Competition and agentic AI

2. Understanding agentic AI and emerging market developments

Understanding how agentic AI systems work, how they are supplied and how they are being adopted is important for assessing their potential effects on competition. Building on earlier OECD work, (OECD, 2024[10]; 2025[11]; 2025[2]) this section identifies the main characteristics and market developments relevant to the discussion of potential competitive effects in Sections 3 and 4.

It first explains how agentic AI differs from many earlier AI applications by pursuing objectives through a series of actions rather than only producing specific outputs. These systems may connect to multiple information sources and tools, retain information over time and operate within controls that shape what they can do. Their design may therefore affect their capabilities, interoperability, users’ ability to switch providers and the extent to which they can understand and control systems acting on their behalf.

The section then considers how agentic AI is being incorporated into consumer, enterprise and industrial applications, together with emerging business models, supply arrangements and factors affecting adoption. Systems can be assembled in different ways, creating opportunities for a range of providers, but may also depend on particular suppliers, services or routes to users. Competition may therefore be shaped by relationships across the supply chain, access to key inputs and services, and whether agentic AI can deliver useful tasks reliably, securely and at an acceptable overall cost.

2.1.  What are agentic AI systems and how do they work?

The potential significance of agentic AI to competition policy lies less in one underlying technology change than a shift in the tasks and decisions that users can delegate to AI systems. Many AI applications have been used primarily to generate specific outputs, for example, answering a query, summarising information, drafting a report or producing a segment of code. Agentic AI systems can go further by breaking an objective into steps, interacting with tools, applications, data sources and other AI agents, and adjusting their approach in response to results (OECD, 2026[1]; 2026[13]). They may therefore have the potential to make and execute decisions rather than only support them. By reducing the effort involved in searching, comparing and acting, they could also make delegation worthwhile for more routine or frequent activities, such as managing subscriptions or arranging repeat purchases (CMA, 2026[15]).

A travel planning example illustrates this difference. Asked to help organise a short trip to Paris, a non-agentic AI chatbot might propose an itinerary and suggest areas in which to stay, leaving the user to search for and book the relevant services. An agentic AI system, on the other hand, could search for offers matching the user’s dates, budget and preferences, revise the plan where necessary and, depending on its permissions, complete the bookings or request approval before payment. Figure 1 illustrates this shift from providing specific outputs to pursuing an objective through a series of actions. In practice, the distinction is not binary: AI assistants can incorporate agentic capabilities to varying degrees, and the boundary between assistants and agentic AI systems is likely to become increasingly blurred as these capabilities develop (Bostoen and Krämer, 2026[16]).

Figure 1. Stylised example comparing task of agentic AI to earlier systems

Note: The comparison is stylised. The figure illustrates differences in function and the degree of delegation, rather than whether a system comprises one agent or multiple co-ordinated agents. At the time of publication, many agentic AI systems still require human verification and interaction at key stages. For example, a system returns live booking options without completing the entire booking process. While the prompts are different, even with the same prompt the earlier system would not be able to deliver the outcome of the second.

Source: Based on OECD (2026[1]), The agentic AI landscape and its conceptual foundations, https://doi.org/10.1787/396cf758-en.

Agentic AI systems combine generative AI models with software, information and connections to carry out such tasks. The following paragraphs describe some of the key components and how they interact. These features underpin the later analysis of access to inputs and services, users’ ability to switch providers, and the consequences of delegating decisions in Sections 3 and 4. Figure 2 draws on existing literature to provide a simplified illustration of these components and their interactions.

In the systems considered here, generative AI models use patterns learned during training to generate outputs from the information they receive. A generative AI model can, for example, process instructions and produce text, propose a plan or request the use of a tool. Organising the sequence of work is the role of “orchestration”. A lead AI agent may plan the task and assign sub-tasks to specialised AI agents. In the travel example, separate AI agents could search for transport and accommodation. The same model need not perform every step: different models may be selected for successive requests according to the capabilities required (Anthropic, 2024[17]; Wang et al., 2024[18]).

The runtime software,1 sometimes termed a “harness”, supports this orchestration and manages how model outputs are translated into actions. It can restrict access to information and tools, impose spending limits and require human approval. For example, the system could be permitted to search for travel options but be prevented from making a payment without confirmation from the user. Some actions may also be confined to restricted execution environments known as “sandboxes”, which can limit access to tools, external systems and data. Such controls can be enforced through the surrounding software rather than relying solely on instructions to the model (AI Security Institute, 2025[19]; Liu, 2026[20]).

