Competition and agentic AI

3. Competition in the supply of agentic AI systems

A definitive assessment of competition in agentic AI is premature given its relatively nascent stage. Nonetheless, effective competition in potentially transformative technologies is important to provide ongoing incentives to innovate and ensure the benefits are effectively disseminated (OECD, 2024[10]). The purpose of this section is therefore not to predict how competition will develop, or where breaches of competition law may occur, but instead to identify potential risks as a starting point for policymakers to consider and monitor.

There have not been many assessments of competition in agentic AI markets by competition authorities to date. One notable example is the French competition authority’s work which is summarised in Box 2 (Autorité de la concurrence, 2026[30]). There have, however, been several studies considering generative AI markets more broadly, including work by the OECD.1 Box 3 below for a summary of some of this work. Further, as a digital technology, there is a broader context of how competition has developed in other digital markets, including discussions on market tipping or “winner takes all/most” dynamics where one or few firms obtain market positions that are hard to displace (OECD, 2022[31]).

This section first considers the nature and state of competition in agentic AI markets. Next, it considers various aspects of the potential evolution of competition in agentic AI, starting with access to inputs and vertical foreclosure within the context of previous commentary on competition in AI markets, before considering additional issues that may be relevant for the development of competition.

Box 2. French Competition Authority’s Opinion on the Competitive Functioning of the AI agents sector

In July 2026, the French competition authority published its study on the competitive functioning of the AI agents sector. As part of the study, the authority consulted relevant stakeholders and received around 40 responses to its public consultation. The study culminated in a published opinion setting out the authority’s findings and its recommendations going forward. The study considered the emergence of AI agents and agentic commerce, even if the latter was not yet available in France given its potential to be deployed quickly.

Findings

The opinion noted the current concentration in the supply of agentic AI, and that while barriers to entry appear lower than for training generative AI models, since agentic AI can be deployed using third party models, there are several potentially significant barriers to expansion in agentic AI. In particular, the ability for suppliers to scale may be constrained by their ability to reach users and access critical data, as well as several other technical barriers which may restrict customer acquisition.

More broadly, the study found that AI agents are emerging as an increasingly important gateway for digital services, and that this “platformisation” could raise a number of risks to competition. This included disintermediation of existing market structures, and concerns around the proper functioning of AI agents in facilitating digital transactions, such as potentially acting in a discriminatory or non-transparent manner. The study noted that barriers to interoperability could further risk how competition develops and the ability for multiple AI agentic solutions to emerge.

The study also examines how agentic commerce could reshape online retail by automating parts or all of the purchasing journey. It identifies risks that merchants become dependent on AI agents to reach customers, while agents’ selection and presentation of offers could restrict product visibility and diversity or favour particular suppliers. The study also reports concerns that automated repeat purchases could lock-in initial choices without reassessing competing offers, reducing opportunities for rival products and new entrants.

Recommendations

The report makes several recommendations following its findings. It notes that the risks to competition could emerge quickly and therefore urges public authorities to fully implement their existing tools to promote competition. In addition, it makes several recommendations to firms operating in agentic AI, including to promote interoperability and portability where possible through contractual and technical terms of use, as well as open standards. The Authority also stresses the need to empower users to carefully compare AI agents and make informed choices, as well as calling for effective enforcement of consumer-protection rules, technical scrutiny and closer co‑ordination between public authorities.

Source: Autorité de la concurrence (2026[14]), Opinion No. 26-A-05 on the Competitive Functioning of the Artificial Intelligence Agents Sector, French Competition Authority, https://www.autoritedelaconcurrence.fr/sites/default/files/commitments/2026-09/26a05_eng.pdf.

3.1. Preliminary look at the nature and state of competition in agentic AI

Before turning to potential competition issues in the supply of agentic AI, the paper first provides a preliminary look at some aspects of the nature and state of competition in agentic AI markets, such as the current market structure, parameters of competition and emerging patterns of substitution.

