate, remain untested (1). The slowness, however, is not irrational: a large firm's profits rest on the billable hour and associate leverage, and both pillars are threatened by a technology that saves precisely the hours that sustained the model (1).

2.4. The innovator's dilemma in its purest form

Shapiro's description tracks Clayton Christensen's innovator's dilemma: the firms with the most to gain from rebuilding are often the ones whose current economics make rebuilding most painful (1). The legal sector adds a further temporal distortion: a managing partner drawing a high annual salary, close to the end of their career, who transforms the firm incurs immediate disruption, compensation friction, and lower short-term billables, while the benefit of the transformation may only materialize after they have left (1). Running out the clock benefits the person leaving; fixing it benefits their successors.

2.5. The parallel with software — and its translation to law

The parallel between software and law is not complete, and Shapiro is aware of it: law has no "compiler." A defective contract does not throw an immediate error message; it appears to function normally until a counterparty exercises a consent right no one considered, or an indemnification clause exposes an unwitting client to unlimited liability (1). This absence of immediate feedback makes legal AI harder to evaluate than coding AI, without implying, in Shapiro's view, that the tool is actually less powerful. The difference lies in the method of validation: where an engineer runs the code, a lawyer must subject the AI's output to substantive scrutiny — citations, reasoning, the commercial impact of clauses — a form of oversight closer to reviewing an associate's work than to executing a script.

2.6. The absence of a "compiler" as a defining factor in legal practice

The absence of a compiler has profound implications for adoption strategy. A coding error is caught instantly; a contractual error may go unnoticed for years and, when it surfaces, its consequences can be catastrophic. This explains the asymmetric fear Shapiro identifies among partners at large firms: someone who quietly rebuilds a workflow gets, at best, a polite nod; someone who signs a filing containing fabricated AI citations may carry a headline that follows them for the rest of their career (1). The Sullivan & Cromwell case, analyzed in Section 10, illustrates that not even the most prestigious firms are immune to this risk (7). The right response, Shapiro argues, is not to abandon the tool but to build robust verification processes.

2.7. Partial conclusion: the divergence of speeds as a structural fact

The distance between the two clocks is not cyclical but structural: it flows from the nature of complex organizations, from the incentive structures of their leaders, and from the absence of immediate feedback mechanisms in legal work. As the historical electrification analogy demonstrates (Section 5), institutions can be redesigned to absorb new technologies — but that redesign demands a deliberate, sustained effort that goes well beyond buying licenses or forming innovation committees.

3. The failure of the first instinct: the problem of generic instructions

3.1. Shapiro's experiment with large firms

The gap between the two clocks is not, for Shapiro, a theoretical exercise but the result of repeated empirical observation: in the months before publishing his essay, he met with partners at some of the country's largest firms and asked them to show him what they had actually tried to do with AI (1). The pattern that emerged was remarkably consistent: a lawyer with twenty or thirty years of experience would upload a document and issue a brief instruction — "review this agreement and flag the issues" — the model would return a competent but generic response, and the lawyer would conclude that the tool was interesting for summaries but not ready for real work (1). According to Shapiro, the failure lies not in the tool but in how the interaction was conceived.

3.2. The fallacy of the brief query

The generic instruction is, above all, a delegation failure rather than a technological one. The model does exactly what it is asked — reviewing the document and flagging issues — but it cannot weigh those issues against the client's business, negotiating posture, or acceptable risk level, because no one told it to: it read the contract, but not the deal (1).

3.3. Rebuilding the instruction

The decisive turn in Shapiro's experiment came when, instead of adding more complex formulas, the team rebuilt the instruction the way an experienced partner would when handing work to a trusted associate: client background, negotiating posture, commercial objective, dynamics with the counterparty, relevant clauses, arguments that should not be raised, the confidence level required, a format useful to the client, and checks the AI should run before returning its answer (1). The result, with the same model and the same document, changed radically: what had been a generic checklist became decision-oriented analysis, complete with priorities, concrete warnings, and alternative drafting proposals (1).

3.4. The anatomy of a serious instruction

Shapiro breaks the effective instruction down into six elements: task, background, criteria, constraints, deliverable, and verification (1). None of these is technical — it is the same natural language a partner already uses with an associate. As Shapiro observes, the lawyers who adapt best to this approach are not necessarily the youngest or the most technical, but the best delegators (1).

3.5. Implications for legal training and practice

This anatomy requires experienced lawyers to externalize tacit knowledge, turning into explicit priorities and constraints things that previously operated almost unconsciously. The process also transforms the partner's role, which shifts from merely reviewing drafts to designing the criteria that govern their production — without diminishing the partner's responsibility for the final result (1).

