In December 2025 The Spectator published a piece by Sean Thomas entitled "AI will kill all the lawyers." It featured an anonymous senior barrister — "James" — who had fed a complex civil appeal to an AI model and watched it produce, in thirty seconds, a document he judged superior to the one that had taken him a day and a half. His conclusion: "With rare exceptions, law is finished for almost everyone, maybe even the judges." He advised his niece not to study law. "Do not get into a lifetime of debt for a job that won't exist in ten years. Or less."
It is a vivid piece of journalism. It captures a real anxiety that runs through the profession. And it is, on the economics, substantially wrong — or at least, wrong about the practitioners it most wants to terrify.
The Error in the Argument
The anonymous barrister's experiment tested AI on a standalone written opinion: a complex civil appeal, stripped of identifying details, fed as a prompt. That is a task which consists almost entirely of what economists call codifiable cognition — legal research, analysis of authority, structured argumentation, prose drafting. It is the kind of task at which large language models already excel. Nobody serious disputes this.
But the barrister then made a leap. Because AI could outperform him on that task, he concluded it would replace the job. That leap — from task-level capability to job-level displacement — is precisely the error that a new paper from the London School of Economics now provides a formal framework for understanding.
Garicano's Bundling Framework
In March 2026 Luis Garicano (LSE), Jin Li and Yanhui Wu (University of Hong Kong) published "Weak Bundle, Strong Bundle: How AI Redraws Job Boundaries." The paper's central insight is deceptively simple: labour markets price jobs, not tasks.
Most analysis of AI and employment starts from "task exposure" — if AI can do a task, the worker who does that task loses. Garicano and colleagues argue this is incomplete. Jobs bundle multiple tasks together. What matters is not just whether AI can perform a particular task, but how costly it is to split that task away from the rest of the job. That cost — the coordination cost of unbundling — determines whether AI replaces the worker or merely makes them more productive.
The framework distinguishes two types of occupation:
Adapted from Garicano, Li & Wu (2026). In a weak bundle, AI peels off Task 1 and the human role contracts. In a strong bundle, AI improves Task 1 performance but the human retains the whole job.
Weak-bundle occupations are those where the codifiable task (Task 1) can be separated from the contextual task (Task 2) at low cost. When AI becomes good enough at Task 1, it makes economic sense to split the job: the machine handles Task 1 autonomously, the human does only Task 2. The human's revenue share narrows. If demand is inelastic, the flood of extra output depresses prices and some workers are displaced entirely. This is the standard "AI replaces jobs" story — and it is real, but it applies only to occupations where the tasks are loosely coupled.
Strong-bundle occupations are those where splitting the tasks destroys significant value. Shared context, accountability, cross-task spillovers, liability, and communication frictions all raise the coordination cost. In a strong bundle, AI still improves Task 1 performance — but it does so inside the existing job. The human keeps both tasks and retains a larger share of gross revenue. Bundling becomes a force that protects labour.
What Makes a Strong Bundle? Context, Liability and Cross-Task Spillovers
Garicano and colleagues identify three main drivers of high coordination cost — that is, three reasons why splitting the job destroys value:
Shared context. When the person who performs Task 1 is the same person who performs Task 2, information flows freely between the tasks. Split them, and the Task 2 performer loses context. The radiologist who also speaks to the referring physician has clinical context that a pure image-classifier lacks. The employment barrister who drafts their own skeleton has absorbed everything from the conference, the documents, the witness dynamics — and that understanding shapes every submission.
Liability. A single professional who is liable for the full service has skin in the game across both tasks. Separating the codifiable task from the contextual one creates a verification problem: who checks the machine's output? Who takes responsibility when it hallucinates a citation — as it certainly will? In the legal context, the answer under professional conduct rules is clear: the practitioner on the record. That accountability binds the tasks together.
Cross-task spillovers. What the worker learns doing Task 1 improves how they do Task 2, and vice versa. The solicitor who has drafted the schedule of loss understands the claim's pressure points in mediation. The barrister who has done their own legal research spots the cross-examination question that the skeleton argument couldn't anticipate. Outsource Task 1 to a machine and these spillovers vanish.
Employment Law as a Paradigm Strong Bundle
Consider the Task 1 elements of employment litigation: legal research, drafting pleadings, citation-checking, document review, calculating basic and compensatory awards, computing ACAS uplift, preparing chronologies and cast lists. Every one of these is codifiable. AI can already do most of them passably and will soon do them well.
Now consider the Task 2 elements: conducting an initial client conference with a distressed claimant who has just been dismissed. Reading the employment judge's body language during a preliminary hearing. Deciding whether to press a point in cross-examination or leave it. Negotiating with a respondent's solicitor at a judicial mediation. Managing lay client expectations — and lay client emotion — through a multi-day final hearing. Exercising professional judgment on whether to accept an offer or fight on. Advising on the tactical implications of a deposit order. Building a relationship with instructing solicitors who will send the next brief.
