Competition and agentic AI

4. Competitive effects of agentic AI

This section considers how agentic AI could impact competitive dynamics in markets where it is adopted, either by enterprises or consumers. Agentic AI appears to offer both potential pro- and anti-competitive effects. Many of these issues, including some in relation to agentic AI, were discussed in the recent OECD paper on AI’s downstream competitive effects (OECD, 2025[2]).

Rather than seeking to consider definitive benefits or harms to competition, including potential breaches of consumer law, the section focusses on identifying and exploring the main categories of potential effects. The section does not consider broader effects arising from agentic AI in other markets, such as the macroeconomic benefits from greater productivity growth, beyond the extent to which they focus on competition. Neither does it consider other risks, such as those to privacy, democratic processes or society more generally (Stucke and Ezrachi, 2017[43]).

This section is organised around three areas where agentic AI could affect competition in downstream markets. First, it considers how the use of agentic AI solutions by consumers, and the emergence of commercial agentic AI, could alter consumer search and demand dynamics in markets. Second, it considers how industrial agentic AI adoption could affect competition in those industries, focussing on how it could raise barriers to entry and expansion. Finally, it considers the impact of agentic AI on collusion between firms, building on an area that the OECD has considered in some depth already, algorithmic collusion.

4.1. Consumer search and purchasing through enhanced delegation

Agentic AI provides potential for widescale automation of agentic services. Human economic agents are not new and have been used in a variety of markets for a long time, providing brokerage, intermediary or consolidation services. Agents are a form of intermediation and as such can also draw parallels from other digital intermediation markets, such as online search. The agentic nature of AI will likely have effects on competition in those downstream markets. Immorlica et al. (2024[44]) argue that AI assistance technologies should be viewed as economic agents, although much is still not known about how they will act as agents. (Hadfield and Koh[45]) urge caution in attributing human characteristics to agentic AI systems due to the emerging nature of the technology, its complexity and the overall difficulty in understanding how generative AI systems reach their outcomes.

As with other principal-agent situations, the business model that agents operate under can be an important part of market functioning. In particular, the remuneration model can risk creating a mismatch between the incentives of the agent and the objectives of the client. For example, settings where agents are remunerated by suppliers rather than clients can create incentives for the agent to favour some suppliers over others, even if these are not the best match for the user’s preferences. Similarly, commission models can create incentives for agents to increase overall spending, even if this is not in the client’s best interest. It is unclear whether agentic systems could be less susceptible to the potentially conflicting incentives caused by remuneration models, although given that they will be developed and deployed by firms that capture the return from their use, it is perhaps difficult to argue that a firm’s agentic AI systems would be immune from acting in their financial interests. For example, if AI agents are given an objective of pursuing profit maximisation they may be as, or more so, motivated by the financial incentive than human agents.

4.1.1. More effective consumer search

One of the most obvious potentially positive effects of agentic AI or agentic commerce, is significant search and transaction cost reductions for consumers in relation to online information.1 Such effects have been argued in several papers, such as Bichler (2026[46]), Hadfield (2025[45]) and others. Indeed, the promise of consumers being assisted in their interaction with markets by technology predates the emergence of agentic AI (Gal and Elkin-Koren, 2017[47]). With agentic AI capable of interpreting an instruction from a consumer, such as a desire to purchase a good or service, then engaging in automated search and execution, consumers could have access to wider, more detailed and deeper searches on goods. This could also include ongoing monitoring of changes in price or other variables. These benefits could accrue to consumers who harness agentic AI services, but may also accrue to those who do not, depending on whether firms are able to discriminate against them.2

The intuition for how a proliferation of automated agents able to analyse and assess substantial amounts of data at relatively low cost, compared to manual search, could significantly enhance the demand side of markets is straight-forward. If consumers have imperfect information, they are unlikely to maximise their welfare, but the competition between firms will also be affected. Better informed consumers generally exercise stronger demand-side pressure on firms, increasing incentives to improve their offering in order to capture more consumers.3 Dong, Luo and Xu (2026[48]) show through a model of search that as the complexity of consumer preferences increases, the potential value of agentic search over manual search increases. This vision of enhanced consumer search hinges on several assumptions, including that the agent is sufficiently capable, robust and objective to act as the consumer’s agent, and that the agent can freely navigate the range of potential options available.

