In every chambers corridor and every City HR department, someone is having the same conversation: do we really need as many juniors? The question sounds novel. It is not. What is novel is that a team of economists has now built a formal model of the answer — and the answer is more troubling than the question suggests.
Friebel, Huang, Li, Shukla, and Zhang published “Pyramids, Diamonds, and Oscillations: AI and the Structure of Internal Labor Markets” on 10 April 2026.1 It is not a legal paper. It is a dynamic general equilibrium model of how firms allocate tasks between senior and junior workers, and what happens to that allocation when AI arrives. But its implications cut directly across the terrain that employment lawyers inhabit: redundancy, workforce restructuring, and the increasingly fraught question of who gets hired in the first place.
Schumpeter Inside the Firm
In 1942, Joseph Schumpeter described capitalism’s engine as a “perennial gale of creative destruction”: new technologies render old industries obsolete, old firms die, new ones are born.2 The concept is usually invoked at the level of markets. Horse-drawn carriage makers gave way to Ford. Kodak gave way to the smartphone. Blockbuster gave way to Netflix. The destruction is real but creative — the economy re-forms around the new technology.
What Friebel and colleagues show is that the same process operates inside the firm, in its workforce structure. When AI arrives, it does not simply replace workers. It reshapes the internal hierarchy. The pyramid of juniors and seniors that has characterised professional services for decades begins to change shape. And the transition is not a clean demolition and rebuild. It oscillates.
This distinction matters. Classical creative destruction is a one-off event: the old structure dies, the new one emerges, the market moves on. Firm-level creative destruction is a rolling process. The firm survives, but its workforce churns through cycles of hiring freezes and surges — each cycle creating winners and losers among the workers caught in its path.
The Efficiency Matrix
The paper identifies three channels through which AI can affect a firm: it can raise senior productivity, raise junior productivity, or accelerate the rate at which juniors learn and qualify for senior roles. In practice, of course, these channels overlap. An AI tool that helps a trainee solicitor draft a skeleton argument faster raises junior productivity and may accelerate learning and may free the supervising partner to take on more complex work. The three channels are analytical distinctions, not mutually exclusive real-world categories.
What matters is which channel dominates, because the long-run consequences are starkly different:
The central finding is this. If AI operates only through the productivity channels — making seniors or juniors better at their current tasks — the firm shrinks but its shape stays the same. A pyramid remains a pyramid. The long-run junior-to-senior ratio is pinned entirely by worker flow rates: how fast seniors leave, how fast juniors learn, how fast juniors depart before they get the chance to learn. Productivity changes the scale of the firm, not its structure.
But if AI accelerates the rate at which juniors learn — turning a three-year trainee into a competent practitioner in eighteen months — the shape changes permanently. Fewer juniors are needed to sustain the senior pipeline. The pyramid becomes a diamond.
What Oscillation Actually Looks Like
The transition from pyramid to diamond is not smooth. Picture a mid-sized law firm with 20 partners, 30 senior associates, and 60 trainees. AI arrives. Each trainee is now twice as productive. The firm doesn’t need 60 trainees to do the work — 35 will suffice. So it freezes the next recruitment round.
Two years later, the problem surfaces. Those 25 trainees who were never hired are also the 25 senior associates the firm will never have. The senior associate pool is thinning. Partners are overworked. The firm panics and over-hires — taking on 50 trainees in a single round. Three years after that, those 50 trainees qualify into a glut of senior associates. The firm freezes hiring again.
This is not a hypothetical. It is the oscillation the model predicts, and the paper’s empirical section confirms it is already visible in the data.
Drawing on 618,471 job postings from 55 tech and platform firms and over a million from finance, the paper shows exactly this pattern emerging after late 2022. The share of firms posting zero junior vacancies has risen sharply. The junior-to-senior postings ratio has fallen. The data are descriptive, not causal. But they are consistent with the model’s predictions in a way that is difficult to dismiss.
The Lost Cohort
The most uncomfortable implication is what the authors call the “lost cohort.” Juniors who enter the labour market during a hiring freeze face sharply reduced opportunities to work and accumulate skills. They do not simply wait in a queue. They miss the pipeline entirely. The cohort that follows, entering during the corrective hiring surge, benefits from the resulting senior shortage. The inequality between adjacent cohorts is not a transitional artefact. Under the rapid-learning extension of the model, it becomes permanent.
I say this as someone whose profession is structured precisely as the model describes. The Bar, law firms, accountancy, consulting — all are pyramids that depend on a continuous flow of talent upward. Barristers are self-employed, so the model applies to chambers as an organisational form rather than through the employment relationship directly. But the dynamic is identical: pupils become juniors, juniors become senior juniors, senior juniors take silk. Freeze the intake and the pipeline starves.
