How does the leverage model actually make money?
By billing associates’ time at rates well above what associates cost, across enough associates per partner that the spread becomes the firm’s profit. Marc Galanter and Thomas Palay set out the canonical account in Tournament of Lawyers: The Transformation of the Big Law Firm (University of Chicago Press, 1991), which tied the growth of the hundred largest firms to a blend of experienced partners and junior lawyers driven by “the race to win ‘the promotion-to-partner tournament’” — and argued that the same tournament that drove the growth could undo the model. Whether that thesis still describes 2026 is a separate question the book cannot answer.
The billing structure that converts that into revenue has not moved. The 2026 Report on the State of the US Legal Market, published in January 2026 by the Thomson Reuters Institute with the Georgetown Law Center on Ethics and the Legal Profession and drawn from the reported results of 184 US-based firms, cites Thomson Reuters Legal Tracker data showing 90% of all legal dollars still flowing through standard hourly rate arrangements.
One number in that report matters more than any forecast about AI, because it describes the base of the pyramid as it already is. The report puts associate realization rates “the lowest on average at 85.6%” and says associates are “producing work that’s already being written off at a significant pace.” It does not label that figure a collection realization rate — the only collection realization figure it labels as such is 90.3% across all segments — and it does not list the other timekeeper levels, so read 85.6% as the report’s own average for associates and nothing more.
Has AI actually reduced law firm headcount?
No. Measured across the firms in the Thomson Reuters and Georgetown data, lawyer full-time-equivalent headcount grew 2.9% in 2025, and the report describes 2025 as the third year of historically strong hiring, with most growth at the firms seeing the strongest demand. Since January 2023 the average Midsize and Am Law Second Hundred firm has grown total headcount by more than 8%, while the average Am Law 100 firm held to what the report calls “a more reserved 5%.”
The report’s own reading of this is worth quoting for its directness: “Whereas other industries may be touting AI-induced layoffs to promote efficiency, the legal industry has chosen the opposite course: If AI augmentation makes their lawyers better and more efficient, then that only makes manpower more valuable, not less.” Spending followed. Direct spending on lawyer compensation rose 8.2% against 2024, with per-lawyer spending on associates up 3.8%, alongside growth of 9.7% in spending on technology and 10.5% on knowledge management tools.
It also offers a mechanism for why automation has not yet displaced juniors, and it turns on that realization figure. Because associate work is already written off at a significant pace, AI can absorb the inefficient portions without touching collected revenue — firms automate what was not getting paid for and keep associates on work that is. Whether that buffer holds once the automation reaches billable work is exactly the open question.
Then what is the evidence that entry-level work is shrinking?
Three separate datasets pointing the same direction at the top of the market, none of which attributes the movement to AI. The National Association for Law Placement’s “Jobs & JDs: Employment for the Class of 2025 — Selected Findings,” published 5 August 2026, reports that firms of more than 500 lawyers hired approximately 540 fewer graduates than the year before, a decline of 7.5%, bringing that segment to 6,588 jobs and marking the first decrease there since the Class of 2014. The same document records a 92.8% employment rate for the class overall, the second-highest the National Association for Law Placement has measured and behind only the Class of 2024 at 93.4%, against a market that “contracted by more than 2,700 jobs compared to the previous year.” The sharpest falls were in federal government employment, down 37%, and public interest, down 14%, but the losses were not confined there: NALP reports that job figures were down across almost every segment, with private practice losing more than 940 entry-level positions and business-sector employment falling to its lowest level in 35 years.
The pipeline into those jobs narrowed first. NALP’s 2025 research on law student recruiting found the average number of 2L summer associates per office at eight in 2025, down from nine in 2024 and ten in 2022 and 2023 — the smallest average class per office since 2020. The offer rate out of those programs held at 97%, so the constraint is the size of the class, not conversion.
The mix of hires shifted too. The NALP Foundation’s Update on Associate Attrition and Hiring (CY 25), published 21 April 2026 from 141 US and Canadian firms covering 6,335 associate hires and 4,442 departures, counted 3,296 lateral associate hires against 3,039 entry-level hires in 2025 — a reversal of 2024 that the Foundation describes as “reverting to norms for the prior years.” Separately, Artificial Lawyer reported on 20 July 2026 that SurePoint’s Law School Hiring Report 2026 put Am Law 200 entry-level hiring at 7,489 in 2022, 7,640 in 2023, 7,417 in 2024 and 7,426 in 2025 — flat, with a small decline across the four years. SurePoint’s report is behind a download form and states neither its population definition nor its method, so what those four numbers count cannot be checked; the Am Law 200 framing is Artificial Lawyer’s. Artificial Lawyer’s own caveat is that the flattening “started before widescale legal AI use.”
Does the wider evidence on AI and entry-level jobs actually measure legal work?
Barely, which is the main thing to know before borrowing it. The most-cited study is “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence” by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab, using ADP payroll records. In an update dated 12 August 2026, extending ADP payroll data through mid-2026, the authors report that employment among workers aged 22 to 25 in highly AI-exposed occupations stands about 19% below where it would be had it kept pace with similarly aged workers in less-exposed occupations, and that the adjustment runs through reduced hiring rather than increased separations.
