Ask a competition lawyer what has to be proved under the Chapter I prohibition and the answer comes without hesitation. There must be an agreement or concerted practice between undertakings (the legal term for businesses), which has the object or effect of preventing, restricting or distorting competition.

Now ask a different question. What do those requirements assume? What has to be true about the world for “find the agreement” to be a sensible way of catching businesses that are fixing prices? That question usually produces a pause. This article is about that pause, because the assumptions are where artificial intelligence does its real damage.

New to the series? The argument is that AI rarely breaks competition law’s rules. In a small number of cases, it dissolves the assumptions underneath them. The opening article has the map, the two previous articles covered the AI stack and the contested flywheel, and this one names the assumptions.

Two Layers in Every Legal System

Legal systems have two layers.

The visible layer is the rules: what must be pleaded, which tests apply, who has to prove what. Lawyers argue about this layer constantly.

Beneath it sits a second layer made of assumptions. These are never stated in a statute, never argued in a case, and never chosen by anyone. They were inherited from a world in which they happened to be true. Competition law rests on four of them, and each one can be drawn out of doctrine that practitioners use every day.

The First Assumption: An Agreement

The Chapter I prohibition in section 2 of the Competition Act 1998, like Article 101 of the Treaty on the Functioning of the European Union, catches agreements between undertakings, decisions by associations of undertakings, and concerted practices. The Court of Justice defined a concerted practice in Suiker Unie as a form of coordination which, without reaching the stage of an agreement, knowingly substitutes practical cooperation for the risks of competition.

Neither provision catches what economists call conscious parallelism, meaning rival firms that each independently decide not to compete hard, simply by watching one another. The Court of Justice stated in its judgment that the requirement that every undertaking determine its own commercial policy independently “does not deprive economic operators of the right to adapt themselves intelligently to the existing and anticipated conduct of their competitors”. What it forbids is direct or indirect contact with the object or effect of influencing a competitor’s conduct or revealing one’s own.

The case law holds that line firmly. In Wood Pulp the Court of Justice held that parallel conduct cannot be treated as proof of concertation unless concertation is the only plausible explanation for it. In the United Kingdom, the Court of Appeal in Argos Ltd v Office of Fair Trading [2006] EWCA Civ 1318 upheld findings of concerted practices conducted through a common supplier. It rested them on a narrower proposition than the Competition Appeal Tribunal had stated: the retailer disclosing its pricing intentions could be taken to intend them to be passed on, the competitor receiving them could be taken to know how they had been obtained, and it used them in setting its own prices. The Court doubted, without deciding, whether reasonable foreseeability of the onward transmission would be enough on its own. The United States arrives at the same place by a different route, at two different stages of a case: Theatre Enterprises holds, at trial, that parallel behaviour, without more, does not prove an agreement, and Twombly requires a claimant, at the pleading stage, to allege facts plausibly suggesting one.

Why does the line sit there? Partly for a reason of principle, which is the one the Court of Justice stated in Suiker Unie and Donald Turner argued for in the United States (Turner, “The Definition of Agreement Under the Sherman Act: Conscious Parallelism and Refusals to Deal”, 75 Harv. L. Rev. 655, 672 (1962), later cited to the same effect by the Supreme Court in Twombly). A firm that reads its market and adapts intelligently to its rivals has done nothing blameworthy.

But there is a second reason, quieter and rarely stated. Coordinating without communicating is genuinely hard. Watching rivals continuously is costly. Punishing a firm that breaks ranks, precisely enough to discipline it and then stopping, is beyond most companies. Every one of those difficulties grows worse as the number of firms rises. Because coordination without contact almost never worked for long, tolerating it cost almost nothing.

The assumption: durable collusion requires an agreement or a knowing concertation, so hunting for the contact catches the collusion.

The Second Assumption: An Author

Competition law attributes conduct to somebody who chose it. The formal requirements differ across the three systems, and the differences are instructive.

In cartel law the requirement is explicit. The concerted practice in Suiker Unie must be entered into knowingly, and in Eturas the Court of Justice presumed participation only where an undertaking was aware of the collusive initiative and failed to distance itself from it. Awareness is the hinge.

In the law of abuse the position is different, and it is worth stating precisely, because it is often misstated. Under Article 102 and the Chapter II prohibition, abuse is an objective concept. The Court of Justice held in Hoffmann-La Roche that it does not depend on the dominant undertaking’s intention. A firm can abuse its position without meaning to, and proof of intent is not an element of the offence.