Pupillage numbers in some sets have already been quietly frozen or reduced. Training contract offers at City firms have plateaued. The justification is invariably framed as “efficiency.” The model suggests the long-term consequence may be a shortage of experienced practitioners that no amount of later recruitment can fix, because the firm-specific human capital — the knowledge of how a practice operates, the relationships with clients and colleagues, the instinct for which arguments work before which judges — cannot be acquired from outside.
From Model to Tribunal: What This Means for Employed Knowledge Professionals
The paper models firms in the abstract, but its natural habitat is the knowledge economy: law firms, accountancy practices, consultancies, engineering firms, financial services. These are the sectors built on hierarchical internal labour markets where juniors learn by doing, seniors supervise, and the pipeline from one to the other is the engine of the business. Crucially, the people in these structures are employees. When AI disrupts the pyramid, employment law applies.
But not every consequence of the model sounds in employment law equally. Where AI simply makes existing workers more productive and the firm stops recruiting, there is no dismissal and no claim. The “lost cohort” the authors describe — juniors locked out during a hiring freeze — is a labour market problem, not a tribunal one. You cannot be made redundant from a job you were never offered.
The position changes, however, the moment AI efficiency gains translate into a reduced requirement for existing staff. If a team of six junior associates can now do the work that previously required ten, the firm’s requirements for employees to carry out work of that kind have diminished. That is a textbook redundancy situation under section 139(1)(b) ERA 1996 — regardless of whether anyone uses the word “freeze.” The trigger is not the technology itself but the organisational decision to reduce headcount in response to it. And the model predicts that this decision will be widespread: firms facing an AI productivity shock will shed junior roles, not merely decline to fill them.
Under section 139(1)(b) of the Employment Rights Act 1996, a redundancy situation exists where the requirements of the business for employees to carry out work of a particular kind have ceased or diminished, or are expected to. Murray v Foyle Meats Ltd [1999] UKHL 31 confirmed the broad, functional approach: the tribunal asks simply whether those requirements have diminished. An AI-driven restructuring that eliminates junior roles will, in most cases, satisfy this definition without difficulty. But the ease of establishing the statutory definition should not obscure the harder questions that follow.
The first is whether the redundancy is genuine or a pretext. Safeway Stores plc v Burrell [1997] ICR 523 set out the three-stage test: has the employee been dismissed? If so, has the reason for the dismissal been established? If the reason is redundancy, was the dismissal fair or unfair? Where a firm eliminates an entire tier of its workforce in response to AI, the second stage will usually be straightforward. But the third stage — fairness — is where the Friebel model becomes relevant.
Williams v Compair Maxam Ltd [1982] IRLR 83 established that a reasonable employer will warn and consult, adopt objective selection criteria, consider alternative employment, and allow union representation. The AI context adds a layer of complexity. If the firm is eliminating all junior positions, there is no pool from which to select and no alternative junior role to offer. Consultation risks becoming performative. A tribunal might reasonably ask whether the employer considered the pipeline consequences — the very oscillation the model predicts — before deciding to eliminate the base of its own hierarchy.
The second question concerns the Polkey v AE Dayton Services [1987] IRLR 503 reduction. If the dismissal is procedurally unfair but the tribunal is satisfied that the employee would have been dismissed in any event, compensation is reduced accordingly. Here, the model offers something unusual: a formal framework for assessing the counterfactual. If the firm would have needed to re-hire juniors within two years to address the predicted senior shortage, the Polkey reduction may be significantly less than 100%. The oscillation is, in a sense, built into the mathematics.
Beyond Redundancy: Reorganisation and SOSR
Not every AI-driven restructuring will be framed as a redundancy. Some firms will reorganise roles rather than eliminate them, merging junior and mid-level functions into a single tier. In such cases, the employer may rely on “some other substantial reason” under section 98(1)(b) ERA 1996. Hollister v National Farmers’ Union [1979] ICR 542 established that a sound business reason can justify dismissal for refusing new terms following a reorganisation.
The difficulty is that the “sound business reason” test requires genuine business logic, not merely a fashionable rationale. A firm that restructures because AI is the flavour of the quarter, without modelling the workforce consequences, may struggle to show that the reason was substantial. The Friebel paper provides exactly the kind of analytical framework a tribunal might expect a reasonable employer to have considered — or at least to have grappled with in principle.
The Wage Structure Problem
The model’s wage predictions are also worth noting. When AI raises senior productivity, the wage gap between seniors and juniors widens: the senior position becomes more valuable, and firms exploit this by offering lower junior wages today in exchange for the higher future prize. When AI raises junior productivity, the gap narrows. When AI accelerates learning, both wages fall but the ratio is compressed.