3.6. Partial conclusion: the model wasn't weak; the institution didn't know how to instruct it

The failure of early AI tests at large firms was not caused by insufficient models but by insufficient instructions. That finding, however, does not by itself resolve the problem of institutional absorption: teaching lawyers to write better instructions is necessary but not sufficient unless that knowledge becomes a reusable institutional asset — a question addressed in Section 9.

4. The innovator's dilemma in the legal sector

4.1. The economic structure of the large law firm

The slowness of large firms is not a matter of ignorance; it flows from the logic of a business model built on the billable hour and associate leverage (1). Precisely the work AI performs best — first drafts, diligence, document review, citation checking, summaries, comparisons, formatting — is exactly the work the BigLaw pyramid exists to sell (1).

4.2. The threat to the two pillars of the business model

The threat operates on two fronts: every hour AI saves is an hour that cannot be billed the old way, and compressed leverage reduces the need for a large junior associate bench, flattening the pyramid and eroding the traditional training pipeline (1).

4.3. The perverse incentive

The individual response — experimenting privately, without altering the billing structure — is rational for the partner but does not transform the institution; the institutional response — committees, policies, pilots — is equally rational and yet perpetuates the problem (1). This is the innovator's dilemma in its purest form: waiting is rational right up until it's fatal (1).

4.4. The time-horizon mismatch

A managing partner drawing a high salary and nearing the end of their career faces a simple calculation: transforming the firm means immediate costs and deferred benefits that a successor, not the partner, will likely reap (1). This mismatch is not a moral failing but a structural feature of large-firm governance, whose compensation horizon — annual — does not match the return horizon of a serious digital transformation.

4.5. Paralysis by fear

The slow clock also feeds on fear: the asymmetric fear of becoming the cautionary tale, and the deeper fear that AI will replace the lawyer altogether. Predictions such as the one made by Anthropic CEO Dario Amodei — in an interview with Jim VandeHei and Mike Allen of Axios on May 28, 2025, warning that AI could eliminate up to half of entry-level white-collar jobs within one to five years (5) — weigh heavily on this second fear. Shapiro disagrees with that prediction as it applies to lawyers, though he acknowledges a partner need not believe it to feel its pressure.

4.6. The institutional response: committees, pilots, policies, and "AI theater"

Faced with this pressure, the typical large-firm response — a working group, a responsible-use policy, a pilot with a vendor, a training day, a panel on responsible innovation — amounts to what Shapiro calls "AI theater": an institutional defense mechanism that creates the appearance of engaging with change without actually altering workflows (1). The test, Shapiro says, is simple: ask these firms which specific workflow has changed, how much faster it has become, what has improved for the client, and what the firm now does differently on a real matter — and concrete answers are scarce (1).

4.7. The S&C case: the risk of hallucinations in court

Sullivan & Cromwell, one of the world's most prestigious firms, was forced in April 2026 to submit a letter of apology to Judge Martin Glenn of the U.S. Bankruptcy Court for the Southern District of New York, after an emergency motion filed on April 9 in the Chapter 15 proceedings of Prince Group contained dozens of inaccurate citations and other errors, including cases invented by the AI model used (7). The error was caught by opposing counsel, Boies Schiller Flexner, and widely reported by Bloomberg Law and Reuters (7). The case shows that institutional prestige offers no immunity to this kind of failure; what protects a firm, in Shapiro's view, is the verification process, not its prior reputation (1).

4.8. Partial conclusion: the incentive structure as the primary barrier

The barriers to AI absorption in the legal sector are neither technical nor even mainly cultural, but structural: hourly billing, leverage, the time-horizon mismatch in incentives, and fear of liability together produce a system that penalizes radical transformation and rewards the simulation of innovation.

5. The lesson of history: electricity, line shafts, and unit drive motors

5.1. The electrification of factories in the early twentieth century

Shapiro turns to the history of industrial electrification to give his argument historical depth. When electricity replaced steam in the late nineteenth and early twentieth centuries, factory owners pulled out the steam engine, installed an electric motor, and left the rest of the factory — including the long central line shaft — exactly as it was (1). The decision was understandable: the electric motor was cleaner and more reliable, and redesigning the plant would have required a reengineering effort that did not seem justified in the short term (2)(3).

5.2. Three decades of persistence in the line-shaft design

That individually rational decision, however, did not translate into aggregate productivity gains: for nearly thirty years, economists observed that productivity improvements fell far short of what the new technology's potential seemed to promise. Paul David documented this paradox in "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox" (American Economic Review, vol. 80, no. 2, 1990), and Warren D. Devine, Jr. expanded on the analysis in "From Shafts to Wires: Historical Perspective on Electrification" (Journal of Economic History, vol. 43, no. 2, 1983): a general-purpose technology only unlocks its potential once the organization of work that uses it is redesigned, and that redesign can lag a full generation behind the invention itself (2)(3).