Another way in which consumer search might be enhanced could be reduced effects of choice architecture or framing on consumer decisions (Hunt, Chissell and Mawar, 2025[34]). Effects of behavioural biases and how firms exploit them are well known (OECD, 2022[49]), and there is potential agentic AI systems will be less susceptible to these, allowing choice to be more closely aligned to preferences rather than the way they are framed. This can include reductions in information overload, allowing decision making over more transactions to incorporate a broader range of information than currently possible under human oversight (Bichler, 2026[46]). The extent to which this is likely to happen is unclear. Firms may also adjust to manipulate the outputs of agentic models, which are themselves trained largely on human data which may limit the reduction of bias they exhibit, and as explored in more detail below, agentic AI may also have other biases of their own.

While not directly related to search, consumer facing agentic AI could also reduce the transaction costs for consumers to join and exercise joint purchasing power. For example, rather than having to identify these opportunities, research each one and provide suitable permissions, an agent could be provided authority to do so on behalf of consumers. This aggregation of demand could provide opportunities for consumers to exercise collective buying power, particularly in markets where there is limited differentiation or where there are large volumes of shared demand, such as in residential energy markets or road fuel. Such aggregation would heighten competition to obtain the large volumes of business, although there could be a risk that at the extreme if all demand were to be joined it could lead to a hollowing of rivals as the market becomes all or nothing for suppliers. If there are barriers to entry in the relevant market, this could even worsen outcomes in the longer run.

4.1.2. Non-neutral or ineffective agents

While agentic AI could lead to significant benefits for consumers, there may be risks to competition if consumer purchasing decisions are increasingly delegated to agentic AI systems (OECD, 2025[2]). Some of the risks relate to the opaque nature of agentic AI systems and the potential for consumers to over-rely on them. Some of these dynamics are already present through the existence of recommender systems or online search, but could be magnified and exacerbated by agentic AI. Evidence of current AI agentic systems suggests that divergences are already emerging in the outcomes of queries depending on the provider used, illustrating the potential importance of how the agentic AI system works in consumer outcomes (Autorité de la concurrence, 2026[14]).

Many of the potential pro-competitive effects from agentic AI rely on the empowerment of consumers with effective agents that work on the consumer’s behalf to better their outcomes. In essence this equates to consumers being able to use agentic AI services to find the best quality matches between their preferences and product offerings.4 As agents lose the ability to do so, not only will consumer welfare decrease, but the incentives for firms to compete will also decrease. This is akin to the more general principal-agent problem relevant for many traditional markets, as explored in more detail above. Namely, how to ensure that an agent acts truly in the best interests of their client. With AI as an economic agent, Immorlica et al. (2024[44]) argue that the value of the outputs will be measured based on perceived helpfulness, potentially providing a reward for being convincing rather than necessarily maximising the welfare of users.5

As with other intermediation services, there may also be risks relating to AI purchasing agents being influenced, steered or manipulated by sellers or platforms. Such non-neutral action from an agent could clearly distort competition, especially if it is hidden or hard to measure.6 There could be particular risks, for example, if agentic AI systems are vertically integrated and used to self-preference other products or services offered by the group. The opacity of agentic AI outputs may make such conduct even more difficult to identify than in existing technologies such as online search (Bostoen and Krämer, 2026[16]). Even where transparency is provided, whether it is truly verifiable is doubtful.

While agents that act on behalf of another party other than their client pose clear risks to competition and outcomes, agentic AI could also exhibit biases of its own, irrespective of intention. As a technology still in development, how agentic AI will perform is hard to determine. Evidence based on simulated marketplaces using recent LLM based agents suggests that even at this stage the capabilities are nearly sufficient to perform welfare-enhancing searches on behalf of consumers, but that performance may suffer as volume increases and that agents appear to exert preferences for the first proposal (Bansal et al., 2025[50]).

An open question will be the extent to which agentic AI can accurately measure and interpret consumers’ preferences and map this to real life product differentiation. It is reasonable to assume that if the proportion of agentic AI use is high and a particular form of differentiation cannot be captured by agentic AI over time, then that differentiation will likely disappear and product offerings will appear more homogenised on that element. While this could intensify competition on price or on other factors, it is likely to ultimately reduce welfare if that differentiation offered increased value to diverse consumers. The propensity for LLM-based models to continue to generate fictional outputs, known as hallucinations, could also reduce the accuracy of LLM-based agentic AI services and affect their ability to perform on behalf of consumers (Alansari and Luqman, 2026[51]).

