Designing AI Agents for EU Antitrust Compliance: From Public Distancing to Machine Distancing

Commentary on ‘AI and corporate responsibility’ generally concludes that businesses must understand the technology they deploy and mitigate its risk.

This article reaches the same conclusion by a doctrinal route. It examines how EU competition law may apply to AI agents that interact - agents deployed by a business that communicate with the agents or systems of competitors - rather than to autonomous algorithms that align pricing without communication.  At each stage, it considers what the case law implies for how agents are designed and controlled.

The outcome is a practical translation of EU competition law doctrine into agent controls: design measures that may bear on whether machine communications amount to participation, whether autonomous conduct was reasonably foreseeable, and whether the undertaking acted negligently.

To achieve this, the article untangles three questions which should be kept separate, not least because agentic autonomy may have different implications at each stage:

• Does the conduct satisfy the requirements of Article 101 TFEU, including the existence of an agreement or concerted practice?

• Is the agent’s conduct attributable to the undertaking?

• If an infringement is established, was it committed intentionally or negligently so as to justify the imposition of a fine?

These questions are analytically distinct, but they are not hermetically sealed. Eturas1 treats awareness as part of participation, and VM Remonts2 makes knowledge and foreseeability part of attribution. Agentic autonomy may therefore matter at more than one stage, and the same evidence (logs, perimeters, testing records) may be relevant for more than one factor.

Infringement: from machine receipt to participation in a concerted practice

In Eturas, the Court of Justice rejected an irrebuttable inference that travel agents knew of an anticompetitive message merely because it had been sent through the platform. A finding of participation required evidence capable of establishing awareness, subject to the presumption of innocence. A recipient could also rebut participation by showing that it had determined its market conduct independently or had publicly distanced itself.

Applied to the agentic AI setting, these principles may suggest that the following should not necessarily establish that the undertaking knowingly substituted coordination for competition:

• The mere fact that a competitor’s agent sent an invitation to collude, or sent competitively sensitive information

• The fact that the company’s agent received or technically processed it

• The fact that the information existed somewhere in the agent’s memory or logs.

However, if the agent is treated as part of the undertaking (which is probable - see below), an authority may argue that machine receipt and operational use constitute receipt and use by the undertaking, even if no natural person reviewed the message.

Moreover, agentic AI is likely to involve interaction between competitors and so, where agents exchange communications aimed at a coordinated outcome, it will be very difficult to deny the existence of a concerted practice.

Nevertheless, one could look to distinguish between, on the one hand, (i) mere technical receipt, without use or alteration of market conduct, or receipt plus autonomous processing, but no resulting change in commercial conduct; and, on the other hand, (ii) receipt followed by responsive pricing or output or reciprocal agent-to-agent communications designed to reach or maintain a coordinated outcome.

Realistically, the Eturas awareness argument may apply only to the first category. That still leaves the public-distancing defence, but what would public distancing mean for agentic AI? The following could form the basis of a “machine distancing protocol”:

• Rejecting the invitation in its reply to the competitor's agent

• Logging the interaction and preserving but not using the information

• Escalating to a human decision-maker; and

• Preventing the information from affecting subsequent commercial outputs or model memory

Attribution: is an AI agent analogous to an employee?

If an antitrust authority were to treat an ‘infringing agent’ as analogous to an employee carrying out the company’s business,  then the bar for avoiding liability is high.  That is because EU competition law holds undertakings personally responsible for infringements, and employees form part of the undertaking's economic unit.

The issue of corporate versus employee liability for infringements of EU competition law resurfaced very recently in Advocate General Szpunar’s 1 October 2026 Opinion in Zapp3, which relied on EU jurisprudence establishing that liability does not require proof that an authorised representative personally engaged in the conduct or expressly authorised an employee’s participation in an anticompetitive meeting.  An employee’s anticompetitive conduct is therefore attributable to the undertaking, which is generally liable for that conduct.

An authority might even argue that these additional factors make a company-deployed AI agent less independent than an employee:

• Has no separate legal personality

• Has no interests distinct from those of the company

• Acts through company-provided infrastructure, data and permissions

• Is deployed to advance the company’s commercial objectives

• Has a degree of autonomy as a feature selected or accepted by the company

From a policy perspective, treating agents differently from employees also creates a challenge: attribution would become harder as businesses delegate more functions to AI.

That said, could the agent nevertheless be compared to an employee who lacked authority to act on the undertaking’s behalf?  Might it be possible to demonstrate to the regulator that the agent exceeded the powers which the company had conferred upon it? Or perhaps this is better argued in support of the argument that the infringement was not foreseeable in a situation where the agent operated beyond the permitted perimeter of commercial operation.

Attribution: can an AI agent’s technical autonomy make it analogous to an independent service provider?

An agent is plainly not a separate undertaking under EU competition law. However, even a separate undertaking may not be genuinely independent if it has little or no autonomy and acts under the direction or control of the undertaking using its services.

