One Out, One Model In: Taking AI Substitution Out of Section 139

In February 2026 Block — the payments company once called Square, founded by Jack Dorsey — told roughly four thousand people, nearly half its workforce, that their roles no longer existed. Its chief executive was open about his reasoning: AI had made the roles unnecessary, he said, and “within the next year, the majority of companies will reach the same conclusion”. Whether he is right is a question for the economist and the economic historian. If he is right, the question for an employment lawyer is how much society should charge him.

A UK employment lawyer reads the same announcement and is already inside section 139: the need for employees to carry out work of a particular kind has just dropped by four thousand, and had Block been a UK employer, dismissal put to a tribunal as redundancy — a potentially fair reason under s.98(2)(c) — would have been the natural first move, though at that scale one requiring collective consultation and a notification to the Secretary of State of the redundancies. How fit is UK (and European) employment law for dealing with such mass redundancies?

The trap, as a game

A paper published this spring by Brett Hemenway Falk of the University of Pennsylvania and Gerry Tsoukalas of Boston University gives the Dorsey manoeuvre a formal skeleton. The AI Layoff Trap (arXiv:2603.20617) models competing firms, each of which keeps the full cost saving of replacing a worker with AI but bears only its own fraction — one divided by the number of firms — of the consumer demand that the lost wage destroys. Every firm can see the cliff edge approaching; no one can afford to hit the brakes.

In the limiting case the model reduces to the prisoner’s dilemma, proved rather than asserted (their Corollary 1). The bimatrix below is that case with whole-number nominal payoffs: +1 for the deviator, −2 for the firm left holding the wage bill, −1 when both automate. Payoffs are per-firm profits relative to mutual restraint.

Why they cannot agree their way out

The paper is equally unsentimental about the exits. Voluntary restraint fails not because of bad faith but because automation is a strictly dominant strategy — the move that pays best whatever the other firm does: a non-binding pact is, in the authors’ words, cheap talk, and no private bargain among firms is self-enforcing even before contracts are drafted. Their Corollary 3 is sharper still — even partial fixes will not arise on their own, because voluntary profit-sharing (giving workers a stake in the AI gains) is itself a dominated strategy, one that always loses to a better strategic play. Supervisory pieties about responsible adoption are not merely inadequate. In game terms they are inadmissible.

What does correct the distortion, in their taxonomy, is exactly one family of instrument, built on a premise that predates AI and economics alike: governments tax the behaviour they want to discourage. Universal basic income does not discourage automation, because it pays the same whether a firm keeps its people or replaces them; a tax on company profits does not discourage it either, because it scales every firm’s gains equally and leaves the choice between people and machines unchanged. The only instrument that works is a tax on automation itself — charge each firm, for every task it automates, a fee matching the demand that its layoffs destroy for everyone else. Hold on to that last idea, because the rest of this article is about giving it a legal name.

What section 139 does with it

Section 139 is where English law sets the price. Notice what the words omit: the work itself need not have stopped. It is enough that the employer’s requirement for employees to carry it out has ceased or diminished, and that the dismissal is attributable to that. Under Murray v Foyle Meats Ltd [1999] UKHL 31 the test is functional, and Safeway Stores plc v Burrell [1997] ICR 523 taught tribunals to run it in three stages.

Run the one-for-one substitution through it. The role persists; the requirement for employees to perform it does not. Swap two employees for a model licence and one supervisor and the requirement for employees has, by definition, diminished. The dismissal is redundancy on the orthodox reading. A fair redundancy then carries an algorithmically fixed cost — the statutory redundancy payment, set by the formula for age and years of service and capped — and that figure is a discount next to what an unfair dismissal would expose the employer to paying.

The three statutory discounts

The first is procedural: redundancy fairness polices process rather than purpose, the Williams v Compair Maxam Ltd [1982] IRLR 83 checklist of warning, consultation, objective criteria and alternative work. A board-level automation contract collapses all four into ceremony.

The second is remedial. Under Lund v St Edmund’s School Canterbury UKEAT/0514/12/KN a genuine redundancy dismissal sits outside the ACAS Code, so the 25 per cent s.207A uplift is unavailable however perfunctory the consultation was.

The third is compensatory. Polkey v AE Dayton Services [1987] IRLR 503 reduces to near nothing the award where dismissal would have happened regardless, and a vendor contract signed in February answers that question by September. The competent one-for-one swap costs a capped redundancy payment and a week of meetings.

Pricing such exits is ordinary work for counsel, mine included; the arithmetic flatters the client.

The elegant fix

The amendment the trap demands is definitional, and short enough to fit in a box:

The definition does the policing. The line is drawn more cleanly by Article 3(1) of the EU AI Act:

The formulation is doing precise work. A harvester moves matter; it does not infer from inputs how to generate outputs, so it stays outside, and agricultural and manufacturing mechanisation keeps its redundancy route. An LLM that infers content from prompts sits squarely inside, as do the classical machine-learning systems that produce “predictions” or “recommendations” — so the definition neither date-stamps the provision to current architectures nor leaks into every tool with a script in it.

One status note on the borrowing: the Regulation is not part of UK law. It postdates IP completion day, so it was never retained EU law under the European Union (Withdrawal) Act 2018, and by August 2026 the UK has still declined to enact a horizontal counterpart, preferring the sectoral regulator-led course. The borrow imports nothing but the drafting: Parliament would be free to lift the definition directly into the 1996 Act, and nothing domestic stands in its way.