3.1.1. State of competition and parameters of competition

The competitive landscape for agentic AI is still emerging, and it is not yet clear whether the levels of concentration seen in model development will extend to the supply of agentic AI. Agentic AI services currently appear largely concentrated around three main operators,2 with a fringe of other firms also providing services (Autorité de la concurrence, 2026[14]). Concentration however, particularly in the early stages of a market, is not necessarily synonymous with ineffective competition (OECD, 2018[32]). As explored further below, given this emerging nature, it is more relevant to consider the likely evolution of the market through an assessment of the ease of entry and expansion rather than initial market shares, although this is not to say that current market dynamics are not relevant.

Agentic AI can be considered an extension of previous agent-led markets, and this may be a useful framing in which to consider the nature of competition. Agents provide a range of benefits and enhance the welfare of users in some markets. However, agent markets may also be characterised by information asymmetries, particularly the difficulty in establishing the quality of the agents’ services. The quality of responses from agentic AI systems is likely to be a key parameter of competition, with consumers placing value on the outcomes achieved (Autorité de la concurrence, 2026[14]). Competitive pressure on quality may be weaker where users cannot readily observe or compare it. Feedback and reputation mechanisms may help where the quality of outcomes becomes apparent over time, but may be less effective where interactions are infrequent or outcomes are difficult to verify.

The lower cost nature of agentic AI compared with human agents may result in a higher volume of agentic services or delegated decisions, with overall lower levels of complexity. This could make assessing outcomes easier and therefore make it more likely that consumers, or other third parties, can accurately evaluate and report on the performance of their agentic AI solution. Further, agentic AI solutions may also offer solutions to evaluate the outcomes of services. On the other hand, agentic AI systems, particularly given their likely reliance on generative AI models, may make decisions in a less transparent way than human agents in some situations, potentially making it difficult to assess their performance and hence undermining competition between providers.

As with many markets, prices will likely be an important feature of competition for agentic AI customers. There is currently a mix of commercial models in the AI space, but for consumers, the main offering appears to be a freemium one, whereby a lower quality service is provided for free, with options to improve the service with paid subscriptions or on a per use basis (Autorité de la concurrence, 2026[14]). As with other digital markets, the ability for customers to predict and assess their expected costs from using these services, including when comparing alternative providers, will be relevant for the effectiveness of competition.3 Another relevant factor will be the extent to which agentic AI services are bundled with other services or become integral to existing digital ecosystems.

3.1.2. Dimensions of competition and patterns of substitution

While still in development, there are indications that competitive conditions could vary across different product or geographic dimensions of agentic AI systems. A formal market definition exercise is not within the scope of this paper, but considering these product and geographic dimensions is useful context for any discussion on competition, including whether agentic AI constitutes a single market or comprises several markets. An initial split on the demand side might be between different use categories as identified in Table 1, namely between industrial, enterprise and end-consumers, although whether this demarcation will mark separate markets is unclear. Industrial and enterprise users may have a greater ability to self-supply but will also face broader challenges around integrating solutions into existing systems and, on the whole, requiring greater security and robustness safeguards.

Industrial and enterprise use of agentic AI may ultimately be further segmented if there are differences in the conditions of supply according to the sector, such as healthcare, manufacturing or finance. Whether agentic AI systems specialised in one or more sectors exercise constraints on others could determine the breadth of agentic AI markets going forward, as may whether general solutions offer an asymmetric constraint on specialised ones.4 Beyond potential demarcation by industry, there could also be groups of customers whose requirements differ significantly enough to make conditions of competition distinct.

The extent of fragmentation of competition between agentic AI systems could also vary depending on the functionality of those systems, such as their capabilities or purpose. There could feasibly be demarcation by quality, such as model reasoning capability.5 For industrial agentic AI, use cases may overlap with different industries, but this need not necessarily be the case. For enterprises and consumers, differences could emerge between different productivity boosting applications, as well as when compared to agentic commerce applications.