For practitioners advising on equal pay and pay equity, this matters. If AI-driven restructuring systematically depresses entry-level wages while inflating senior pay, the distributional consequences will not be evenly felt. Junior cohorts in AI-exposed sectors are disproportionately younger, more diverse, and more likely to be women or ethnic minorities entering professions where the senior ranks remain demographically narrow. The model does not address discrimination directly. But the structural wage effects it predicts are the kind of thing that, in a different paper, would be called indirect discrimination by another name.
Firm-Specific Human Capital and the External Hire Problem
The paper’s extension on firm-specific human capital is, for legal services, the most directly applicable section. When internally promoted seniors possess knowledge that external hires lack — knowledge of workflows, protocols, client relationships — the firm captures a rent on every internal senior it retains. In equilibrium, the firm promotes exclusively from within and hires no external seniors at all.
This has a sharp implication for the talent pipeline. When the hiring freeze disrupts the flow of juniors, the future supply of high-productivity internal seniors falls. The firm’s value declines. And this erosion is not visible immediately — it manifests years later, when the cohort that was never hired fails to materialise as the experienced practitioners the firm now needs.
Any managing partner who has watched a practice group struggle after a period of low trainee intake will recognise this dynamic instantly. The model merely proves what experience already teaches: you cannot outsource the development of institutional knowledge.
After the Diamond: Is the New Shape Sustainable?
The paper models the diamond as a new steady state. But steady states in economics are equilibrium constructs, not predictions about durability. A diamond — a bulging middle tier with a narrow base of juniors — raises an obvious question that the model does not fully confront: where do tomorrow’s seniors come from?
If the firm needs fewer juniors because AI accelerates learning, it can maintain its senior pipeline with a smaller intake. That is sustainable, at least in principle. But if the firm needs fewer juniors because AI has replaced the tasks through which juniors learn — the document review, the first-pass research, the witness statement drafts — then the pipeline does not merely shrink. It degrades. The juniors who remain have fewer opportunities to develop the judgment and expertise that qualifies them for senior roles.
A diamond that rests on a genuine acceleration of learning is stable. A diamond that rests on the elimination of learning opportunities is not. It is a shape that looks efficient today but hollows out at the top within a decade, as the narrow cohort of juniors fails to produce enough qualified seniors to replace those who retire. The honest assessment is that we do not yet know which kind of diamond AI will create — and most firms adopting these tools are not asking the question.
Practical Takeaways
- For employers considering AI-driven restructuring: the Friebel model suggests that eliminating junior roles captures immediate efficiency gains but may create oscillating workforce instability and erode the talent pipeline. A reasonable employer should model the long-term workforce consequences, not merely the short-term headcount saving.
- For employees facing redundancy: where an entire junior tier is eliminated, the fairness of the process under Williams v Compair Maxam may turn on whether the employer genuinely considered alternatives — including partial reductions, retraining, or phased restructuring — rather than a blanket cut.
- For tribunals assessing the Polkey reduction: the model provides a formal basis for arguing that re-hiring would have been necessary within a defined period, limiting the appropriate deduction.
- For the profession: the legal sector is itself a pyramid-structured internal labour market. The pressure to reduce pupillage and training contract numbers in response to AI tools is real. The model predicts that the firms which resist that pressure will, in the long run, hold the stronger position — provided the learning channel dominates the productivity channel.
The framework can see the pipeline clearly enough. Whether it can persuade the people holding the budget is another matter entirely.
Table of Authorities
| Case | Citation | Point |
|---|---|---|
| Murray v Foyle Meats Ltd KB → | [1999] UKHL 31 | Broad functional definition of redundancy |
| Safeway Stores plc v Burrell KB → | [1997] ICR 523 | Three-stage redundancy test |
| Williams v Compair Maxam Ltd KB → | [1982] IRLR 83 | Fair redundancy procedure: warning, consultation, criteria, alternative employment |
| Polkey v AE Dayton Services KB → | [1987] IRLR 503 | Reduction for chance dismissal would have occurred in any event |
| Hollister v National Farmers’ Union KB → | [1979] ICR 542 | Sound business reason can justify SOSR dismissal on reorganisation |
Notes
- Friebel, G., Huang, Y., Li, J., Shukla, S. & Zhang, A., “Pyramids, Diamonds, and Oscillations: AI and the Structure of Internal Labor Markets” (10 April 2026). Working paper, Goethe University Frankfurt / University of Hong Kong / Harvard Business School.
- Schumpeter, J.A., Capitalism, Socialism and Democracy (New York: Harper & Brothers, 1942), ch. VII (“The Process of Creative Destruction”), pp. 81–86. Schumpeter argued that capitalism’s essential fact is not price competition between existing firms but structural change driven by new products, new methods, and new forms of organisation — a “perennial gale” that incessantly revolutionises the economic structure from within.