As already documented in the context of the digital economy, high usage of recommender systems has the potential to overly aggregate demand and hence dilute diversity between suppliers, meaning firm exit and success for a few leading firms (Fletcher et al., 2023[52]). It has been speculated that agentic AI will worsen these problems, as emerging evidence suggests that LLMs are more likely to steer consumers towards more expensive products independently of perceived quality (Zac and Gal, 2025[53]). This may not lead to better outcomes for consumers dynamically, particularly if there is more diversity in preferences than is picked up in the models.

Further, firms that rely on customer acquisition through agentic systems will begin to compete for agents rather than consumers. While this might drive better outcomes, if parameters of competition vary between agent and consumer, outcomes may worsen. Control over which sources, products and services are presented to consumers could also give agent providers greater influence over the allocation of demand. Advertising, paid placements, commercial partnerships and technical integrations may shape visibility and selection, creating risks of discrimination, self-preferencing and a narrowing of consumers’ effective choices (Autorité de la concurrence, 2026[14]). Competition may therefore increasingly centre on access to, and control over, these interfaces, with implications for suppliers’ ability to reach customers and for the barriers to entry and expansion discussed below (FCA, 2026[54]).

More generally, some research suggests that generative AI agents may exert preferences for the first proposal retrieved, which could suggest that suppliers will seek to compete over the speed at which their offer is retrieved by consumer agents rather than other aspects of product quality or price (Bansal et al., 2025[50]).7 Firms will also have access to agentic AI to produce and market their products, and some evidence suggests that this could lead to increased noise on product quality compared to reality, making the ability to find good matches harder (Bichler, 2026[46]). This, as well as other mechanisms, could provide ways in which agentic AI models are manipulated to favour market outcomes other than those that best match user preferences.

4.2. Effects of agentic AI on barriers to entry and expansion in other markets

4.2.1. Lowering barriers to entry or expansion and disrupting markets

With increased abilities to operate autonomously, agentic AI has the potential to fundamentally reshape a whole host of tasks, including applications by firms to boost productivity. The potential for AI adoption to reduce labour requirements, lower minimum efficient scale, automate tasks and increase productivity has been examined in OECD (2025[2]). These mechanisms may lower barriers to entry and expansion where agentic AI further extends the range of tasks that can be delegated. Lowering barriers to entry generally would improve competition across markets, if agentic AI is widely available and accessible to a range of potential users who do not face overly restrictive costs or inability to utilise, for example due to a lack of relevant data or users. Broader productivity improvements can also boost output and improve competition (Hadfield and Koh, 2025[45]).

Another mechanism through which agentic AI may reduce barriers to entry differs from the labour-substitution effects discussed in earlier OECD work. Previous analysis highlighted how AI systems can reduce the need for certain categories of labour, lower minimum efficient scale and increase productivity (OECD, 2025[2]). In addition, Agentic AI may further reduce barriers to entry not only by automating individual tasks but also by reducing the managerial, organisational and co-ordination capabilities required to operate a firm. By planning tasks, monitoring progress, interacting with software systems and executing multi-step workflows, they may allow smaller firms to perform activities that previously required larger managerial, administrative or support functions. The effect could be particularly relevant in knowledge-intensive sectors where growth has traditionally depended not only on specialised expertise but also on the ability to co-ordinate increasingly complex operations. The extent to which such systems can reliably substitute for managerial and organisational functions remains uncertain and will depend on their capabilities, reliability and the degree of supervision required.

Disruption of existing business models could also provide pro-competitive benefits. This could include reducing the need for large scale intermediaries or platforms in some sectors – online search or online marketplaces for example. This would be beneficial if these markets are characterised by low levels of competition now.