Another relevant dimension of competition to consider for agentic AI systems is the geographic dimension. Geographically, as with other markets of AI, initial indications appear to suggest that markets are likely to be broad, at least national but perhaps wider, with many of the main operators active across jurisdictions in the OECD and beyond. However, regulations, specialisation or adoption practices may vary across jurisdictions, and could lead to more localised competitive conditions. For example, the transactional services described by the French Competition Authority were not yet available in France at the time of its July 2026 opinion (Autorité de la concurrence, 2026[14]).6

Finally, the discussion above is unlikely to be exhaustive and other demarcations may emerge as the technology and use cases unfold. Whether separate economic markets exist may change over time and will require careful assessment. Key aspects for how broad or narrow agentic AI markets will be may be the extent to which they can easily perform a range of tasks, the demand for specialisation and how crucial proprietary industry or use case specific data will be to providing a competitive service relating to that industry or use case.

3.2. The applicability of previous AI competition analysis to agentic AI

As detailed in Section 2, Agentic AI builds on earlier generations of AI, particularly generative AI, while adding orchestration, tool use and other components that enable systems to take actions. Therefore, from a competition perspective it is likely that many of the same potential competition issues considered for generative AI will apply to agentic AI, particularly given that the latter is still emerging making it all the more difficult to identify unique trends. Nonetheless, it is worth considering which issues appear similar or potentially different compared to previous iterations of AI. This section therefore considers the extent to which some of the potential competition issues identified for other AI markets, as summarised in Box 3, apply to the same extent for agentic AI.

While several similarities appear, not all potential competition issues are likely to be identical. For example, the French competition authority’s report highlights that the barriers to entry in the agentic AI market appear to be lower than for training generative AI models, as agentic AI suppliers can gain access to third party generative models rather than having to train their own which can be prohibitively difficult (Autorité de la concurrence, 2026[14]). In this sense the supply of agentic AI may develop to be modular, allowing independently sourced components to be brought together in a solution, at least in theory (Bostoen and Krämer, 2026[16]). Building an effective product that can enter a market does not guarantee commercial success by itself however, with the French report also noting that barriers to expansion for agentic AI services may be higher than barriers to entry.

Box 3. Summary of main potential risks to competition in AI markets

Following the emergence of ChatGPT in November 2022, generative AI rose to prominence and highlighted the potentially transformative role for AI as a general-purpose technology. Given the importance of preserving and promoting competition in a potentially crucial sector, the competition community also devoted significant time to assessing how competition was emerging in AI markets. This included several reports and studies from competition authorities within and outside of the OECD, and several roundtable discussions and reports from the OECD itself. While there were variations across studies, a general picture emerged regarding common potential risks to competition in the supply of AI. In identifying such risks, the intention was not usually to assess the current competitive conditions in these emerging and developing markets, but instead to conduct forward-looking assessments of areas to monitor with a view to tackling issues promptly if they were to emerge.

The main potential risks to competition for the supply of AI broadly focussed on the importance of accessing several key inputs in order to compete, such as compute, talent and data. Different parts of these inputs may also raise potential competition issues, for example in the supply of compute and access to infrastructure that is vital for supplying AI services. As well as inputs, the importance of accessing distribution methods in order to reach users has also been acknowledged. In this context, the discussion has also noted the vertical integration across many parts of AI supply potentially giving rise to input or customer foreclosure. In this regard, AI’s integration into existing digital ecosystems has been considered a potential mechanism by which incumbent digital firms may be able to leverage their market positions into AI markets.

It has also been noted that some parts of the value chain appear concentrated and concerns have been raised whether AI markets might exhibit tendencies toward concentration given the high costs of training models. In this context, competition authorities have also considered the role of various forms of partnership, investment or contractual relationships between firms in the context of merger control, including whether different forms of non-traditional acquisitions, such as “acqui-hires”, may lead to lessening of competition whilst not being adequately captured by merger control.