5.3. Redesigning the workflow: a motor on every machine

The decisive shift came in the 1920s, with the abandonment of the central line shaft in favor of individual motors on each machine — which literally required "tearing up the floor" of the factory to redistribute production according to the product's logical flow rather than proximity to the drive shaft (1). The result was the productivity leap economists had been expecting for decades.

5.4. The Coca-Cola versus General Electric parable

When mechanical refrigeration became cheap in the early twentieth century, the obvious bet seemed to be on the machine manufacturers — General Electric, Westinghouse, Frigidaire. The biggest beneficiary, however, turned out to be Coca-Cola, a regional Atlanta-based fountain-syrup company that, under Robert Woodruff's leadership, set out in the 1920s to put its product within arm's reach of desire in every corner of the world (7)(16). Coca-Cola did not manufacture a single refrigerator; it understood, before anyone else, what cheap cold made possible, and rebuilt its business around that understanding.

5.5. Translation to the legal sector

Translated to law, "tearing up the floor" means abandoning the cascading delegation of routine tasks and hourly billing, and instead reorganizing practice around the models' ability to perform those tasks with reduced supervision and high reliability — not to replace the lawyer, but to reconfigure the relationship between production and judgment (1).

5.6. Strategic implications

The historical lesson relativizes the debate framed solely around model capability: the competitive advantage will not lie in access to the best models — presumably available to any firm — but in the ability to build them into workflows that reflect a firm's own judgment and standards, an asset that should remain under the firm's control rather than the technology vendor's (1).

6. The Kirkland & Ellis case: the $500 million strategic bet

6.1. Context

Kirkland & Ellis closed 2025 with $10.56 billion in revenue and $11.1 million in profits per equity partner, record figures for the sector (9). That same exposure to the private-equity segment — the engine of its growth — makes it especially vulnerable to pressure from clients who execute the same deal structures dozens of times a year and are starting to ask why work a machine can draft is still billed at associate rates (1).

6.2. Client pressure: Blackstone

In February 2026, Bloomberg Law and Law.com reported that Blackstone — Kirkland's largest private-equity client — paid the firm $87.8 million in fees in 2025, down from a record $101.3 million in 2024, even as Kirkland's overall revenue grew by 20% (10).

6.3. The threat from the other side: Norm Law and AI-native competitors

Norm Law, an AI-native legal platform, launched in November 2025 backed by Blackstone through its parent Norm Ai, with Bain Capital and Vanguard among its investors; in January 2026 it named Michael Schmidtberger — who had chaired Sidley Austin's executive committee for seven years — as its president (11). Crosby has raised more than $85 million from Sequoia, Index, and Lux, and Eudia raised a Series A of up to $105 million before launching an AI-augmented legal firm in Arizona (11).

6.4. The announced investment

Kirkland announced in May 2026 an investment of roughly $500 million over three to four years, with an initial outlay of about $100 million in 2026 (8). Its first product followed a week later: on June 4, 2026, Kirkland and Palantir Technologies unveiled a fund-formation engine, exclusive to the firm, built to serve the more than 1,000 lawyers in its investment funds practice, with an architecture Kirkland describes as model-independent (12).

6.5. Critical analysis

The question Shapiro raises is whether this investment builds a durable competitive advantage or, instead, installs a very expensive electric motor on top of the old line shaft: a proprietary platform is only worth as much as the transformed practice it is connected to (1).

6.6. Shapiro's warning: acquisition is not absorption

Buying the best platform and hiring the best engineers is not enough if lawyers keep working the same way. The risk, Shapiro notes, is that an investment of this size becomes a substitute for real change rather than its catalyst (1).

6.7. Palantir's role and the data architecture

Shapiro cites an interview with Palantir CEO Alex Karp on CNBC on July 1, 2026, given on the occasion of the announcement of a sovereign AI partnership between Palantir and Nvidia, in which Karp argued that companies should own the means of production of their own AI — compute, models, data, and competitive advantage — rather than renting them from third parties (15). Shapiro shares that diagnosis but adds a caveat: choosing a platform does not, by itself, settle who designs the workflows that run on top of it, and that design should remain in the hands of the firm's own lawyers, not the vendor's engineers (1).

6.8. Partial conclusion

The Kirkland case is, for Shapiro, a proving ground for the entire sector: if the world's most profitable law firm succeeds in transforming its practice, it will demonstrate that absorption is possible even for the most entrenched incumbents; if it merely installs technology without redesigning the work, it will serve as a warning for everyone else.

7. The absorption business: a new professional function

7.1. Definition

Absorption, in Shapiro's terminology, is the process of moving capability from the fast clock to the slow clock without breaking the institution along the way (1). Today, he argues, that task usually falls to a single person who knows the work well enough to do it the old way and knows the tools well enough to rebuild it the new way (1).

