AI and the Legal Career Path: What's Changing, What's Not, and What to Do

This article explores how generative AI is reshaping the legal career path — not simply by changing legal tasks, but by challenging how professional expertise is developed. Through the lens of legal work, apprenticeship, and learning science, it examines what happens when AI begins to absorb the entry-level work that traditionally trained future experts.

The Changing Apprenticeship: How LLMs Are Rewriting the Legal Career Path(When AI Reads Every Case: The Quiet Remaking of the Legal Profession)

It took three moments for an industry to notice the ground moving beneath it. A junior lawyer at 2 a.m. A trading floor in February. An apology letter in April. None of the three made the news as a story about the legal profession. Together, they are the story.

This is not a collapse. Courtrooms still fill, contracts still close, clients still pay. What is happening is quieter and more consequential: a migration of value — away from the scarcity of information, toward the scarcity of judgment. This piece asks three questions. Which layer of legal work is being eroded? Where is the value moving? And for the people still standing on this career path, how do you rebuild your skills when the ladder you were supposed to climb is being dismantled from the bottom?

Three Nights That Changed Everything

Century City, Los Angeles, 2021. Two in the morning.

Winston Weinberg is still at his desk at O'Melveny & Myers. Less than two years out of law school, he's a junior associate in the firm's securities and antitrust group, which means his days are spent inside a framework drawn by partners: research memos, redlined contract clauses, citation checks — and an endless supply of 2 a.m.s. His roommate, Gabriel Pereyra, a research scientist at Meta, has a habit of saying you need to see this at dinner. One night, the thing is GPT-3.

What actually changes the trajectory is a case Weinberg doesn't understand at all: a landlord-tenant dispute. The standard path — pull the statutes, run the case law, call someone senior — would take him two or three days. He takes a different one. He scrapes a hundred real tenant questions from Reddit's legal forums, uses structured reasoning prompts to generate responses, then hands the outputs to three practicing lawyers for blind review. Nobody knows which answers came from the machine.

A large majority of the responses were rated highly useful by practicing lawyers under blind review.

In 2022, Weinberg leaves the firm and founds Harvey with Pereyra. By March 2026, Harvey is valued at $11 billion, used by more than half of the AmLaw 100, with Paul Weiss, PwC, and KKR on the client list (per reporting by Fortune, TechCrunch, and Observer, March 2026).

Tuesday, February 3, 2026. European markets open.

Traders in London and Amsterdam watch the same names fall through the floor: RELX, parent of Westlaw, down 14%. Thomson Reuters, parent of LexisNexis, down 15%. Wolters Kluwer down 13%. LegalZoom, down nearly 20%. Across software and financial-data names, some $285 billion in market value disappears in a single session. Financial media gives the day a name: the SaaSpocalypse (as reported by Reuters and Bloomberg at the time).

The fuse had been lit the Friday before. On January 30, Anthropic shipped a legal plugin for Claude Cowork — autonomous contract review, compliance tracking, legal briefing. Every feature a direct hit on the core product that Westlaw and LexisNexis have sold for decades: annual licenses, priced per seat. A Morgan Stanley note from analyst Toni Kaplan turned the market's unease into panic.

New York, April 2026.

Sullivan & Cromwell — founded 1879, partner rates north of $2,000 an hour, outside counsel to OpenAI — files an open letter of apology with the court. In a closely watched bankruptcy matter, a filing from the firm contains AI-fabricated case citations. The cases do not exist. The reporter numbers resolve to nothing. The letter concedes the filing violated the firm's own internal AI use policy (per the Financial Times; also covered by Sina Finance, April 22, 2026).

Three scenes, three positions. One man found the future at 2 a.m. One market repriced the old model in a single trading day. One 147-year-old institution exposed the crack in an apology letter.

How Generative AI Is Actually Changing Legal Work: A Four-Layer Map

"Will AI replace lawyers?" is the wrong question, because a lawyer's work is not one thing. It is four layers stacked on top of each other, and each layer is absorbing a very different amount of shock:

The adoption numbers say this is not prophecy; it is current practice. The Thomson Reuters Institute's 2026 Report on AI in Professional Services finds that 40% of legal organizations have formally deployed AI tools, 82% of professionals use them at least weekly, and 87% expect AI to become core to their business processes within five years. Agentic AI — systems that execute multi-step tasks autonomously — sits at just 15% deployment today, but 53% of organizations have it in planning.