These tasks are not merely different in kind. They are deeply intertwined. The barrister who has personally read every document in the bundle brings that knowledge into the hearing room. The solicitor who has drafted the witness statements understands which factual narrative will survive cross-examination. The practitioner who has personally worked through the case law knows which authorities to deploy on the hoof when the judge raises an unexpected point.
This is what a strong bundle looks like. The coordination cost of splitting these tasks is enormous. You cannot farm out the legal research to a machine and then hand the contextual output to a different human — or even to the same human who hasn't engaged with the research — without destroying significant value.
What the Model Actually Predicts
Garicano's model produces four results that matter for practitioners:
First, in bundled production the human keeps a larger share of gross revenue than in unbundled production. This is intuitive: a bundled worker contributes to both tasks, so the market pays them for both. An unbundled worker who performs only Task 2 earns only the Task 2 share.
Second, better AI pushes the frontier towards unbundling — but only where the coordination cost is low enough. In strong-bundle occupations, the job stays bundled even as AI capability increases substantially.
Third, the strong bundle mitigates the capacity shock that destroys weak-bundle jobs. In a weak bundle, when AI takes over Task 1, the surviving workers suddenly devote all their time to Task 2, flooding the market with output and depressing prices. In a strong bundle, the worker's time allocation barely changes — they simply produce better Task 1 output with AI assistance, without the sudden release of capacity that drives displacement.
Fourth, even individual workers in strong-bundle occupations see their earnings protected. The price compression effect that hammers weak-bundle workers is largely absent.
Where the Spectator Piece Goes Wrong
The anonymous barrister tested AI on what was, functionally, an unbundled task: a standalone written opinion, detached from any client, any hearing, any instructing solicitor, any court. He then extrapolated to the profession as a whole. In Garicano's terms, he demonstrated that AI is excellent at autonomous Task 1 production — and concluded that this would destroy all legal jobs.
But the question is not whether AI can do the task. It is whether the task can be peeled away from the job without destroying the value of the bundle. For most contentious employment work, the answer is clearly no.
This is not to deny that the barrister's experiment was impressive. AI is already very good at drafting legal opinions. The point is that "very good at drafting opinions" does not entail "capable of replacing the professional who drafts opinions, advises the client, appears at the hearing, cross-examines witnesses, and carries professional liability for the outcome."
The Honest Caveat: Weak Bundles Are Real
None of this means that AI poses no threat to any legal work. Garicano's framework is symmetrical: it predicts that weak-bundle legal occupations face genuine displacement. These are roles where the codifiable task can be separated at low cost — where there is little shared context, limited cross-task spillover, and no binding liability constraint.
Standard-form conveyancing. Compliance checklists. Process-driven probate. Template-based employment contract drafting. Routine disclosure review. Regulatory filings. These are weak bundles. AI will take the codifiable elements, the human role will narrow, and some displacement will follow. The Spectator's anonymous barrister is not wrong that "process lawyers are obviously doomed." He is wrong to extrapolate from that to the whole profession.
The profession will bifurcate. Practitioners whose work sits in strong bundles — litigators, advocates, advisers who combine technical knowledge with judgment, persuasion, and client management — will find that AI makes them more productive without threatening their role. Practitioners whose work is primarily codifiable, delivered without significant contextual integration, will face the full force of the substitution effect.
Implications for Employment Practitioners
If Garicano's framework is right — and it is grounded in serious economics, not wishful thinking — the strategic implications for employment lawyers are relatively encouraging.
First, lean into the bundle. The practitioners who will thrive are those who combine technical excellence with deep client relationships, courtroom presence, and strategic judgment. A solicitor who both drafts the particulars of claim and manages the client through the hearing is in a strong bundle. A solicitor who only drafts documents on instruction — with no client contact, no hearing attendance, no strategic input — is in a weak one.
Second, use AI for Task 1. The model predicts that strong-bundle workers who adopt AI see productivity gains without earnings losses. Use it for research, drafting, calculation, document review. The economic logic is clear: better Task 1 output improves the quality of the bundle as a whole, and you keep the revenue share.
Third, be sceptical of the doom narrative — but not complacent. The "AI will kill all the lawyers" story confuses task exposure with job displacement. The more rigorous analysis is about bundle strength. But bundle strength is not fixed. If professional regulation loosens, if liability rules change, if clients become willing to accept unbundled services — coordination costs fall, and strong bundles weaken. The protection is structural, not eternal.
Conclusion
"With rare exceptions, law is finished for almost everyone," said the anonymous barrister. With respect, the economics do not support that. What the economics support is a more nuanced — and more interesting — conclusion: that AI will redraw the boundaries of legal work along the fault line between weak and strong bundles.
Employment litigation, with its deep interweaving of codifiable cognition and contextual judgment, sits firmly on the strong-bundle side of that line. The rumours of our death are — to borrow from a rather more reliable authority than an anonymous barrister on his second espresso martini — greatly exaggerated.
Sources: L. Garicano, J. Li & Y. Wu, "Weak Bundle, Strong Bundle: How AI Redraws Job Boundaries" (LSE/University of Hong Kong, March 2026); S. Thomas, "AI will kill all the lawyers", The Spectator, 16 December 2025.