When Prices Think for Themselves: Rethinking Coordination in UK Competition Law

I. Introduction

In recent years, the Competition and Markets Authority (CMA) has confronted pricing behaviour in digital markets that cannot be attributed to any purposive human decision, most notably in situations where autonomous pricing tools adjust prices across competing firms in rapid succession without communication or awareness on the part of the actors involved.¹ These systems operate by continuously analysing market conditions and competitor behaviour, generating outputs that no firm selects or revises at the moment they occur. When such outputs produce aligned pricing, the CMA is left with behaviour that mirrors the economic effects of coordination while lacking the intentional and reciprocal qualities that the Chapter I prohibition in the Competition Act 1998 presupposes.²

The doctrinal framework of UK competition law was constructed for markets in which harmful outcomes can be traced to human conduct that expresses a shared understanding or cooperative engagement; it was not designed for environments in which autonomous systems generate outcomes that remain outside the conceptual boundaries of agreement, concerted practice, or restriction by object.³ The CMA’s 2021 Algorithms report acknowledged that certain pricing tools “learn to coordinate” without human direction, and its 2024 AI Strategic Update emphasised that autonomous systems adjust outputs through processes that “cannot be fully explained” even by developers.⁴ These developments reveal a deeper structural difficulty: the legal categories that define the Chapter I prohibition rely on forms of human agency that autonomous systems do not exhibit, leaving the law unable to classify behaviour that undermines competition without expressing the predicates on which liability depends.

This article argues that UK competition law is structurally ill-equipped to address aligned pricing generated by autonomous systems and requires reform to establish a coherent and workable approach.

II. Doctrinal Boundaries Within the Chapter I Prohibition

A. Agreement

Agreement under the Chapter I prohibition has always been understood as something rooted in human intention. Courts look for signs that undertakings reached a shared position, whether through explicit communication or conduct that clearly reflects a common plan.⁵ In Ping Europe Ltd v Competition and Markets Authority, the Court of Appeal had no difficulty finding an agreement because Ping had intentionally built an online sales ban into its dealer contracts and enforced it.⁶ The restriction made sense within the traditional framework because it came from a deliberate human decision.

This familiar picture becomes harder to apply when pricing is generated by autonomous systems. Reinforcement learning tools adjust prices automatically in response to data, and the resulting alignment can look very similar to collusion. The missing input is the human intention that the doctrine expects. The CMA’s 2021 Algorithms report openly acknowledges that some pricing tools “learn to coordinate” without any human direction.⁷ If the alignment comes from a system responding to market signals rather than from undertakings forming a shared intention, the legal test for agreement becomes difficult to satisfy. The doctrine expects intention; the systems do not provide it. As a result, agreement becomes hard to establish not because the outcome is ambiguous, but because the mental element the law relies on is absent.

B. Concerted Practice

Concerted practice was developed to capture situations where firms fall short of agreement but still act with a common understanding.⁸ UK courts have applied this idea by reconstructing how undertakings communicated and reacted to each other. In Argos Ltd and Littlewoods Ltd v OFT, the Tribunal examined complaints, pressure and subsequent behaviour to decide whether retailers had reached a common view on price increases.⁹ In JJB Sports plc v OFT, the Tribunal pieced together communications between JJB, Umbro and Sports Soccer to determine whether they had coordinated on replica football kit prices.¹⁰ In both cases, the analysis depended on being able to interpret human exchanges as evidence of a shared understanding.

Algorithmic pricing tools do not generate this kind of evidence. A pricing system may track competitors’ prices and adjust its own outputs accordingly. If several firms use similar tools, the resulting pattern may look coordinated, yet there may be no complaints, no pressure and no explicit messages. The systems respond to data, not to each other’s intentions. The concerted practice doctrine expects behaviour that can be interpreted through human reasoning; algorithmic pricing produces behaviour that cannot be explained in those terms. The doctrine can describe the pattern, but it cannot explain it in the way the test requires. This leaves a gap: behaviour that looks coordinated may fall outside the doctrine simply because it does not produce the kind of evidence the law is built to recognise.

C. Restriction by Object

The distinction between restrictions “by object” and “by effect” also becomes unstable when pricing is generated algorithmically. Object restrictions are those whose harmful nature is clear enough that regulators do not need to examine market impact. In Ping, the online sales ban qualified because its purpose was evident from the policy itself.¹¹ The classification depended on intentionality: Ping deliberately restricted online sales.