Why does the information-processing layer fall first? That requires one fact about how large language models actually work. At the bottom, an LLM is a next-token predictor: given everything before it, it estimates the most probable next unit of text. Two consequences follow. First, retrieval, comparison, and synthesis are tasks where language models can provide significant assistance, especially when combined with external knowledge sources and verification workflows. Second, hallucination is not merely an occasional bug; it is a fundamental limitation of generative models that requires verification mechanisms. When the model doesn't know the answer, it doesn't stop. It keeps generating the statistically most plausible next tokens. A case that never existed is, to the model, just a very convincing string of tokens.

That is the root of the April apology letter — and the technical premise for everything that follows: the risk AI poses in law is not that it is sometimes wrong. It is that when it is wrong, it looks exactly like when it is right.

Will AI Replace Junior Lawyers? What the Hiring Data Shows

The Stanford Digital Economy Lab's study Canaries in the Coal Mine? (Brynjolfsson et al., built on ADP payroll records covering millions of U.S. workers) finds that in occupations with high AI exposure, employment among workers aged 22–25 has declined relative to less AI-exposed occupations, while employment among senior workers in the same occupations has remained comparatively stable. The title borrows the miners' warning system: entry-level jobs are the canaries. They fall first because what they do is closest to what AI can already do.

Law is among the most AI-exposed professions there is. Over the past year, Reuters, The American Lawyer, and other industry outlets have documented a consistent pattern: big firms shrinking first-year associate classes, deferring start dates, and — in some cases — effectively freezing campus hiring. The numbers vary; the direction does not.

On the surface this is an efficiency story: if AI does in minutes what a junior associate does in three days, why keep hiring juniors? One layer down, what's being hollowed out is the century-old economics of the big law firm — the pyramid leverage model. A partner sits atop a column of junior lawyers, and the spread between the juniors' billable output and the partner's rate is the firm's profit engine. When AI absorbs the base of the pyramid, the fulcrum of the lever comes out.

And there is a third layer the industry has barely begun to discuss. A partner's judgment was built the hard way — through thousands of "low-value" contracts and memos drafted as a junior. When the training ground disappears, where do the partners of 2036 come from? This is a delayed fuse. It will not appear in any quarterly earnings report. But it determines what this profession looks like a decade from now — which is where we pick the thread back up in the final section.

Billable Hours in the Age of AI: Why the Pricing Model Is Cracking

Return to that Tuesday in February. The RELX and Thomson Reuters rout was, on the surface, a story about two stocks. In substance, it was the market repricing an entire business model: charging for the scarcity of information no longer works. For fifty years, Westlaw and LexisNexis were not selling software. They were selling I can find what you can't. When AI performs research and first-pass analysis of comparable quality in seconds, that moat is being filled in.

For law firms, the shock arrives in a more insidious form — what we'll call the efficiency penalty: AI saves lawyers time → fewer billable hours → less revenue, while the firm's spending on AI tooling and internal systems climbs. The better you deploy the technology, the less you earn. This paradox is putting new pressure on the billable hour, the industry's dominant pricing model for decades.

But the cracks admit light. Fixed fees, value-based billing, subscription legal services — alternatives that have been advocated for years without traction suddenly have a forcing function. And AI is helping in a counterintuitive way: precisely because AI can estimate standard work hours and data costs for a given matter type with precision, firms can quote a fixed price that is competitive without being a loss-leader. The anchor of pricing power is shifting from how much time did you spend to what is your judgment worth.

That shift means radically different things depending on where you stand — which is the subject of the next section.

Where Legal Value Is Moving: Judgment, Accountability, and the New Compliance Frontier

The most revealing detail in the S&C apology letter is not that a top-tier firm erred. It is what the letter inadvertently proves: the signature — and the responsibility behind it — is something AI cannot take. Courts require a human to sign and stand behind every filing. Clients pay $2,000 an hour for the privilege of someone is accountable when this goes wrong. In an era of exploding AI output, underwriting that output — verifying it, signing it, bearing its errors — is converting from an obligation into a high-value function.

The second destination of migrating value is a continent currently being mapped: AI compliance. Under the EU AI Act — on the timetable adjusted by the European Commission's November 2025 "Digital Omnibus" proposal — transparency obligations, including disclosure of AI-generated content, take effect in December 2026; high-risk AI system obligations apply from December 2027; and obligations for high-risk components embedded in regulated products begin in August 2028. Penalties run up to €35 million or 7% of global annual turnover, whichever is higher, and the Act has extraterritorial reach: if your product serves the EU market, you are in scope. Within the next few years, every multinational will need to build or scale an AI governance and compliance function — and people who speak both law and AI systems are the scarcest species in that market.