7.2. The forward-deployed engineer as archetype

Palantir coined the "forward-deployed engineer" role twenty years ago — an engineer embedded in a client's operations who rebuilds the work around the software (1). What long looked like a Palantir peculiarity has, by spring 2026, become the position every investor is trying to replicate.

7.3. The rise of AI-native services firms

OpenAI announced on May 11, 2026, the creation of the OpenAI Deployment Company, an independent unit with more than $4 billion in committed capital led by TPG, launched alongside the acquisition of Tomoro, a consultancy that brought roughly 150 forward-deployed engineers on day one (13). Days earlier, on May 4, 2026, Anthropic announced, together with Blackstone, Hellman & Friedman, and Goldman Sachs, an AI-native enterprise services firm capitalized at roughly $1.5 billion (13). Both moves respond, as Shapiro notes, to a 2025 MIT study finding that 95% of organizations saw no measurable return on their generative AI investments despite accumulated spend (13)(14).

7.4. The conflict of interest: who does the engineer work for?

In these models, a forward-deployed engineer works for the vendor, not the client. That relationship is acceptable for a factory whose competitive edge does not rest on its logistics, but not for a law firm whose work depends on confidentiality and whose procedures encode its own method: if those procedures are written onto the vendor's rails, they tend to migrate into the vendor's product, where the firm next door can subscribe to them (1). For a law firm, Shapiro argues, that person should work for the institution, not the vendor.

7.5. The need to internalize redesign capability

Internalizing absorption does not necessarily require building proprietary platforms from scratch — Kirkland has chosen that path, but it is not the only one — it requires that the substantive knowledge embedded in workflows belong to the firm, even if the infrastructure is rented from a third party (1).

7.6. Change management as a new discipline

Shapiro reclaims change management — "the least glamorous phrase in business" — as one of the most valuable disciplines of the moment, not in its traditional form of stakeholder maps and adoption dashboards, but as the process of turning expert judgment into automated procedures a machine can execute and an institution can oversee (1). A generalist consultant can sketch a reorganization of reporting lines from across a table; they cannot design the thousand small decisions that make up a specific legal matter, because those decisions increasingly are the practice of law itself (1).

7.7. The impossibility of outsourcing tacit knowledge

The knowledge of the clause that once cost a client money, or the argument a court will not accept, is not written in any manual: it is transmitted through daily practice and can only be extracted while the work is being done, with the active participation of the professional who holds it (1).

7.8. Partial conclusion

Absorption cannot be fully outsourced without risk: a firm's tacit knowledge is its most valuable asset, and handing it to a technology vendor puts that advantage within reach of competitors, sooner or later.

8. The transformation of legal work: from production to judgment

8.1. Legal production versus decision-making

Shapiro distinguishes between production — gathering, drafting, comparing, summarizing, checking, and organizing legal information — and decision-making — the moment a lawyer converts that material into a concrete recommendation to a client. Frontier models are extraordinarily effective at production; judgment about which risk matters or which argument to make remains, in Shapiro's view, a fundamentally human domain (1).

8.2. The fiction of hourly billing

For a century, firms have billed visible, quantifiable work — research, drafting, diligence, cite-checking — even though clients mainly valued the decision-making that work supported. Shapiro argues that hours were, in reality, simply the way firms chose to bill for what the client actually bought: the judgment of the partner signing the advice (1).

8.3. AI's effect on production

AI attacks production first: it drafts first drafts, compares documents, summarizes case files, checks citations, and performs tedious review imperfectly but fast enough that the old relationship between time and value stops holding (1).

8.4. The shift of value toward judgment

As production gets cheaper, the scarce input becomes the person capable of directing the machine, evaluating its output, and taking responsibility for the advice. Judgment, Shapiro argues, was always what the client was buying; the invoice merely obscured it (1).

8.5. The flaw in mass-displacement predictions

Shapiro distances himself from predictions like Dario Amodei's (5) and argues that the more likely arithmetic is not the replacement of half of all jobs, but the replacement of half of every job: the machine absorbs production, and at the top of the market, work intensifies around judgment, while at the bottom — where work is routine and stakes are low — legal services can indeed be commoditized (1).

8.6. The crisis of traditional training

Mechanical work was not just what firms sold — it was also how juniors learned. If AI compresses that repeated exposure to the raw material of judgment, firms will need to deliberately design training around decision-making, rather than relying on it to accumulate through sheer hours worked (1).

8.7. New hiring and training criteria

Shapiro suggests that hiring will need to weigh early signs of judgment, initiative, and business sense more heavily than traditional academic credentials, and that experiences such as a judicial clerkship or a stint at a bank or corporation may prove more formative than the mechanical work AI is absorbing (1).

8.8. The revaluation of judgment and a new pricing model

Firms that manage to decouple from hourly billing and price judgment — the scarce good — could substantially improve their margins, in a legal-sector echo of the Coca-Cola parable: putting elite legal judgment within reach of every difficult decision (1).