The United States runs the other way, and makes intent explicit for completed monopolisation. Grinnell speaks of the “wilful acquisition or maintenance” of monopoly power, as distinguished from growth as a consequence of a superior product, business acumen, or historic accident. Attempting to monopolise sits on a separate, stricter footing: the Supreme Court held in Spectrum Sports v McQuillan that a plaintiff must prove specific intent to monopolise, together with a dangerous probability of achieving monopoly power.

So intent is not what the three systems have in common. What they share is something weaker and more fundamental. The conduct must be attributable to an undertaking that chose the course it took. Objective abuse still requires conduct that can be attributed to an undertaking and adopted as a commercial strategy, rather than arising from nowhere. The concerted practice still requires undertakings that knowingly substituted cooperation for the risks of competition. Even where intent drops out of the elements, authorship does not.

That is why the evidentiary culture looks the way it does. Investigators look for the strategy memo, the incautious email, the document that reveals what somebody meant to do. Even in a system where intent is not an element, those documents establish that a course was chosen and by whom. Intel and the factual record in Google Shopping, including an internal Google email on the purpose of the display changes at issue, both drew on internal material of that kind.

The assumption: conduct is attributable to an undertaking that chose it.

There is a useful precedent buried in that formulation. Abuse law has already dispensed with the stronger requirement, which is intent, while keeping the weaker one, which is attribution. The law has shown it can shed one and retain the other. The question the later articles raise is what happens when even attribution fails, and no undertaking chose the outcome at all.

The Third Assumption: A Consumer

The consumer-welfare standard is the test that asks whether conduct makes consumers worse off, through higher prices, lower quality or less choice. Article 102 itself speaks of abuse “to the prejudice of consumers”, and the CMA’s prioritisation principles put consumer harm at the centre of enforcement. But the standard does not merely measure harm to consumers. It quietly relies on them to do a job.

Consider how markets are defined in the first place. The CMA and the European Commission both use the hypothetical monopolist test, sometimes called the SSNIP test, which asks whether enough customers would switch away if a hypothetical monopolist raised prices by a small but significant amount. The switching customer is not simply a figure in the analysis. The consumer who compares, defects and walks away is the discipline that keeps sellers honest during the long stretches when no enforcer is watching.

That assumption is no longer hypothetical to test. Since late 2025, payment networks and leading AI companies have built the infrastructure for agents to make purchases on a person’s behalf, including without a human being present at the moment of purchase. The effect on discipline could run either way. An agent that compares every offer tirelessly would be a more demanding customer than any person has ever been. An agent that answers to someone else would be no customer at all. Which of those emerges is the subject of later articles. What matters here is that the figure the hypothetical monopolist test has always pictured, a person deciding whether to switch, is ceasing to be the one who decides.

Competition law has met a version of this before. The European Commission’s 2024 Market Definition Notice itself observes that demand is sometimes shaped not by the final consumer but by someone whose interests are not necessarily aligned with the consumer’s. It gives the example of the doctor who chooses a medicine the patient will take and an insurer will pay for. But the prescribing doctor comes with a professional duty to the patient. The agent, so far, comes with none.

The strain will not fall evenly. The largest buyers have an exit that no individual has: they can run their own models. In September 2026 it was reported that Latham & Watkins had bought its own Nvidia servers to fine-tune open-weight models in-house. A buyer with that option disciplines its suppliers whether or not it uses it. A consumer, or a small business, has no equivalent. So the demand side is not ceasing to police the supply side everywhere at once. It is ceasing to do so for the buyers least able to replace it.

The same Market Definition Notice supplies a way of seeing this. It recognises that where suppliers offer different groups of customers different terms for the same product, for reasons unrelated to cost, a group can form a narrower, separate market. That is typically so where three conditions are met together: the supplier can tell, at the moment of sale, which group a customer belongs to; resale between customers is unlikely; and the difference in treatment lasts. Pricing systems that know exactly who they are dealing with, and agents that can tell them, make the first condition easy to satisfy, and personalised terms for services are often hard to resell.

The assumption: the demand side is human, acts for itself, and polices the supply side.

The Fourth Assumption: Time

No statute says that enforcement must arrive before a harm becomes permanent, because nothing ever needed to say it.