The authors’ own caveats are load-bearing. They describe these as descriptive patterns rather than causal estimates, note that education differences explain part of the gap and that some trends predate generative AI, and flag that the ADP sample produces larger estimated effects than national survey benchmarks. The update does not break out lawyers or paralegals.
The number most often cited against that is a forecast, and should be read as one. Goldman Sachs’ 26 March 2023 report “The Potentially Large Effects of Artificial Intelligence on Economic Growth,” bylined Jan Hatzius, Joseph Briggs, Devesh Kodnani and Giovanni Pierdomenico, put US legal occupations at 44% exposure, second only to administrative work at 46%. The report describes that number in its text as a share of current work tasks that could be automated and labels the underlying exhibit as a share of industry employment exposed to automation, which are not the same quantity. Either way it is an exposure estimate produced in March 2023, not a measurement of anything that has happened to a law firm.
The closest thing to a measurement inside legal work is a randomized controlled trial run on students rather than associates. Daniel Schwarcz, Sam Manning, J.J. Prescott, Patrick Barry, David R. Cleveland and Beverly Rich published “AI-Powered Lawyering: AI Reasoning Models, Retrieval Augmented Generation, and the Future of Legal Practice” in the Journal of Law and Empirical Analysis on 9 April 2026. Recruitment began in September 2024 at the University of Minnesota Law School and the University of Michigan Law School; 137 upper-level students completed at least one of six tasks, each student working under all three conditions — a retrieval-augmented legal AI tool, an AI reasoning model, and no AI — rather than being sorted into three separate groups. The abstract reports that the tools “significantly boost productivity in five out of six tested legal tasks, with statistically significant gains of anywhere from 50% to 130%,” and that they perform “exceptionally well in complex tasks like drafting persuasive letters and analyzing complaints.” The five-of-six figure belongs to the retrieval-augmented tool; the reasoning model improved four of the six, and neither helped on the transactional drafting task. On hallucinations the paper is careful and the numbers are small: three hallucinations in tasks done with the retrieval-augmented tool, four with no AI at all, and eleven with the reasoning model, across 768 tasks, which the authors present as descriptive rather than as a basis for statistical inference.
What it does not establish is the thing a managing partner would want to know. Law students are not first-year associates, six assignments are not a matter, and quality was scored by graders rather than tested against a client’s outcome. Set it alongside the Institute’s account of how often legal AI still gets it wrong and what a controlled evaluation found about general-purpose tools, where results were strong on summarizing and poor on precision work.
What breaks in training if juniors stop producing first drafts?
The mechanism by which a junior lawyer learns to recognize a bad draft, which is producing bad ones and having them corrected. This is the part of the subject with no dataset behind it, and it should be labeled that way: the claim that supervision capability is built by production is an argument from how the apprenticeship has worked, not a measured finding, and nobody has run the counterfactual.
The argument is specific enough to be worth stating precisely. Reviewing AI output requires knowing what the right answer looks like and where this kind of work usually goes wrong. A lawyer who has drafted forty disclosure schedules has that; a lawyer who has reviewed forty machine-drafted schedules without ever having built one may have something different and thinner. If firms remove the reps, they are relying on the judgment transferring some other way, and no firm has yet published evidence that it does.
The counter-argument is equally real. Supervising output at volume is itself a skill, arguably a more valuable one, and first drafts were never the only way juniors learned. The honest position as of September 2026 is that this is unresolved. The Institute’s talent pyramid area covers what leverage looks like when the bottom is the part that automates, and training and competence covers what firms are doing about it.
Which moves first, the billing model or the headcount?
The stronger case is that billing moves first, because the conflict is already visible in the numbers and the headcount is not. The Thomson Reuters and Georgetown report frames the tension plainly: firms are deploying technology that can accomplish in minutes what once took hours, then trying to bill for it by the hour, and it illustrates the impasse with the sticker shock a client would feel at a $2,000 hourly bill from an associate even where the work would previously have taken ten hours. Client interviews in the report describe legal departments asking firms to propose billing structures that incorporate AI efficiencies while firms complain that clients convert every proposal back to hourly rates.
The case for headcount moving first rests on precedent rather than on present data. The same report recalls that after 2007 firms shed partners and associates, corporate legal departments absorbed that talent, billing scrutiny went from cursory to forensic, realization rates fell and rate growth flatlined for nearly a decade — and warns that generative AI could trigger a similar revolution if a downturn arrives while clients are already squeezing budgets. That is a forecast, offered as one.
What is measured, as of September 2026, is a narrow set of facts: hourly billing at 90% of legal dollars, headcount growing 2.9% in 2025, entry-level hiring at the largest firms down 7.5% for the Class of 2025, and 2L summer classes at their smallest per office since 2020. Anyone presenting a resolution of those into a single trend is forecasting. The Institute’s writing on the billable hour and its business model area cover where that pressure lands first.