8.9. Partial conclusion

Production is being commoditized; judgment is being revalued. Firms that grasp this distinction will gain a significant advantage; those that cling to the hours-and-leverage model will face mounting pressure from clients and AI-native competitors.

9. Workflows and institutional knowledge: the infrastructure of the future firm

9.1. A prompt versus a workflow

A prompt tells the model what to do on a particular matter; a workflow tells it how a specific lawyer, practice group, or firm performs that category of work, reusably across every matter without needing to be re-explained (1). Some procedures are mechanical — how to mark up a document without corrupting it, how to run citation checking as a separate process; the most valuable ones are substantive — how a particular partner reviews a contract, when to ask for more data, when to reject a premise because the record doesn't support it (1).

9.2. Externalizing tacit knowledge

The main challenge is not technical but cognitive: getting senior lawyers to articulate what they do almost unconsciously — the phrase that raises their suspicion, the clause they reread because it once cost a client money — extracting that knowledge as the work happens (1).

9.3. The standing playbook as an intangible asset

The output of that process is what Shapiro calls the standing playbook: an asset specific to a given firm, practice group, or even individual partner, that encodes judgment, priorities, and quality standards, and that constitutes the true competitive moat of absorption (1).

9.4. The continuous-improvement loop

When the model misses a relevant clause, the partner corrects the error and folds it back into the workflow, so the next run already accounts for it: the fix benefits every future matter, not just the one that surfaced it (1).

9.5. Ownership of the method

The durable asset is not the technological wrapper but the firm's own method, written with enough precision that the model can follow it, lawyers can oversee it, and the institution can improve it over time — which is why it should not live inside a third party's product (1).

9.6. The risk of technological dependence

The model independence Kirkland's platform pursues (12) mitigates, but does not eliminate, the risk of dependence on a particular platform. Shapiro's proposed solution is intellectual sovereignty: the method must belong to the firm and be written so that it can run on whichever platform the firm chooses.

9.7. The role of technology infrastructure

AI labs and vendors like Palantir or Snowflake can build the data architecture, but the layer of work sitting above that infrastructure belongs to lawyers, because no software company holds the accumulated know-how of a specific firm's actual practice (1).

9.8. Partial conclusion

Workflows constitute the infrastructure of institutional knowledge. This work is slow, personal, and invisible on any org chart — but it is, in Shapiro's view, the only kind of AI spending that actually changes what a firm does (1).

10. Litigation and disputes in the age of AI: procedural risk and professional responsibility

10.1. The duty to supervise and the signing lawyer's responsibility

The duty to supervise, traditionally centered on the work of associates and paralegals, now extends to output generated by AI models. The Sullivan & Cromwell case (7) confirms that delegating to the machine does not relieve the lawyer of ultimate responsibility for the accuracy of filings submitted to courts.

10.2. The S&C case: lessons on hallucinations and quality control

As discussed in Section 4.7, the emergency motion Sullivan & Cromwell filed on April 9, 2026, in the Prince Group Chapter 15 proceeding contained inaccurate citations and cases invented by the model; partner Andrew Dietderich sent a letter of apology to Judge Martin Glenn on April 18, 2026 (7). The case confirms that institutional prestige does not prevent error; the verification process does.

10.3. U.S. case law on AI hallucinations has consolidated in 2026

Between February and July 2026, U.S. courts have accumulated numerous rulings on AI-generated hallucinations in court filings. The most significant for appellate practice is Fletcher v. Experian Information Solutions, Inc., 168 F.4th 231 (5th Cir. 2026), in which the Fifth Circuit sanctioned appellant's counsel for filing a response brief containing numerous AI-fabricated citations and assertions, compounding the error with a less-than-candid response when the court asked her to explain the mistakes; the Circuit imposed a monetary sanction and stressed that the lack of candor — more than the use of AI itself — justified the reprimand. Other jurisdictions have followed a similar logic, including the Alabama Supreme Court in Ibach v. Stewart, 2026 WL 1110659 (Ala. 2026), and various state courts in New York and Ohio. The common lesson is that a lawyer's candor about the error — rather than the mere existence of the hallucination — proves to be the decisive factor in the severity of the sanction.

10.4. The burden of independent citation and source verification

The anatomy of instruction Shapiro proposes includes verification as a separate element: the model should check its own references before returning an answer (1). This automation does not eliminate the lawyer's responsibility, but it reduces the risk that errors like those in the S&C case ever reach the court.