An interesting question is whether the converse could apply: could an agent with substantial autonomy - and therefore not acting under an undertaking’s direction or control - be treated as an independent third party?

If so, one could argue that the company would only be liable (in line with the reasoning of the Court of Justice in VM Remonts) where it was aware of collusion and intended to contribute; or it could reasonably have foreseen that the agent would collude and was prepared to accept that risk

However, this argument appears weak for proprietary agents. Even an agent with extensive technical autonomy may lack economic independence where, for example, it:

• does not offer services on a market in pursuit of its own economic interests

• bears no commercial risk

• has no assets or revenues of its own

• cannot choose its principal; and

• cannot itself be fined or deterred

The argument is stronger for agents supplied or operated by a third party. Treatment as an independent provider is most plausible where a third party supplies and substantially operates the system; the customer has limited visibility into or control over the system’s operation; and the provider designs the mechanism that facilitates coordination.

Fault: foreseeable autonomy or unforeseeable deviation?

Even where a substantive infringement is established, could the undertaking avoid a fine because it acted neither intentionally nor negligently?4

The threshold is high. In Schenker5, even legal advice that the conduct was lawful did not shield the undertaking from a fine. The reasoning is that an undertaking’s mistaken legal characterisation of its conduct does not exempt it from a fine if it could not have been unaware of that conduct’s anticompetitive nature.

But with agentic AI, one could argue that there is an important distinction between, on the one hand, knowing the relevant facts but wrongly believing them to be lawful and, on the other hand, being genuinely unaware that the relevant conduct had occurred at all because an autonomous system departed unpredictably from its design and safeguards.

Schenker limits reliance on legal mistake. Arguably, it does not necessarily resolve whether an undertaking acted negligently where an agent engaged in materially unforeseeable conduct despite appropriately designed and implemented controls.

The argument for distinguishing that line of case law would be: the undertaking neither intended the anticompetitive conduct (or even class of conduct) nor negligently permitted it, because the conduct was not reasonably foreseeable given the agent’s defined function, technical constraints, testing, monitoring and escalation controls.

However, the evidential threshold would be high, and an authority would likely examine:

• Had comparable models exhibited emergent coordination?

• Were system outputs monitored?

• Were there hard restrictions on communications concerning price, output, customers or strategy?

• Did the company test for collusive behaviour before deployment?

• Were warnings ignored and did the company continue using the system after suspicious behaviour emerged?

Finally the greater the agent’s autonomy and commercial importance, the stronger the controls an authority might expect to see. A company may struggle to argue that conduct was unforeseeable if it intentionally deployed a system whose commercial value lay in autonomous strategic interaction.

Conclusion: from doctrine to agent controls

An undertaking is likely to struggle to avoid attribution by characterising its own AI agent as a rogue employee or independent third party. Technical autonomy is unlikely to equate to economic independence, particularly where the undertaking selected the agent, defined its objectives, supplied its data and permissions, and deployed it in pursuit of its commercial interests.

But attribution is not the end of the analysis. An authority must still establish an infringement, and the imposition of a fine requires intentional or negligent conduct. That leaves two potentially important distinctions.

First, technical receipt of a collusive communication should be distinguished from its adoption or use in determining market conduct.

Second, a mistake about the legality of known conduct should be distinguished from materially unforeseeable conduct by an autonomous system operating outside a properly designed and enforced perimeter.

Compliance by design matters because it supplies the technical and evidential architecture for those distinctions. Appropriate controls can enable an agent to reject and quarantine collusive communications, prevent them from influencing commercial decisions, preserve an auditable record, escalate the interaction and permit timely human intervention. They may also demonstrate that the undertaking identified foreseeable risks and took proportionate steps to address them.

These controls will not be a safe harbour of course. Nor will they overcome evidence that the undertaking tolerated foreseeable coordination or continued deployment after warning signs emerged. But they may help prevent an infringement, rebut an inference of participation in appropriate cases, inform the assessment of negligence, and reduce the duration of any infringement that nevertheless occurs. In an agentic environment, compliance by design is therefore not simply good governance. It is part of the undertaking’s legal defence architecture.

What an enforcer will ask, and the controls that answer it

• 1Case C‑74/14, Eturas UAB and Others v Lietuvos Respublikos konkurencijos taryba, EU:C:2016:42.
• 2Case C‑542/14, VM Remonts and Others v Konkurences padome, EU:C:2016:578.
• 3Opinion of AG Szpunar in Case C‑347/25, Zapp Precision Metals GmbH and Zapp AG v KU, 1 October 2026, paragraphs 49-50
• 4Council Regulation (EC) No 1/2003, Article 23(2)(a) (fines for infringements committed "intentionally or negligently")
• 5Schenker (Case C‑681/11, EU:C:2013:404)
• 6Musique Diffusion française (Joined Cases 100–103/80, EU:C:1983:158)