Agentic AI could provide significant disruption to these models by removing the need to use intermediaries, for example by combining information on one platform, as they make it easier to intermediate via the solution itself. Some have argued this could lead to significant disruption to existing large platforms, for example (Marro and Torr, 2025[55]). The effect on market structure may depend on whether agentic AI primarily disintermediates existing platforms or instead becomes a new intermediary layer, and whether operators at that level obtain market power. By acting on behalf of users in search, comparison, negotiation and purchasing, agentic AI systems could reduce dependence on some existing intermediaries and lower switching costs between suppliers.

4.2.2. Increased barriers to entry and risk of foreclosure

While productivity boosting technology may make entry easier, whether it provides the ability to do so, and indeed whether it also provides the environment to allow expansion after entry, is not inevitable.

Different firms are likely to be better positioned to take advantage of the benefits of agentic AI. This in itself is not necessarily a competition issue. In fact, it is an example of the competitive process working and of the productive and dynamic efficiencies of markets, reallocating to the most efficient and innovative firms. However, this may not be the case if certain firms have unassailable advantages in their potential ability to leverage agentic AI, for example due to their existing market positions in the relevant market or in an adjacent one. More broadly, if agentic AI leads to large disintermediation that creates new platforms, it may create new gatekeepers for consumers. In such a scenario, firms that wish to access consumers may have little choice other than to transact with agentic AI suppliers, providing them with the ability to set the terms of trade. While this can lower costs and simplify the distribution of profits across firms, it could also raise similar competition issues as seen with other digital platforms, such as exclusionary behaviour and self-preferencing, or new forms of intermediation power.8 This could be particularly the case if consumers do not multi-home across suppliers.

Further, agentic AI has the potential to become a new gateway to the internet and become the means by which consumers access digital services (ARCEP, 2026[56]). As such, if a small number of agentic AI systems become important routes through which users access products and services, they may themselves acquire intermediary power, leading to a form of platformisation of agentic AI. In that case, firms could increasingly compete for visibility to agents rather than directly for consumer attention, potentially making access to agent ecosystems, user relationships and accumulated behavioural data important competitive assets. The same technology could therefore reduce barriers to entry in some circumstances while creating new bottlenecks in others. This remains an open empirical question and may depend on adoption patterns, business models and interoperability between competing agent ecosystems.

If agentic AI systems become the main gateways, this could also reshape competition in online search and advertising. Where users obtain answers or access services without leaving an agent’s interface, traffic to underlying websites may decline, putting pressure on advertising- and subscription-funded business models and potentially affecting the variety and quality of online content (Autorité de la concurrence, 2026[14]).

Beyond potential market power in agentic AI itself, competition in downstream markets could be affected negatively by agentic AI if incumbents in adjacent markets have advantages that others do not. For example, data is likely to be both a key input into an agentic AI system as well as an enabler of its capabilities. This latter point is particularly likely to be the case the more specialised the task. For enterprise or industrial users, if downstream incumbents have unique access to data that cannot be easily replicated, this could make it difficult for other firms in that sector to utilise agentic AI as effectively, potentially reducing their ability to compete. As with other feedback effects, this is self-reinforcing: the incumbent continues to grow market share, acquiring more data, further improving the service, and so on.

4.3. Agentic collusion

Co-ordination between competitors can weaken incentives to compete on price, quality, innovation and other dimensions.9 Agentic AI systems may change both the ability and likelihood of firms to co‑ordinate as well as impact the evidentiary trail for assessing the legality of such conduct.

Concerns about the role of algorithms in facilitating co-ordination predate agentic AI. This section therefore begins by highlighting key findings of previous OECD work in Box 5. It then examines how agentic AI systems may affect three mechanisms through which co-ordination can arise or be sustained: the implementation of agreements reached by human decision makers; common-provider arrangements; and co-ordination arising through autonomous communication or adaptation. In each case, the focus is on how greater autonomy and the delegation of linked commercial tasks may change the way co-ordination develops, is implemented and is sustained.

Box 5. Previous OECD work on Algorithms and Co-ordination

Previous OECD work identified several mechanisms through which algorithms and software systems could facilitate co-ordinated outcomes:

• Implementation of cartel agreements. Firms may use algorithms to implement cartel arrangements more efficiently, for example by applying agreed pricing rules, monitoring rivals' conduct, detecting and responding to deviations and automatically adjusting prices. Algorithms may also make cartel conduct harder to detect. Faster and more accurate monitoring may reduce opportunities to deviate and make an existing cartel easier to sustain.