10.5. The professional standard of care in the use of AI tools

In Spain, the General Council of Spanish Lawyers (Consejo General de la Abogacía Española) approved, on April 10, 2026 (with amendments incorporated on April 13), Interpretive Circular 3/2026 of the Code of Professional Conduct, issued under Article 23 of Organic Law 5/2024 of November 11 on the Right of Defense, which governs the drafting, signing, and delivery of legal filings assisted by generative AI. The Circular treats AI as an auxiliary function subject to human oversight, never as a substitute for the professional, and places responsibility for AI-related errors on the signing lawyer's failure to verify and control the output — conduct that, in the most serious cases, can trigger the disciplinary regime of the General Statute of the Spanish Bar. It is the third circular in a series that began with Circular 1/2025 on professional secrecy (November 2025) and continued with Circular 2/2026 on client funds (January 2026).

In the United States, American Bar Association Model Rule 5.3, on lawyers' responsibility for the work of non-lawyer personnel, has been invoked by analogy to AI system output. In California, the state Senate passed, on January 29, 2026, by a vote of 39 to 0, Senate Bill 574 (Umberg), which would require lawyers not to input confidential information into public AI systems, to personally verify the accuracy of any AI-generated citation before including it in a court filing, and to prevent AI use from producing discriminatory outcomes; as of this report's closing date, the bill remained pending in the California Assembly, with a deadline of August 31, 2026 for passage, and so cannot yet be described as enacted law.

10.6. Ethical and disciplinary implications

The duty of confidentiality requires not entering client information into AI models that cannot guarantee data privacy; the duty of competence requires understanding the limitations of the tools used, including their propensity to hallucinate. The S&C case shows that breaching these duties can carry severe reputational consequences even when a firm formally apologizes to the court.

10.7. The need for internal verification protocols

A minimum protocol should include independent verification of every AI-generated legal citation, a check that the reasoning is consistent with applicable doctrine and case law, validation that the facts invoked match the record, and final review by a lawyer who assumes ultimate responsibility for the content.

10.8. Partial conclusion

The risk of hallucinations is real and can affect even the most prestigious firms. The answer is not to ban the tool but to build robust verification protocols that catch errors before they reach the court.

11. The regulatory and competition dimension

11.1. The European regulatory framework: the AI Act and its recent amendment

Regulation (EU) 2024/1689 on Artificial Intelligence (the AI Act) entered into force on August 1, 2024, and rolls out its obligations in stages. Its original text provided that obligations for high-risk AI systems under Annex III would begin applying on August 2, 2026, and those for high-risk systems embedded in products regulated under Annex I on August 2, 2027.

It is worth noting — and this is one of the points this report corrects relative to earlier versions of the analysis — that this timeline has since changed. On May 7, 2026, the Council of the European Union and the European Parliament reached a provisional political agreement on the so-called "Digital Omnibus on AI," part of the simplification package the Commission presented on November 19, 2025. The agreement pushes back application of the Annex III high-risk obligations to December 2, 2027, and the Annex I obligations to August 2, 2028; it also introduces a new prohibition, under Article 5, of AI systems designed to generate non-consensual intimate content ("nudification" tools) or child sexual abuse material, set to take effect December 2, 2026; and it delays the Article 50.2 synthetic-content labeling obligations from August 2 to December 2, 2026. The remaining Article 50 transparency obligations, along with the prohibitions already in force since February 2025, are unaffected. As of this report's closing date, the agreement was still pending formal adoption and publication in the Official Journal of the European Union — a step EU institutions expected to complete before August 2, 2026; until that happens, the AI Act's original timeline remains, formally, the law in force. Firms planning their compliance should therefore continue to treat August 2, 2026 as the operative reference date, while keeping in mind that the substantive deadline for Annex III high-risk systems has likely been extended to December 2027.

11.2. Impact of the amendment on legal-support systems

The AI Act classifies AI systems into four risk tiers. AI systems used in the administration of justice are among those that may qualify as high-risk under Annex III, which would impose transparency, data-governance, technical-documentation, and human-oversight obligations on their developers and users. For firms building proprietary tools — such as Kirkland & Ellis's fund-formation platform (8)(12) — whether the system falls into that category remains a first-order question, even though the deadline for actual enforcement has been extended under the Digital Omnibus. The AI Act does not replace data-protection law; it supplements it with a risk-based approach.

11.3. Data protection and client confidentiality

A law firm acts as a data controller under Article 4(7) of Regulation (EU) 2016/679 (GDPR) with respect to its clients' data and that of third parties affected by a matter, which requires it to ensure that any AI system it uses respects lawfulness of processing, data minimization, and purpose limitation. Using consumer versions of AI tools, whose terms of service may permit retraining on submitted data, poses significant confidentiality risks — consistent with the CGAE's Circular 3/2026, cited above, which warns that a breach of professional secrecy can arise not only from the outcome but from the risk assumed in the process itself.