In such emerging and complex markets, other risks to competition may emerge and have been raised across previous studies. For example, while there are several legitimate policy issues that may warrant regulation, competition in AI markets may be affected by regulation.

Source: OECD (2024[10]), “Artificial intelligence, data and competition”, https://doi.org/10.1787/e7e88884-en; OECD (2026[28]), Artificial intelligence markets: Recent developments and competition issues, https://doi.org/10.1787/d531d73f-en; OECD (2025[11]) , “Competition in artificial intelligence infrastructure”, https://doi.org/10.1787/623d1874-en.

3.2.1. Access to inputs and vertical foreclosure

As noted in Box 3, many of the identified potential risks to competition in generative AI are linked to foreclosure of key inputs due to vertical integration. As an extension to the same supply chains, with several vertically integrated layers or with vertical ties between suppliers – such as partnerships, investments or minority shareholdings – this discussion appears highly relevant for agentic AI as well (Colangelo and Ezrachi, 2026[33]). While vertical integration can provide significant benefits, it may, under certain conditions, give firms the ability and incentive to foreclose rivals from key inputs, with anticompetitive effects.

Healthy competition in agentic AI, as with generative AI, depends upon the ability of a sufficient number of firms to have access to key inputs. Many of these inputs are similar as for other types of AI, such as requiring access to compute for inference, as well as data, skills, associated software and other necessary inputs. However, the exact mix of inputs required differs. Perhaps most significantly, agentic AI requires generative AI models as a crucial input. On the one hand this means another input which must be accessed. On the other hand, not having to train generative AI models, but instead relying on third-party will reduce the requirement to access compute or data for training.7 Agentic AI also has additional layers in the supply chain, such as the orchestration and harness layer, as explained in Section 2. Currently many orchestration tools appear to be supplied on open-source basis, potentially lowering certain barriers to entry and reducing foreclosure risks.8

Interoperability could be even more important for the competitive functioning of agentic AI than previous versions of AI, due to its likely reliance on integration with other devices (Hunt, Chissell and Mawar, 2025[34]). As such, there may be potential for vertically integrated suppliers to leverage control over adjacent layers to make independent agent providers less viable over time. As Thomas (2026[35]) argues, each supply level must perform well for the overall agentic solution to be effective and thus the inability to access just one layer could be determinative for the overall competitive dynamics. For example, access to APIs providing real-time prices, together with data that are accessible and machine-readable, is therefore likely to influence how effectively services at different levels of the supply chain can interact.9 In this regard, while input foreclosure can involve completely removing access to parts of a market, it can also affect competition if access is technically allowed but partially restricted in ways that prevent access from being fully effective.

This could be particularly relevant for commercial agentic AI where the purpose of the service will be to execute transactions on behalf of consumers. Such a journey will require interactions with a range of different tools to be completed, and inability to access some could render a service ineffective. For example, if an agentic AI tool is unable to access product information of several connected large retailers, then its services will presumably be of lower value compared to an agentic AI solution that could do so.

Beyond access to inputs, another risk of vertical foreclosure relates to access to customers through distribution. If agentic AI suppliers face limitations in accessing distribution networks or existing ecosystems, such as browsers, messaging apps, search engines or operating systems, amongst others, they may face competitive disadvantages that prevent them from achieving scale irrespective of their competitive merit.10 Box 4 highlights how access to distribution can be important in the context of AI markets by summarising the European Commission’s ongoing investigation into Meta regarding third-party access to WhatsApp for AI developers.

Box 4. Investigations into allegations of exclusions of AI competitors from WhatsApp

An ongoing investigation by the European Commission (EC), which took over a similar investigation from the Italian competition authority, into Meta’s conduct in relation to WhatsApp highlights the potential role for vertically integrated firms to affect access to distribution in the AI supply chain. Similar investigations are also ongoing in Brazil and by COMESA. It also illustrates the potential role of interim measures to tackle conduct in a growing market to prevent what the competition authority considers to be potentially irreparable harm.