Competition law is overwhelmingly retrospective. An authority investigates, proves and then remedies, and that architecture assumes the market will still be capable of correction when the decision finally lands. The interim measures available under section 35 of the Competition Act and Article 8 of Regulation 1/2003 are faster, and the Commission’s WhatsApp order in June 2026 shows they can bite within months. But they too are triggered by a proceeding, which presupposes that the conduct has already been detected. They are quicker afterwards, not before.

The assumption has held historically because monopoly erodes. High profits attract entry, new technology routes around incumbents, and enforcement is only one corrective force among several.

The assumption: harms are reversible on enforcement timescales.

What the Four Have in Common

An agreement, an author, a consumer, and time.

Notice what they share. Not one of them is an element that anybody proves. Each is instead the reason the elements work. Each is a piece of the world that makes hunting for agreements a sensible way to catch collusion, and acting after the event a sensible way to repair harm.

Why name them now? Technology has pressed on them before. In the 1990s the US Department of Justice alleged that airlines used a shared fare-publishing system to signal and coordinate fares without ever meeting, and resolved the case by consent decrees. But that was a strain at the edge, reachable because a shared system sat in the middle. Artificial intelligence is the first technology to strain the assumptions themselves, and it does so selectively.

Pricing systems have been shown, in the laboratory, to coordinate with no agreement and no author, and the field evidence is consistent with that without yet proving it. Software agents now buy on consumers’ behalf, and may not act for the consumer at all. Frontier dynamics could, in the worst case, outrun correction entirely, although that possibility is so far modelled rather than observed.

A Test You Can Actually Use

Later articles take each of these in detail. This one offers something smaller and more immediately useful: a test to apply to any AI competition claim that crosses your desk.

Does the claim merely intensify conduct the law already reaches, through more scale, more speed, more precision or more opacity? Or does it dissolve one of the four assumptions?

If it is the former, it is old wine. It may be real and it may be urgent, but it is a problem of capacity, meaning cases, resources and proof.

If it is the latter, it is genuinely novel, and no increase in budget will fix it, because the failure sits in the foundation rather than in the funding.

And ask one further question of any claim that passes the test: for whom? The assumptions do not fail uniformly. A harm that dissolves an assumption for consumers may leave it intact for a buyer large enough to build its own alternative, and the right response depends on which of them the harm reaches.

The Uncomfortable Part

The principles built on these assumptions were chosen and defended: Suiker Unie explains why intelligent adaptation to rivals is lawful. The assumptions underneath them never were. Nobody weighed them against alternatives or adopted them after argument. They were inherited from a world that made them true.

A legal system defends its rules constantly. Its assumptions it rarely notices at all, and the first signs that it is beginning to do so are revealing. In October 2025 California amended its antitrust statute to target common pricing algorithms and lowered the bar for pleading a conspiracy, and New York barred landlords from using pricing software to set rents. Both respond to real harms. But look at what they changed. California’s central prohibition still requires a contract, combination or conspiracy, and aims at algorithms fed with pooled competitor data, the ground the existing law already reaches. It made the agreement easier to allege. Its one provision that dispenses with an agreement, aimed at coercing others to follow an algorithm’s recommendations, still requires somebody to do the coercing. New York’s law went further, and has now been halted. On 29 September 2026 a federal court enjoined it, finding RealPage marginally likely to show that it restricted protected commercial speech, in part because it targeted software without distinguishing between public and non-public data, which is the line the existing law already draws. Neither asks what happens when there is no agreement and no one to attribute the coordination to. The system has started to patch its rules. It has not yet revisited its assumptions.

That is why, when an assumption fails, nothing breaks loudly. The system keeps running, confidently, on foundations that are no longer there. The cases still get pleaded. They simply stop describing the world.

Food for thought. For in-house counsel: which of the four assumptions does your product strain first, and does your compliance programme know it? For practitioners: which element do you prove in every case without noticing you are proving it, and what is your move the day it is not there? For policymakers: what breaks in your framework if the demand side stops being human?

Next in this series: the good news, most AI competition harms are old wine, the tools mostly work, and that boring conclusion is the strongest card an enforcer holds.

Join the conversation. If you think the four assumptions are the wrong four (too many, too few, mis-named), this is the article in which to say so, because everything that follows in this series stands on them.