• Information exchange and enhanced transparency. Algorithms and digital systems can increase the speed, frequency and granularity with which firms observe rivals’ prices and other competitively relevant information, reducing strategic uncertainty and potentially making co-ordination easier to sustain. The competitive significance depends on the nature, timing, frequency and granularity of the information, whether it is genuinely public in practice, and the characteristics of the market. Earlier assumptions that aggregated, historical or publicly available information is necessarily low risk may therefore not hold in all settings.

• Common-provider and hub-and-spoke arrangements. Competing firms may rely on a shared provider of pricing or decision-support software. Concerns may arise where the provider uses non-public information from several clients in setting or recommending prices, facilitates information exchange between competitors, or aligns their commercial decisions through a common strategy or methodology. Use of the same provider does not by itself establish co-ordination: relevant factors include the information flows, the role of the provider, firms’ awareness and the applicable legal framework.

• Autonomous learning and tacit co-ordination. Research has shown that reinforcement-learning pricing algorithms can sustain supra-competitive prices without explicit communication or instructions to collude, using patterns of monitoring and retaliation like those that sustain traditional tacit co-ordination. Evidence from Germany’s retail gasoline market for example found higher margins following adoption of algorithmic pricing, providing real-market evidence that algorithmic interaction can materially alter competitive outcomes.

Source: OECD (2017[7]) Algorithms and Collusion: Competition Policy in the Digital Age; OECD (2026[57]) Information sharing in Competition Policy; OECD (2025[9]) Algorithmic pricing and competition in G7 jurisdictions; Assad et al (2024[58]) Algorithmic Pricing and Competition: Empirical Evidence from the German Retail Gasoline Market.

Beyond the theoretical concerns around algorithmic co-ordination, there have also been examples of competition authorities taking enforcement action against firms for co-ordination involving the use of software and algorithms, including software used to implement cartels and common-provider arrangements involving information exchange and pricing alignment. While not widespread, these cases illustrate the application of established competition rules where software facilitates familiar forms of prohibited conduct. 10 At the time of writing there does not appear to be any published enforcement decisions involving agentic AI systems and collusion, but it is worth considering where risks might arise as the technology continues to develop.11

4.3.1. Cartel implementation and common provider arrangements

Agentic AI systems may extend the role of the algorithmic tools discussed above by allowing firms to delegate more broadly decisions on how cartel agreements are implemented. In this setting the agentic AI system would remain a tool for giving effect to the human agreement, but with greater discretion and autonomy over its implementation (CMA, 2026[59]; OECD, 2017[7]). Drawing on the capabilities discussed in Section 2, agentic AI systems may reduce the cost of organising, implementing and maintaining collusive agreements. For example, if a firm introduced a new sales channel covered by the agreement, the system could determine which price lists needed updating, carry out those changes and incorporate the channel into monitoring of adherence. This could reduce the need for participants to specify and programme each adjustment themselves.

Natural-language interaction and high-level instructions may also affect willingness to delegate commercially sensitive tasks and reliance on automated recommendations.12 Experiments indicate that delegation to AI can increase dishonest behaviour in some settings, while model-level safeguards may reduce but not eliminate responses to problematic instructions (Köbis et al., 2025[60]). The AI system controls described in Section 2 may on the other hand support compliance, but their effectiveness in preventing the facilitation or implementation of cartels remains an empirical question.

Common-provider arrangements can also raise concerns without a prior horizontal agreement between competing firms’ human decision makers. Agentic AI systems may extend this influence to planning, communications and actions across several commercial variables, rather than only pricing recommendations. For example, instructions set by the provider could shape which customers systems approach, the offers they make and the terms they accept in negotiations. A common co-ordinating strategy could therefore influence several linked activities through the same service. Use of the same model or service does not itself establish co-ordination. Technical and organisational separation may reduce cross-client information flows but would not address a common strategy aligning competitors’ decisions. The broader significance also depends on concentration among providers and, more directly, on the extent to which competing firms rely on the same service (OECD, 2026[57]).