The third destination is the most interesting, because it puts lawyers on the other side of the table. Harvey employs some 200 practicing lawyers whose daily work is no longer serving clients but coaching the model — evaluating outputs, correcting legal error, injecting senior reasoning patterns into training. Thomson Reuters has committed hundreds of domain experts to post-training its legal AI. Legal AI evaluation and domain-expert roles are emerging as a new area where lawyers contribute to improving AI systems.

One demand-side datapoint is worth recording for the emotional weather: a Reuters/Ipsos poll finds 53% of American respondents worry AI will eliminate their own jobs. The anxiety isn't unique to law. But law feels it first and deepest — because the two assets this profession has historically priced, information asymmetry and text-production capability, are precisely the two things large language models do best.

Old value is depreciating; new value is being priced. So for an actual human being on this career path — a junior associate, a senior partner, an in-house counsel eyeing a pivot — where should the chips go?

Building a Legal Career Alongside AI: Skills Worth Betting On

Before mapping any path, you have to see the true shape of the problem. This is not a "which tools should I learn" question. It is a question with a cognitive-science answer.

Bastani and colleagues at the Wharton School ran a randomized controlled trial with nearly a thousand high-school students in Turkey, published in 2024 as Generative AI Can Harm Learning. Students tutored with GPT-4 performed significantly better during practice. When the AI was removed for the exam, those same students performed significantly worse than peers who had never used it. The researchers call it the crutch effect: the AI performs the cognitive work — retrieval, trial and error, getting stuck, working back out — that the brain was supposed to do. The appearance of practice remains; the growth of capability stops. A second study, published in 2025 in the journal Societies (Gerlich, DOI: 10.3390/soc15010006), identifies the same mechanism from the other direction: frequency of AI use correlates negatively with critical-thinking ability, mediated by cognitive offloading — the more thinking you outsource, the more autonomous cognitive effort you skip.

Set those findings against the picture from the hiring-data section and the shape of the problem is clear: what the legal profession is losing is not just the workload of junior positions. It is the learning mechanism itself — the system by which high-value judgment was accumulated through low-value work. The traditional apprenticeship has broken. And rebuilding it cannot be outsourced to institutions. It falls to each individual standing on the path.

Rebuilding the Apprenticeship: How Junior Lawyers Can Still Get Reps

One governing principle: use AI to train; don't let AI train for you. In daily practice, that resolves into three moves.

First: answer it yourself, then pull the AI draft, then compare. When a research question lands, force yourself through the full arc — retrieval, reasoning, drafting — however slow, however flawed. Only then call up the model's answer and diff them, point by point. Where do they diverge, and why? The comparison is the rep. The AI's output is a reference answer, never a ghostwriter.

Second: cast the AI as the adversary. There was a time when a junior who wanted to rehearse how would opposing counsel attack this clause or where would the judge interrupt my argument had to consume a partner's hours — or simply never got the chance. Now, at 2 a.m., you can have the model play opposing counsel, the presiding judge, the furious client, and run the full red-team exercise. For the first time in history, the practice opponent is no longer scarce.

Third: position yourself as the reviewer, not the author. When reviewing an AI draft, hold yourself to one hard rule: find at least three problems and write down why each is one. If you can't, you haven't actually read what it wrote.

Now look back at the first scene. Weinberg's method — a hundred real questions, blind-rated by three practicing lawyers — was, in essence, a self-built deliberate-practice loop. He wasn't outsourcing his thinking to GPT-3. He was using it to multiply, by a factor of a hundred, the volume of training a junior associate could access. What let him found Harvey was not that he could use GPT-3. It was that he trained with GPT-3 until he had a judgment nobody else had: this technology can survive blind review by practicing lawyers.

The Senior Lawyer's Leverage: Judgment, Risk Allocation, and Pricing Authority

For senior practitioners, the question is not "how do I avoid replacement" but "where does the fulcrum of my leverage move to." Four areas where human responsibility remains difficult to automate.

Signing and standing behind. The S&C apology was signed by a managing partner, not by a model. Putting your name on a document and staking your license and professional reputation on it — that act cannot be outsourced, and the market pays a premium for it. As the volume of AI output requiring human underwriting grows, the premium grows with it.