Autonomous pricing tools complicate this logic. When a pricing algorithm produces supracompetitive outcomes, the harmful effect may be obvious, but the purpose is not. The CMA’s 2024 AI Strategic Update notes that certain systems operate as “black boxes,” generating outputs that even developers cannot fully explain.¹² If the harmful outcome arises from emergent behaviour rather than deliberate design, it cannot be characterised as an object restriction. Effects analysis is not simple either. It requires reconstructing causal chains that may not be accessible, especially when the system adapts over time in ways the deployer cannot predict.

The current framework does not specify how to treat behaviour that is harmful in outcome but opaque in purpose. If the object category is limited to conduct with a clear anticompetitive aim, algorithmic pricing outcomes that are functionally similar to object restrictions may fall outside it. If the category is expanded to include certain algorithmic configurations, the basis for distinguishing object and effect becomes less clear. Either way, the doctrine struggles to classify behaviour that does not express a purpose in the way human-generated restrictions do.

III. The Challenges Posed by Algorithmic Pricing

A. Intention and the Boundary Between Parallel Conduct and Coordination

Competition law has traditionally distinguished lawful parallel conduct from unlawful coordination by looking at what firms meant to do. When undertakings independently react to market conditions and end up behaving similarly, the law treats this as parallel conduct. When they reach a shared understanding and adjust their behaviour because of it, the conduct becomes coordination.¹³ This distinction sits at the centre of both Article 101 TFEU and the Chapter I prohibition, and it works reasonably well in markets where pricing decisions can be traced back to identifiable actors.

Algorithmic pricing makes this line harder to draw. A firm may only intend to use a pricing tool to optimise its prices, yet the tool may track competitors and adjust outputs in ways that soften competition. The CMA’s work on algorithms shows that some systems can learn patterns that produce stable, aligned pricing without any agreement between firms.¹⁴ The firms intend to use the tool, not to collude, but the outcome can resemble collusion closely enough to raise concerns.

One way of dealing with this would be to treat the decision to adopt a particular pricing system as the relevant act of intention. On that view, a firm that chooses a tool capable of producing collusive outcomes accepts responsibility for those outcomes. This shifts the focus from individual price changes to the upstream choice to deploy the system. The difficulty is foreseeability: reinforcement learning tools adapt to data in ways that firms cannot always predict. It is not obvious when a firm can be said to intend coordination if the behaviour that produces alignment is generated by the system rather than directed by the firm.

Another approach would be to place less weight on intention and focus more directly on outcomes. For example, certain pricing patterns with sustained alignment at supracompetitive levels could be treated as presumptively problematic even if firms did not aim for them. This would move competition law towards a more outcome-based model, where the key question is whether the behaviour harms competitive processes. Such a change would be significant, and it would need safeguards to avoid penalising firms for behaviour they could not reasonably anticipate. However, it would reflect markets where pricing is driven by systems rather than human deliberation.

B. Evidence, Interpretability and the Role of Explanation

Competition law depends heavily on evidence that can be interpreted as showing coordination. In traditional cases, this includes communications, complaints, pressure and patterns of behaviour that make sense when viewed as the product of a shared understanding. In Argos and JJB Sports, courts reconstructed communications and reactions to determine whether firms had coordinated.¹⁵ In Ping Europe, the Court examined the purpose and context of the internet sales ban to classify it as a restriction by object.¹⁶ In each case, the analysis relied on being able to explain the behaviour by reference to human decisions.

Algorithmic pricing challenges this reliance on explanation. The CMA’s 2024 AI update notes that some AI-driven tools may influence market outcomes in ways that are difficult to explain even for their developers.¹⁷ A pricing algorithm may adjust prices based on complicated patterns in data that are not transparent. The behaviour may be visible—since prices move together and undercutting stops—but the process that produced it may not be. The law expects evidence that can be interpreted as showing a common understanding or purpose; algorithmic pricing produces behaviour that is observable but not easily interpretable in those terms.

This raises a critical question about what is considered evidence of coordination. If courts insist on traditional forms of evidence, including communications, complaints and explicit policies, algorithmic pricing outcomes may escape scrutiny because they do not generate that kind of material. If courts treat certain pricing patterns as evidence in themselves, they risk blurring the line between outcome and process. The challenge is to recognise algorithmic behaviour as relevant without abandoning the requirement that liability must rest on more than similarity of outcomes.

One way forward would be to treat certain system characteristics as evidence. Instead of asking whether firms communicated, courts could ask whether firms adopted systems that predictably produce aligned outcomes in particular market structures. The focus would shift from reconstructing human decisions to examining the design and deployment of systems. This approach would still require careful analysis, but it would align the evidential enquiry with the realities of algorithmic behaviour and allow competition law to treat certain system configurations as inherently risky.