The investigation involves an announcement Meta made in October 2025 which changed their policy regarding access to WhatsApp for third-party general-purpose AI assistants. The EC opened formal proceedings in December 2025, noting that “Meta's conduct risks raising barriers to entry and expansion, and irreparably marginalising smaller competitors on the market for general-purpose AI assistants”. In February 2026, the EC sent Meta a Statement of Objections and notified it of its intention to impose interim measures. In March 2026, Meta revised its policy, allowing access by initiating a pricing framework for third-party general purpose AI assistants.

The revision did not resolve the concerns held by the EC, which imposed interim measures in June 2026, ordering Meta to restore free access to WhatsApp for third-party general-purpose AI assistants. In the interim measures decision, which is not an infringement decision, the EC noted that at first sight Meta held a dominant position in the market for consumer communication applications, and that the refusal to provide access was a first sight abuse of that dominant position. It held that there was an urgent need to prevent serious damage to the competitive structure of a growing market, general-purpose AI assistants, with the policy risking harm to competition at a key moment in time. It was therefore found that interim measures were warranted to prevent serious and irreparable harm.

Source: European Commission (2026[36]), Commission sends Meta fresh charge sheet on possible interim measures to reverse exclusion of third-party AI assistants from WhatsApp, Press Release, https://ec.europa.eu/commission/presscorner/detail/en/ip_26_805; European Commission (2026[37]), Commission imposes interim measures on Meta to preserve free access to WhatsApp for rival AI assistants, Press Release, https://ec.europa.eu/commission/presscorner/detail/en/ip_26_1276; Autorità Garante della Concorrenza e del Mercato (2025[38]) (2026) A576 – Meta AI: The Italian Competition Authority orders Meta to suspend the terms excluding competing AI Chatbots from WhatsApp, Press Release, https://en.agcm.it/en/media/press-releases/2025/12/A576; Conselho Administrativo de Defasa Econômica (2026[39]), CADE upholds fine against Meta and WhatsApp, Press Release, https://www.gov.br/cade/en/matters/news/cade-upholds-daily-fine-against-meta-and-whatsapp-for-failure-to-comply-with-interim-measure; COMESA (2026[40]) Notice Of Commencement Of Investigation Into Alleged Abuse Of Dominant Position By Meta Platforms Ireland Limited, https://comesacompetition.org/wp-content/uploads/2026/03/NOTICE_OF_COMMENCEMENT_OF_INVESTIGATION_WEBSITE.pdf.

While this case does not relate specifically to agentic AI, it raises issues that may be applicable to it, if not more so. This is because agentic AI marks a potential turning point in the ability of AI to provide end-to-end services within a particular ecosystem (Colangelo and Ezrachi, 2026[33]). With the ability to complete tasks, rather than just provide outputs, this could increase the propensity for user journeys to remain where they start, as opposed to having multiple entry points or calling on different services. If this occurred, it would increase the importance of accessing consumers in the first place and underline the necessity of access to existing distribution systems. Further, for consumer uses of agentic AI in particular, as the competences of agentic AI systems emerge, the lines between inputs and distribution may become increasingly blurred as the systems require increased integration and interoperability with their route to market, such as a device’s operating system or a browser (Bostoen and Krämer, 2026[16]). This not only increases the potential importance of access to distribution methods, but broadens the scope for limited effective access to be provided.