Pricing risk. AI can enumerate ten risks in a clause. Whether a given risk is likely, whether its downside is survivable, whether it's worth torpedoing the deal over — that is a trade-off made on a client's behalf, and trade-offs require someone to bear consequences.

Translating across rooms. Rendering legal language into commercial language and commercial demands back into legal frames; brokering between the board, the regulator, and the engineering team. This is organizational coordination work, and it runs on trust and context — not text generation.

Defining the problem. Turning a client's vague we want into that market into an answerable legal question is the first step of all legal work — and the step LLMs are weakest at, for a structural reason: a model can only answer the question it is asked. Defining the question requires a three-dimensional read on the business, the regulator, and the client's actual situation.

New Roles on the Other Side of the Table: AI Governance, Compliance, and Legal AI Product Work

Three career exits are emerging. They matter less as job listings than for what they share — their common features trace the direction of value migration.

Corporate AI governance and compliance functions. The December 2027 and August 2028 high-risk tiers of the EU AI Act are an explicit hiring driver: every company placing AI systems on the EU market needs people who can read the Act and read model documentation, to build the governance framework. The role did not exist three years ago. It is now in multinational org charts.

Legal AI product and evaluation roles. Harvey's ~200 lawyer-coaches and Thomson Reuters' hundreds of post-training domain experts prove that "lawyers training AI" has professionalized. The work includes model evaluations, setting domain data-labeling standards, and legal engineering — decomposing firm workflows into units a model can learn from.

Process design for agentic AI. The Thomson Reuters report's 15%-deployed / 53%-planning gap for agentic AI says the bottleneck is not model capability. It is process reconstruction: someone must design the new workflow — agent drafts, junior verifies, partner signs — and define the accountability boundary at each handoff. That is the oldest skill in the legal toolkit, designing procedure and allocating responsibility. The objects have simply changed from humans to human-machine hybrids.

All three exits share one property: they stand at the boundary of law and technology, and they demand the ability to evaluate AI, not merely to use it. Usage is the lowest barrier there is. Evaluation is the moat.

The story can end where it began. In March 2026, Winston Weinberg — founder of a company now worth $11 billion — sat down for a Fortune podcast and was asked for career advice for the AI era. He didn't talk about technology. He didn't talk about vision. What he said was: employees must constantly prove their own value.

The man saying it was, five years earlier, the junior associate redlining contracts at 2 a.m. He wasn't replaced by AI. He just got there one step ahead of everyone else — and turned AI into the instrument he proved his value with.

Key Takeaways

• AI is not hitting "lawyers" as a monolith. It is hitting the information-processing and text-production layers of legal work; judgment, accountability, and client trust remain firmly human.
• Stanford's payroll-data research shows employment among workers aged 22–25 in high-AI-exposure occupations has declined significantly in relative terms — entry-level legal jobs are the canary in the coal mine.
• Two pillars are cracking simultaneously: the billable hour and the pyramid leverage model. Pricing power is shifting from how much time you spent to what your judgment is worth.
• The EU AI Act's December 2027 and August 2028 high-risk milestones are actively creating corporate AI governance and compliance as a new function.
• The first rule of skill-building: use AI to train, not to train for you — a principle grounded in the empirical "crutch effect" research, not in exhortation.

Read More of Intelligenr

• AI in Healthcare Jobs: A Practical Adaptation Guide

• How to Learn with AI: A Four-Quadrant Framework Beyond Prompt Engineering

• Beyond Chatbots: How GenAI Is Rewiring Finance’s Core Engine

• How AI Is Changing White-Collar Work: Customer Service & Administrative Work

References

• Brynjolfsson, E., Chandar, B., & Chen, R. (2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab.
  https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/

• Thomson Reuters Institute. (2026). Future of Professionals Report 2026.
  https://www.thomsonreuters.com/en/institute/future-of-professionals-2026/report

• Bastani, H., Bastani, O., Sungu, A., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences.
  https://doi.org/10.1073/pnas.2422633122

Author Note: This article was developed from research into the changing economics of legal work, the evolution of professional apprenticeship, and emerging evidence on how AI affects learning. We use the legal profession as a case study to examine a broader question: when AI takes over parts of the work through which expertise was traditionally developed, how should people rethink the path from practice to judgment?