11.4. The U.S. regulatory landscape: state-level fragmentation

The United States lacks a general federal framework for artificial intelligence; regulation instead proceeds through fragmented state initiatives and the application of pre-existing ethics rules. Colorado passed Senate Bill 26-189 on May 14, 2026, repealing and replacing the 2024 Colorado Artificial Intelligence Act (Senate Bill 24-205) with a regime focused more on transparency and disclosure than on a substantive duty of care against algorithmic discrimination, set to take effect January 1, 2027. California, as noted in Section 10.5, is still considering SB 574, not yet enacted as of this report's closing date. Courts, meanwhile, have begun sanctioning AI misuse without needing a specific statute, relying on existing disciplinary authority and inherent judicial power, as the Fletcher case cited in Section 10.3 illustrates.

11.5. Entry barriers and the risk of technological concentration

Shapiro does not use the language of competition law, but his warning about a firm's method migrating into the vendor's product (1) points to a market dynamic that could raise entry barriers and entrench dominant positions at the level of foundation models and deployment platforms. Concern over concentration in AI infrastructure markets has drawn growing attention from competition authorities on both sides of the Atlantic throughout 2026, though a detailed assessment of those proceedings falls outside the scope of this report and should, in any event, be verified against primary sources before being cited in an academic context.

11.6. Converging ethical trends

Beyond general AI and competition frameworks, an international convergence is emerging around a shared principle: any significant legal decision must remain the responsibility of a legal professional; AI may assist with analysis, research, or preparation, but it cannot substitute for human judgment or accountability. This principle, which Shapiro implicitly endorses by positioning judgment as the profession's scarce resource (1), underpins both the CGAE's Circular 3/2026 and pending U.S. legislative initiatives.

11.7. Partial conclusion: regulation as accelerator or brake on absorption

Regulatory clarity can give legal certainty to firms investing in their own platforms and workflows; uncertainty and jurisdictional fragmentation, by contrast, can discourage investment. The recent extension of deadlines under the European AI Act illustrates how the regulatory framework itself can shift while a firm is mid-execution on its absorption strategy — which calls for continuous regulatory monitoring, not a static reading of the Regulation as of its original entry into force.

12. Looking ahead: scenarios and strategic recommendations

12.1. Baseline scenario: gradual absorption and consolidation among leaders

The most likely scenario is gradual but accelerating absorption, with the advantage consolidating among early movers. That advantage will not be technological — access to comparable models will tend to equalize — but organizational: the ability to encode professional judgment into reliable automated procedures.

12.2. Disruptive scenario: new AI-native operators

Private-equity-backed operators like Norm Law could capture a significant share of the market in repetitive, standardized work segments — especially in private equity — if large firms fail to complete their own transformation in time.

12.3. Stagnation scenario: the "AI theater" loop

The most pessimistic scenario is one where firms entrench themselves in committees, policies, and pilots without any real change to workflows, while pressure from clients and competitors intensifies and competitive advantage erodes gradually but persistently.

12.4. Recommendations for law firms

Investing in technology is necessary but not sufficient: the priority must be redesigning workflows, not merely acquiring tools. Training should be reoriented toward instructing and supervising AI; firms should retain control over their method, accepting third-party dependence only for infrastructure; and change management should be treated as a strategic priority, not a public-relations exercise, subject to the test Shapiro proposes: which workflow has changed, how much faster has it become, and what has improved for the client.

12.5. Recommendations for in-house legal departments

In-house departments should use AI to cut costs without sacrificing quality, set clear criteria for evaluating outside counsel's effective use of AI, and train their own teams to take on production work that was previously outsourced.

12.6. Recommendations for regulators and bar associations

Regulators should offer clear guidance on the standard of care applicable to AI use; bar associations should integrate AI competence into continuing education; and competition authorities should monitor concentration and practices that could entrench technological dependence — bearing in mind, as the 2026 European experience shows, that regulatory calendars themselves can shift quickly.

12.7. The window of opportunity

The time to act is now, while transformation remains a choice rather than an emergency: firms that begin while revenues are hitting records will be able to rebuild with room to maneuver; those that wait will do so under pressure, with fewer resources and greater urgency.

13. General conclusions

13.1. Recap of the central thesis

Zack Shapiro's essay offers a lucid analysis of generative AI's impact on the legal sector. His thesis — two clocks advancing at diverging speeds — locates the real bottleneck not in technological capability but in institutional capacity to absorb it.

13.2. The inevitability of change and the timing of the choice

The historical electrification analogy and the Coca-Cola-versus-General-Electric parable illustrate that general-purpose technologies only unlock their potential once institutions redesign their workflows around them. The question is not whether large firms will transform, but whether they will do so by design or by emergency.

13.3. Production and judgment as the key to the new business model

AI makes production cheaper and shifts value toward judgment: the decision about which risk matters, which concession is acceptable, and which argument to make. Firms that manage to decouple billing from production hours and price judgment instead will emerge from this transformation more profitable, not less.