3.2.2. Other potential barriers to entry and expansion

There has been much discussion about the state of competition in digital markets over the past decade (OECD, 2022[31]). Several factors have been discussed as altering competitive dynamics compared to traditional markets, including the existence of multi-sided networks, strong network effects and feedback loops, economies of scale and scope, and zero monetary prices on at least one side of the market.11 These features may combine to create high barriers to entry and expansion, making incumbents entrenched and difficult to displace. Given their potential significance regarding how competition will evolve, this is therefore a useful lens from which to consider agentic AI, and in particular possible barriers to entry and expansion. At the same time, agentic AI is a new technology and therefore it is not necessarily the case that the frameworks that worked for other digital technologies will apply to it, with some arguing that it requires a different, more cautious approach from competition authorities (Colangelo, 2026[41]).

As with other digital technologies, the barriers to entry and expansion for agentic AI are likely to be strongly affected by whether there are incumbency advantages that offer benefits to scale and adjacent operations. If rivals cannot access these benefits, while entry may technically be possible, it may make significant expansion difficult. This can be particularly true if users do not spread their demand across multiple suppliers (i.e. multi-home) but instead select one main provider. While it remains to be seen whether users of agentic AI prefer sole or multi-supply, there appear to be some frictions for users having multiple suppliers. This may be particularly true for industrial or enterprise users where security, integration and training issues raise the costs of using more than one system. However, consumers could also exhibit preferences for single suppliers for example if the interfaces take time to adjust to or if inputting preferences takes time and cannot easily be interchanged between services.12

A key distinction of AI, including agentic AI, compared to other digital technologies is the non-negligible marginal cost that comes from the compute required to run queries on generative AI models (OECD, 2024[10]), suggesting that economies of scale may not be as significant compared to some digital technologies with near zero marginal costs.13 Nonetheless, while setting-up agentic AI systems may be considered a fixed cost, potentially providing economies of scale, this fixed cost is likely smaller than training generative AI models. Further, with increased inference costs as volumes increase, and potentially increasing costs in verification and quality assurance of outputs, it is unclear how the costs of supplying agentic AI itself will scale. Further, as noted above agentic AI links closely to other layers of the value chain and there may be significant economies of scale in supplying various inputs to agentic AI, such as in cloud computing services, in generative AI models or through distribution methods. Such scale is particularly relevant in the context of vertical relationships noted above.

Agentic AI suppliers may also derive potential scale benefits beyond the spreading of fixed costs. For example, if agentic AI systems become important intermediaries for commercial activity providers serving more users may aggregate demand and thereby acquire stronger bargaining power. Agranat and Gal (2025[42]), argue that this could create a reinforcing loop in which stronger bargaining power enables suppliers to secure more favourable terms, attracting further users and strengthening their market positions.

Another feature of digital markets has been the importance of data including the benefits gained from economies of scope and feedback effects (OECD, 2022[31]). Economies of scope may apply to agentic AI systems if operating in adjacent markets provides advantages to suppliers. For example, a firm operating a search engine or productivity service may be able to reuse assets (such as data on user behaviour) from those activities to reduce the costs of their agentic AI service. Such advantages, if realised, may make it harder for stand-alone suppliers to compete on equivalent terms.

Agentic AI systems may also give rise to new forms of user-specific data, such as consumer preferences or accumulated organisational knowledge. Beyond training data and conventional customer information, deployed agentic systems may develop histories of interactions, workflows, approvals, preferences and operational procedures associated with a particular organisation for their users. The ability for users to take data with them if they move suppliers could be especially important for agentic AI services (Hunt, Chissell and Mawar, 2025[34]). Where such information is difficult to export or reuse with alternative providers, switching costs could increase and incumbent agentic AI suppliers may benefit from a degree of lock-in. One example is at the enterprise level, where data security is likely to be a major factor in selecting suppliers, and firms will require confidence in the security of any data transfers. Conversely, interoperability and portability mechanisms that allow users to transfer agent memories, configurations and operating histories may reduce dependence on individual suppliers and support contestability. As discussed in earlier OECD work, the competitive significance of such effects will depend on whether the information is unique, transferable and important for effective competition (OECD, 2026[12]).