C. Responsibility and the Allocation of Liability

Traditional competition law assumes that responsibility for coordination lies with the undertakings that participate in it. In Argos and JJB Sports, liability attached to firms that engage in communications and behaviour that reflects a common understanding.¹⁸ In Ping Europe, responsibility is attached to Ping for adopting a policy that restricted competition.¹⁹ In each case, the firm’s decisions provided a clear basis for liability.

Algorithmic pricing complicates this allocation of responsibility. The developer designs the algorithm, the firm deploys it, and the system then adjusts prices according to its internal logic. The behaviour that produces aligned outcomes is generated by the system rather than by any individual decision. Assigning responsibility solely to the firm that deploys the tool may seem straightforward, but it does not fully reflect the role of design and configuration. Assigning responsibility to the developer may also be problematic, particularly where the tool is used in ways the developer did not anticipate.

The Digital Markets, Competition and Consumers Act 2024 (DMCC) moves towards a more system-focused approach by giving the CMA stronger powers over firms with strategic market status, including obligations relating to transparency and fairness.²⁰ However, it does not directly address how responsibility should be allocated when algorithmic pricing produces outcomes that resemble coordination. The existing framework still assumes that liability attaches to undertakings that make decisions about how they compete. In markets determined by algorithmic systems, this assumption may be too narrow.

A more realistic approach would recognise that responsibility for algorithmic pricing outcomes is shared between design, deployment and oversight. Firms that choose to use tools capable of producing collusive outcomes could be subject to obligations such as monitoring pricing behaviour or ensuring that tools are configured to avoid certain patterns. Developers could be required to provide information about how tools operate and the risks they pose. Competition law would then treat failures to meet these obligations as relevant to liability, rather than focusing solely on traditional concepts of agreement or concerted practice.

This would not replace existing doctrines, but it would supplement them with a more accurate account of how pricing decisions are made in digital markets. It would also align competition law more closely with its underlying objective: protecting competitive processes rather than policing only intentional collusion.

IV. Comparative Perspectives

A. The European Union

The European Commission has taken a more direct interest in algorithmic pricing than the United Kingdom. Its policy brief recognises that pricing tools can influence market conditions even when firms do not intend to coordinate.²¹ The Commission has suggested that certain forms of algorithmic monitoring may fall within Article 101 TFEU where they reduce strategic uncertainty, and it has encouraged firms to assess the risks created by pricing tools before deployment. Although this does not create a new legal category, it shows a willingness to treat system behaviour as relevant to the analysis of collusion.

B. The United States

The United States has approached algorithmic pricing through antitrust enforcement rather than doctrinal reform. The Department of Justice has pursued cases involving pricing tools that were designed or used in ways that produced aligned outcomes.²² Academic commentary has argued that certain algorithmic configurations should be treated as inherently risky because they create predictable patterns of supracompetitive pricing. Stucke and Ezrachi note that reinforcement learning tools can stabilise prices in ways that resemble tacit collusion, even without communication between firms.²³ This outcome-focused perspective differs from the UK’s emphasis on intention and communication.

C. Implications for the United Kingdom

The comparative picture highlights a common theme: regulators elsewhere are beginning to treat the design and deployment of pricing systems as part of the competitive analysis. The European Union emphasises reduced strategic uncertainty; the United States emphasises system-driven outcomes. Both approaches recognise that autonomous pricing tools can influence market dynamics in ways that traditional doctrines struggle to classify. For the UK, this suggests that reform should incorporate system obligations and clearer expectations around transparency.

V. Regulatory Obligations for Algorithmic Pricing Systems

A. Obligations on Firms Deploying High-Risk Pricing Tools

The CMA’s 2021 Algorithms report acknowledges that certain systems can “learn to coordinate” without human direction.²⁴ This suggests that firms deploying high-risk pricing tools should be subject to monitoring obligations. Such obligations could require firms to maintain records explaining how pricing outputs are generated, to track whether systems produce sustained alignment, and to intervene where necessary. These measures would ensure that firms retain oversight of tools that may influence market outcomes in ways they cannot fully predict.

B. Responsibilities of Developers

Developers also play a significant role in influencing market outcomes. The CMA’s 2024 AI Strategic Update notes that some systems operate as “black boxes,” generating outputs that even developers cannot fully explain.²⁵ Developers could therefore be required to provide information about how tools operate, the risks they pose, and the configuration options available to reduce the likelihood of supracompetitive pricing. Clearer obligations on developers would help ensure that deployers and regulators have access to the information needed to assess system behaviour.