4.3.2. Autonomous co-ordination between agentic AI systems

A further question is whether agentic AI systems can develop and sustain co-ordinated strategies through their own interactions, even without instructions to co-ordinate from the firms using them or a common provider. As summarised in Box 5 earlier research showed that pricing algorithms could sustain supra-competitive outcomes without explicit communication or instructions to collude. Building on this, more recent experiments using large language model-based pricing agents have found similar supra-competitive outcomes. This includes agents responding to rivals’ price cuts in a way which may help sustain higher prices (Fish, Gonczarowski and Shorrer, 2025[61]).13 These findings are based on experiments and outcomes may depend on market conditions and the objectives, instructions and information given to the systems (Calvano et al., 2026[62]).

The mix of systems used by competing firms may also matter. Differences in objectives, information and decision making methods may make co-ordination harder to sustain. Experimental work finds that differences in the weight given to future profits, access to data or underlying algorithms can weaken co-ordination. Differences in model size do not necessarily have the same effect: in some experiments, they produce leader‑follower relationships that sustain co-ordinated outcomes (Keppo et al., 2026[63]). Wider adoption therefore need not have a uniform effect across markets.

How agentic AI systems retain and use information may also affect co-ordination. In the pricing experiments discussed above, agents’ written plans and insights were carried forward alongside market histories to inform subsequent decisions (Fish, Gonczarowski and Shorrer, 2025[61]). Research has also considered that agentic AI systems could condition their decisions on other systems’ source code or simulated responses or even strategically choose which information to retain for future interactions. These possibilities could alter the scope for commitment and co-ordination, although their practical feasibility and implications for competition remain uncertain (Hadfield and Koh, 2025[45]).

In addition, buyer-side agentic AI systems may further change these interactions. Rather than behaving as passive purchasers, they may respond strategically to sellers’ prices and alter the demand those sellers face. In simulations involving two algorithmic sellers, a learning buyer able to vary its demand obtained lower prices and greater buyer surplus than a passive-buyer benchmark (Calvano et al., 2026[62]). This illustrates how buyer behaviour can change market outcomes, not that buyer-side systems necessarily eliminate seller co-ordination. Their effects may be reinforcing or disruptive, depending on their design and behaviour and the choices available to buyers and sellers (Autorité de la concurrence, 2026[30]; Mukherjee and Chang, 2025[64]; Cheng and Aseri, 2026[65]).

4.3.3. Legal assessment and compliance

Autonomous co-ordination even without human involvement is not necessarily tacit. Systems may exchange strategic information or proposals for aligned conduct, including through messages that their operators do not review or understand (Motwani et al., 2024[66]).14 The distinction can become difficult to establish where market actions themselves convey signals. Whether an interaction is autonomous from human decision makers is therefore a separate question from whether the systems communicate to co-ordinate their conduct.

This distinction also matters to the legal assessment. Where systems acting for competitors exchange and accept proposals to align future conduct, authorities would need to establish whether the interaction meets the applicable requirements for an agreement, concerted practice or other prohibited co-ordination, and on what basis it can be attributed to the firms concerned. Where systems only adapt to observed conduct, the familiar distinction between collusive outcomes and conduct prohibited by competition law remains relevant. The applicable tests differ across jurisdictions, and the absence of a prior human agreement does not by itself resolve these questions (OECD, 2023[8]; 2026[57]).

Relevant evidence may include system instructions, data flows, communications and records of actions. Complete records may help authorities reconstruct conduct, while missing logs or short-lived exchanges may make that harder. Depending on the issue under investigation, establishing what information a system received, what it communicated and how it acted may be more useful than explaining every internal model operation. Assessment may also need to examine interactions between systems, since testing an agent in isolation may not reveal co-ordination risks or weaknesses in safeguards that emerge when it encounters other agents. (Hammond et al., 2025[6]).

The practical significance of these mechanisms depends on which decisions firms delegate and how extensively competing firms use the systems. A firm may allow an agentic AI system to gather information or recommend prices while retaining approval over changes or authorise it to act and respond to rivals without reviewing each decision. Human involvement can affect how co-ordination operates but does not necessarily prevent it. Delegation also does not in itself relieve firms of responsibility for compliance with competition law: conduct that would be unlawful if undertaken through a human intermediary does not become lawful merely because an agentic AI system performs the same role. (Ohlhausen, 2017[67]) Greater capability may make broader delegation feasible; whether firms choose to entrust sensitive commercial decisions to their systems, and how they supervise those decisions, remain important questions.15

These issues also raise questions about prevention. Beyond enforcing existing prohibitions, should firms be expected to assess and mitigate risks of harmful co-ordination when developing and deploying agentic AI systems? Such expectations could concern model training, system design or the way systems are used in particular markets. Their scope, effectiveness and legal basis would need to be assessed, particularly where the concern is tacit co-ordination rather than prohibited communication. Section 5 considers policy questions more broadly.