The repricing follows from the definition, mechanically. The substitution dismissal loses the redundancy label and must be defended as some other substantial reason, tested for reasonableness section by section under s.98(4). The ACAS Code reattaches and the s.207A uplift returns. Polkey survives but has to be properly proved: an employer must now persuade the tribunal not that the meetings were polite but that dismissing this employee, now, was within the range of reasonable responses, with the genuine alternatives honestly examined. The tribunal’s question changes from whether the exit was ceremonially correct to whether it was defensible. That is a materially harder question.

One doubt remains: will the employer simply reframe? Bar the redundancy label and a dismisser can still call the same substitution a reorganisation, raise it as some other substantial reason, and win under the reasonableness test precisely because the saving is so large. The amendment closes that second door with a single sentence: the same enabling subsection can provide that a dismissal attributable to such substitution is not capable of being justified as some other substantial reason for the purposes of s.98(1)(b). Cut off from both redundancy and SOSR, the substitution must be run through ordinary unfair dismissal, or left to the more ambitious instruments — an automatic-unfairness ground, or a fuller special regime on the model of TUPE. The minimal change is the definition alone; the robust one adds the SOSR bar. The article argues for the robust version, and sets out the wider menu because the widest is where the argument ends.

Rerouting to unfair dismissal is not therefore a price too low to bite. The objection is that compensation tracks mitigated loss: under Norton Tool Co Ltd v Tewson [1973] 1 WLR 45 the compensatory award covers the dismissed worker’s actual and prospective loss, so an analyst who is in a new role by the next quarter costs the employer little beyond the basic award. The objection assumes the market absorbs displaced labour, and that is exactly the assumption the trap denies. Where demand for the displaced is genuinely drying up, there is no comparable role to move into; the reasonable-mitigation search fails for want of workplace openings, and the award tracks a loss that is real and continuing. Compensation is not uniformly weak: it is light exactly where the trap is absent, and heavy exactly where it bites. A remedy scaled to the reality of the labour market is the right remedy, and the tribunal reaches it by ordinary evidentiary means.

And on the paper’s own test, this works the same lever. An automation tax charges the price automatically, on every task, from the day it becomes law. A redefinition of s.139 charges it only when a tribunal finds that one particular substitution was not really redundancy, and was not justified either. Both raise the price of firing; the tax collects in advance, the statute collects case by case.

A permission route instead

The redefinition has one weakness: it is still litigation. A dismissed analyst with no appetite for a tribunal, or a time limit already spent, lets the price go uncharged, and the taxpayer funds the tribunal that collects it for everyone else. There is a cheaper, earlier place to take the price, and it is not ours to claim. In the Netherlands, redundancy dismissals run through the UWV, which must grant permission on the ’a’ and ‘b’ grounds of Art. 7:669 of the Civil Code before the employer may terminate, and the employer who gets through pays a transition compensation capped at €94,000. The dismissal is the prize at the end of the process, not the start: see the comparison of the two regimes.

A UK parliament prepared to go further could borrow the shape for the hardest cases: a special category of AI redundancy that requires the employer to apply for leave to dismiss, with a surcharge levied when permission is granted. It would collect the price up front, administratively, whether or not the worker ever sued, and it would spare the tribunal the cost of collecting it case by case. The Dutch system has its own evasions — the ‘disturbed relations’ route can become a supervised way to an outcome already expected — and a permission regime for AI alone would be new machinery where none now stands. That is why this article sticks, for now, to the definitional first move: it works with the law as it stands and costs nothing to run. The permission route is the more ambitious second act, for a parliament with more time and courage than a minor definition requires.

Why not make it automatically unfair?

The rival lever — a new automatically unfair dismissal for “replacement by automated means” — tempts, and should be declined. Automatic grounds are binary: proof of the reason ends the employer’s case outright, with no reasonableness range to moderate it, and every automatic ground generates a treasury of motive litigation because the reason must be proved or imputed before it bites. Substitution is provable the ordinary way, through disclosure of the vendor contract, the reorganisation chart and the headcount plan. What it lacks is not proof but price. Automatic prohibition also invites avoidance by relabelling: the dismissal “for performance” of the analyst who would not be automated on time. Rewriting the s.139 definition achieves the same deterrence through a less distorting mechanism, preserves the reasonableness filter for substitutions that genuinely are justified — and does so by deletion rather than accretion.

Timing flirts with the proposal. From 1 January 2027 section 25 of the Employment Rights Act 2025 cuts the qualifying period for ordinary unfair dismissal from two years to six months. Re-routing substitution dismissals from the redundancy shelf to s.98 territory dramatically enlarges the class of employees who can bring the harder claim. A s.139 amendment would land just as the floodgates for AI redundancies open up.

None of that requires mass collapse to be under way, and the 2026 UK data, read carefully, shows both things at once. PwC’s AI Jobs Barometer records the most AI-exposed organisations growing headcount by 52 per cent against a 2018 baseline, evidence that the aggregate collapse the doomsayers promised has not yet arrived.^[1] Yet Morgan Stanley, looking at firms that have used AI for a year or more, found net job losses of 8 per cent over twelve months, double the international average.^[2] The pattern is not contradiction but concentration.

Where firms merely adopt generic tools, headcount holds; where they substitute at depth, the losses are real, and they cluster at entry level. That is precisely the population this amendment prices.

The trap needs no villain. It needs only that the cheapest lawful exit continues to run through section 139.

Table of Authorities