C. A Targeted Regime for High-Risk Systems

The Digital Markets, Competition and Consumers Act 2024 moves towards a more system-focused approach by giving the CMA enhanced powers over firms with strategic market status.²⁶ A similar model could be extended to pricing systems that create predictable risks of aligned outcomes. A targeted regime would allow regulators to intervene where system design or deployment creates structural conditions that undermine competition, even if traditional evidence of coordination is absent. This would ensure that competition law remains capable of addressing risks that arise from autonomous system behaviour rather than human decision-making.

VI. Conclusion

Algorithmic pricing exposes a fundamental gap in UK competition law. The Chapter I prohibition is built on assumptions about human intention and reciprocal conduct, yet autonomous pricing systems do not operate through intention, communication or awareness. They generate outcomes that can soften competition while remaining outside the conceptual boundaries of agreement, concerted practice and restriction by object. As a result, behaviour that undermines competition may fall beyond the reach of existing doctrine simply because it does not produce the kinds of evidence the law is designed to recognise.

The analysis in this article shows that the predicates of liability under Chapter I rely on forms of human agency that algorithmic systems do not exhibit. When pricing decisions are produced by adaptive tools rather than deliberate human choices, traditional evidential markers disappear, and responsibility becomes harder to allocate. Comparative developments in the European Union and United States demonstrate that other jurisdictions are beginning to incorporate systemic behaviour into the competitive enquiry.

Reform is therefore essential. A framework centred on system design, deployment and oversight would allow UK competition law to address algorithmic risks while preserving its core principles and ensuring that competitive processes remain protected in digital markets.

Bibliography

Cases

• Argos Ltd and Littlewoods Ltd v Office of Fair Trading [2005] CAT 18
• JJB Sports plc v Office of Fair Trading [2004] CAT 17
• Ping Europe Ltd v Competition and Markets Authority [2020] EWCA Civ 13
• Suiker Unie v Commission (Joined Cases 40–48/73) EU: C:1975:174
• United States v Topkins (ND Cal, 2015)

Legislation

• Competition Act 1998
• Digital Markets, Competition and Consumers Act 2024
• Treaty on the Functioning of the European Union, Article 101

Books

• Maurice E. Stucke and Ariel Ezrachi, Virtual Competition: The Promise and Perils of the Algorithm-Driven Economy (Harvard University Press 2016)

Official Publications

• Competition and Markets Authority, Algorithms: How They Can Reduce Competition and Harm Consumers (2021)
• Competition and Markets Authority, AI Strategic Update (2024)
• European Commission, Competition Policy Brief: Algorithms and Collusion (2017)

Footnote(S):

¹ CMA, Algorithms: How They Can Reduce Competition and Harm Consumers (2021) 12.

² Competition Act 1998, s 2.

³ Suiker Unie v Commission (Joined Cases 40–48/73) EU:C:1975:174.

⁴ CMA, AI Strategic Update (2024) 7.

⁵ Suiker Unie v Commission (Joined Cases 40–48/73) EU:C:1975:174.

⁶ Ping Europe Ltd v Competition and Markets Authority [2020] EWCA Civ 13 [45].

⁷ CMA, Algorithms: How They Can Reduce Competition and Harm Consumers (2021) 18.

⁸ Ibid.

⁹ Argos Ltd and Littlewoods Ltd v OFT [2005] CAT 18 [125]–[130].

¹⁰ JJB Sports plc v OFT [2004] CAT 17 [210]–[215].

¹¹ Ping Europe (n 6) [52].

¹² CMA, AI Strategic Update (2024) 10.

¹³ Suiker Unie v Commission (Joined Cases 40–48/73) EU:C:1975:174.

¹⁴ Competition and Markets Authority, Algorithms: How They Can Reduce Competition and Harm Consumers (2021) 18.

¹⁵ Argos (n 9).

¹⁶ Ping Europe (n 6) [52].

¹⁷ Competition and Markets Authority, AI Strategic Update (2024) 10.

¹⁸ Argos (n 9).

¹⁹ Ping Europe (n 6).

²⁰ Digital Markets, Competition and Consumers Act 2024, Part 1.

²¹ European Commission, Competition Policy Brief: Algorithms and Collusion (2017) 3.

²² United States Department of Justice, United States v Topkins (ND Cal, 2015).

²³ Maurice E. Stucke and Ariel Ezrachi, Virtual Competition: The Promise and Perils of the Algorithm-Driven Economy (Harvard University Press 2016) 145–147.

²⁴ CMA, Algorithms: How They Can Reduce Competition and Harm Consumers (2021) 18.

²⁵ CMA, AI Strategic Update (2024) 10.

²⁶ Digital Markets, Competition and Consumers Act 2024, Part 1.