Notes

← 1.  It is unclear at this stage how agentic AI would affect consumer search for offline retail if information on such offerings is not available online.

← 2.  If firms can discriminate between users with agentic AI and those without, it could lead to situations with two-tiered pricing and worse outcomes for those unable or unwilling to access technology. In an online environment, there appears to be more potential for firms to discriminate between consumers, although whether this materialises with agentic AI remains to be seen.

← 3.  As explored further below, on the other hand, there is a risk that if all consumers are too well informed and act on all information (such as price dispersion), this could actually erode incentives to compete on price as firms may also be able to react to lower prices and adjust accordingly, making any gains short-lived.

← 4.  The ability to gauge product quality will also be an important dimension of this and is likely to be aided by trustworthy and robust feedback mechanisms for online goods and services, such as verifiable and auditable reviews.

← 5.  This also relates to the discussion above around the difficulty in evaluating the performance of agents in general when there are information asymmetries, which can generally be assumed when delegating to an agent.

← 6.  LLMs are notoriously hard to decipher (Zhao et al., 2024[88])}. In addition, with agentic AI, the various layers interact in ways to efficiently perform tasks, adding further complications for transparency (Thomas, 2026[35]).

← 7.  This effect appeared to exist across the range of models studied in the paper, which the authors suggest could imply a general limitation in how current large language models interact with choices that contain a temporal element.

← 8.  For a discussion of competition risks associated with digital platforms and ecosystems, including exclusionary conduct, self-preferencing and the exercise of intermediation power, (OECD, 2021[26])). For an assessment of how similar risks may arise in relation to agentic AI, see (Autorité de la concurrence, 2026[14]).

← 9.  Throughout this section, “co-ordination” is used as a broad term encompassing both explicit and tacit forms of alignment between competitors. Where relevant, the discussion distinguishes between conduct involving an agreement, concerted practice or information exchange, and tacit co-ordination arising through interdependent behaviour.

← 10.  Examples include cases involving automated pricing software used to implement the Trod/GBE cartel and software supporting customer allocation between energy suppliers (CMA, 2016[74]; OFGEM, 2019[96]). Proceedings concerning RealPage have also addressed allegations of information exchange and pricing alignment through a common provider (US DOJ, 2025[82]). The examples so far tend to concern established forms of co-ordination implemented or facilitated through software, rather than demonstrating that agentic AI systems have independently generated co-ordinated strategies through autonomous interaction (OECD, 2026[57]).

← 11.  It is likely that the practical significance of these issues will depend in part on the extent to which firms delegate commercially sensitive decisions to agentic AI systems, ranging from gathering market information and recommending prices to setting or adjusting prices and other terms with limited human approval.

← 12.  Related behavioural research examines how presentation and retained user control affect reliance on automated systems, and whether explanations help users distinguish appropriate from inappropriate recommendations (Dietvorst, Simmons and Massey, 2018[73]; Bansal et al., 2021[72]). These findings concern human‑algorithm interaction rather than cartel participation specifically.

← 13.  Recent theoretical work also illustrates that collusion is not an inherent consequence of algorithmic learning: some classes of learning algorithms can be shown not to sustain collusive outcomes in particular benchmark environments (Nisan and Zerah, 2026[91]).

← 14.  (Motwani et al., 2024[66]) examine capabilities relevant to covert communication among AI agents, rather than evidence of collusion between firms in commercial markets.

← 15.  The 2026 PYMNTS Intelligence and Visa Acceptance Solutions study surveyed 1 185 merchants in the United States, Brazil and the United Arab Emirates, and found that 46% identified pricing and the final price paid as the function they were least willing to cede to AI agents (PYMNTS, 2026[93]). However, it should be noted that these refer to reported preferences, rather than the proportion of firms that have or have not delegated pricing decisions in practice.