13.4. Absorption as the new strategic imperative

Absorption — moving capability from the fast clock to the slow clock without breaking the institution — requires organizational redesign at every level of practice: workflows, quality standards, verification procedures, junior training, and pricing models. That knowledge must remain under the firm's control, not the technology vendor's.

13.5. The lawyer's role in the new economics of intelligence

The transformation does not threaten the profession's existence; it redefines its core: production is delegated to the machine, judgment concentrates in the human. Junior training and hiring criteria will need to adapt accordingly.

13.6. Professional responsibility in the age of AI

The Sullivan & Cromwell case and recent U.S. case law — including Fletcher v. Experian Information Solutions — confirm that prestige does not prevent error; the verification process does. The CGAE's Circular 3/2026 and pending U.S. legislative initiatives move in the same direction: human oversight and the signing lawyer's responsibility are non-negotiable.

13.7. The first-mover opportunity

Whoever understands first what the new technological capability makes possible, and rebuilds their organization around that understanding, will secure a leadership position that is hard to erode. Concentrated judgment will remain the legal profession's most valuable asset.

13.8. Final reflection

Law is not destined to disappear, but to rediscover its essential core: judgment, decision-making, and accountability. The question is not whether AI will transform the legal sector, but who the transformers will be, and who will be transformed. The clock keeps advancing — and now, no less significantly, so does the regulatory landscape that frames that transformation.

Bibliography

Regulations

(A) Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence, OJ L, 12 July 2024.

(B) Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data (GDPR).

(C) Colorado, Senate Bill 26-189, signed by Governor Jared Polis on May 14, 2026, repealing and replacing Senate Bill 24-205 (Colorado Artificial Intelligence Act).

(D) California, Senate Bill 574 (Umberg), passed by the state Senate on January 29, 2026, and pending in the Assembly as of this report's closing date.

Case law

(E) Fletcher v. Experian Information Solutions, Inc., 168 F.4th 231 (5th Cir. 2026).

(F) Ibach v. Stewart, 2026 WL 1110659 (Ala. 2026).

Doctrinal, institutional, and press sources

(1) Shapiro, Zack. "The Two Clocks." Post on X, July 8, 2026. Available at: https://x.com/zackbshapiro/status/2074852265354022983.

(2) David, Paul A. "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox." American Economic Review, vol. 80, no. 2, 1990, pp. 355-361.

(3) Devine, Warren D., Jr. "From Shafts to Wires: Historical Perspective on Electrification." Journal of Economic History, vol. 43, no. 2, 1983, pp. 347-372.

(4) Anthropic Institute. "When AI Builds Itself." Marina Favaro and Jack Clark, June 4, 2026.

(5) Amodei, Dario. Interview with Jim VandeHei and Mike Allen. Axios, May 28, 2025.

(6) Thompson, Clive. "Coding After Coders: The End of Computer Programming as We Know It." The New York Times Magazine, March 2026.

(7) Dietderich, Andrew (Sullivan & Cromwell LLP). Letter to Judge Martin Glenn, U.S. Bankruptcy Court for the Southern District of New York, April 18, 2026. Re: In re Prince Group, Chapter 15 Proceedings.

(8) Financial Times / Bloomberg Law. "Kirkland & Ellis to Spend $500 Million on AI Platform." May 2026.

(9) The American Lawyer. "Kirkland & Ellis 2025 Results: $10.56 Billion Revenue, $11.1 Million Profits Per Partner." March 2026.

(10) Bloomberg Law / Law.com. "Blackstone Legal Fees to Kirkland Fall to $87.8 Million in 2025." February 27, 2026.

(11) Bloomberg Law. "Norm Law Launches AI-Native Legal Platform with Blackstone and Bain Capital Backing; Michael Schmidtberger Named Chairman." January 22, 2026.

(12) Kirkland & Ellis and Palantir Technologies. Joint press release on the launch of the fund-formation platform, June 4, 2026.

(13) OpenAI / Anthropic. Announcements of AI-native deployment services firms: OpenAI Deployment Company (May 11, 2026) and Anthropic Applied AI (May 4, 2026).

(14) MIT. "State of AI in Business 2025." 2025.

(15) Karp, Alex. Interview on CNBC, July 1, 2026.

(16) Pendergrast, Mark. For God, Country and Coca-Cola: The Definitive History of the Great American Soft Drink and the Company That Makes It. Basic Books, 1993.

(17) Council of the European Union. "Artificial Intelligence: Council and Parliament Agree to Simplify and Streamline Rules." Press release, May 7, 2026.

(18) General Council of Spanish Lawyers (Consejo General de la Abogacía Española). Interpretive Circular 3/2026 of the Code of Professional Conduct of the Spanish Bar, on the drafting, signing, and delivery of legal filings assisted by generative AI, approved by the Plenary on April 10, 2026, with amendments of April 13, 2026.