High-quality data created by the use of agentic systems could also potentially create feedback loops which enhance the quality of agentic AI systems and enhance their ability to provide meaningful and reliable outputs for users. This could occur for agents destined for individual consumers, such as commercial agentic AI, as incumbents learn how to navigate consumer search behaviour and preferences, or with sector specific or industrial applications, that gradually become tailored to the individual firm. For industrial and enterprise users, having access to operational workflow data and organisational knowledge could provide a strategic resource for product refinement. Feedback loops are perhaps most likely when user experiences with a system create unique data that allows user-specific service improvement, for example on what service features users most value or information that allows an improved product. Bostoen and Krämer (2026[16]) argue that whilst feedback effects could be less relevant to generative AI or AI assistants due to their reliance on pre-trained foundation models, as agentic AI is more likely to rely on reinforcement learning and persistent memory, the ability to learn good and bad outcomes could be important for improving the quality of agentic AI offerings.

Stucke and Ezrachi (2017[43]) raised the potential for significant network effects in the context of digital assistants as early as 2017, arguing that as more users operate on a digital assistant, more services will seek to be available on it, attracting more users, etc. It is currently unclear how substantial direct network effects will be for agentic AI, although there could be more potential for indirect network effects, particularly if commercial agentic AI develops such that it operates with platform-like tendencies matching users with suppliers. If there are costs to access an agentic AI system operating as a platform, in particular for sellers, this would likely give rise to indirect network effects since more users attract more sellers, and so forth.

Notes

← 1.  For example, competition authorities in over 13 countries have conducted market studies on AI or aspects of AI markets.

← 2.  Drawing on Sensor Tower usage data, the French Autorité de la concurrence reported that OpenAI, Google and Anthropic together accounted for more than 84% of the measured global usage of AI agents in May 2026.

← 3.  See, for example, dynamics as discussed in the context of cloud computing services and its effect on competition (OECD, 2025[75]).

← 4.  While generalised systems could offer close alternatives to specialised applications in some circumstances, the opposite is less likely.

← 5.  For example, if there were different demand and supply profiles for lower cost, or compute requirements, agentic AI systems with lower capabilities compared to higher cost, more powerful systems.

← 6.  Restrictions of trade between countries could also limit the scope of geographic markets, as well as other government policies that affect the importation or exportation of agentic AI or related technologies.

← 7.  As explained in Section 2 however, this does not mean that compute and data will not still be required. Further, the terms on which a generative AI model is accessed become crucially important.

← 8.  In this sense it is unclear if there are specific competition issues at a horizontal level in relation to orchestration services, although this could evolve over time.

← 9.  It is unclear how significant the need for APIs will be with agentic AI systems in theory able to manually navigate to retrieve information. In particular, even if APIs were not needed, this would require potentially significant amounts of manual navigation by agents within the agentic AI system, which could have sizeable cost implications.

← 10.  While focussing on broader AI assistants rather than just agentic AI, the European Commission’s recently issued binding specification measures regarding interoperability with the Google Android operating system under the Digital Markets Act to Google, provides an example of the potential importance of distribution systems in promoting competition between AI services. As mobile devices are likely to be important distribution routes for agentic AI solutions, it appears to be highly relevant for it as well (European Commission, 2026[95]).

← 11.  Economies of scale and scope concern how average costs of production varies with volume, within the same or adjacent markets respectively. Related concepts such as network effects and feedback effects are considered below. The discussion focusses on the supply of agentic AI itself, rather than the supply of inputs, such as compute, including cloud services, or generative AI models, which are discussed in other reports. Many of these inputs demonstrate significant returns to scale. See, for example (OECD, 2024[10]; 2025[11]; 2025[75]).

← 12.  Raising again the potential importance of data portability between service providers.

← 13.  Note however that this does not mean that agentic AI services would still not benefit from economies of scale, just that in comparison to services with zero marginal costs the economies would appear to be less strong.