Tools and services provide the functions through which the system retrieves information or takes action. A hotel-search tool could return current availability, while a booking tool could submit a reservation and return a confirmation or an error. Connections often use application programming interfaces (APIs), which allow software to request information or actions from other software in a standardised way; systems may also interact through websites or application interfaces. Common protocols can support interoperability between components supplied by different providers. These connections may involve internal applications, external services or other AI agents, depending on technical compatibility and access permissions (Autorité de la concurrence, 2026[14]). The information returned must also be suitable for the task, including in its accuracy, currency and format. In the travel example, the system needs to identify whether a quoted price is per night or for the whole stay, which dates it covers and what cancellation conditions apply. Clearly labelled fields and consistent formats can help it interpret and compare offers, while information in documents or webpages may need to be extracted and checked before use.

As results return, the system needs to retain relevant information for subsequent steps. A language model can consider only a limited amount of information in each call, so the harness may select or summarise material for its current context. This could include the dates and itinerary under consideration, while a preference for refundable bookings could be stored separately as persistent memory for later retrieval. Reusable procedures, workflows or “skills” can supply standardised approaches for tasks, such as checking dates and cancellation terms before booking (Wang et al., 2024[18]). Returned results can then inform the next action: a failed reservation might prompt a revised search or a request for human input. The process can repeat until the objective is met, a limit is reached or human input is required.

While controls are widely used in agentic AI systems, their effectiveness remains an active area of research, practice and debate in AI safety and governance.2 Recent evaluations have found that advanced models, when tasked with escaping sandbox environments, can exploit deliberately introduced vulnerabilities and misconfigurations, illustrating challenges in containing capable systems (Marchand et al., 2026[21]). Security concerns may also affect the adoption of agentic AI systems and, potentially, competition between providers.

Agentic AI systems may also generate logs and records of actions taken, the requests made to tools and information retrieved. However, monitoring such systems remains challenging. Because model outputs are probabilistic and agent behaviours can emerge through sequences of interactions between models, tools and external services, it may be difficult to fully explain why a particular outcome occurred. Logging and observability tools can improve oversight, but they do not necessarily provide complete transparency into system behaviour or decision making processes (Microsoft, 2026[22]; Rao et al., 2026[23]). These limitations may complicate oversight by firms and investigations by regulators, although records of information flows, communications and actions can help reconstruct conduct without fully explaining the model’s internal operations. Importantly, difficulties in fully explaining model behaviour do not necessarily remove firms’ responsibility for outcomes.

Figure 2. Simplified illustrative structure of agentic AI components

Note: This figure provides a simplified illustration of selected components and interactions in agentic AI systems to support the competition analysis in this paper. Configurations vary across implementations and continue to evolve. The figure should not be interpreted as an OECD-endorsed definition or reference architecture for agentic AI systems.

Source: Based on Model AI Governance Framework for agentic AI IMDA (2026[24]), https://www.imda.gov.sg/-/media/imda/files/about/emerging-tech-and-research/artificial-intelligence/mgf-for-agentic-ai.pdf.

2.2. Emerging applications, supply arrangements and adoption

Agentic AI systems are being incorporated into general-purpose assistants, software-development tools, enterprise applications and sector-specific services. Some firms offer ready-to-use systems, while others supply platforms and components that developers and organisations can use to build or integrate their own systems. Practitioner interviews also identify early industrial trials involving manufacturing and supply‑chain optimisation, including co-ordination of logistics and production timing (OECD, 2026[13]).

The requirements of these systems vary according to who uses them, the task performed and the authority delegated. Organisations may use agentic AI systems for internal tasks or to interact with customers, while individuals may use them to act on their own behalf. Organisational deployments may require integration with internal data and workflows, differentiated staff permissions, records of actions and contractual arrangements governing performance and data handling. Services through which individuals delegate actions may place greater emphasis on accessible ways to express preferences, set limits and approve or challenge actions.

Commercial offerings and reported use have expanded rapidly, although the extent of autonomy remains uneven. Agentic capabilities are now offered in general-purpose assistants, software-development tools and enterprise software, while specialist suppliers provide systems for functions such as customer service and workflow automation. Survey evidence also points to growing workplace use: In an April 2026 Stack Overflow pulse survey of developers and working professionals, 59% of respondents reported using AI agents, compared with 31% in its 2025 survey, although 63% still rarely or never allowed agents to operate entirely autonomously (Stackoverflow, 2026[3]). A separate multi-country survey of business and IT leaders found that 74% expected their organisations to be using AI agents at least moderately by 2027, while only 21% reported mature governance arrangements for agentic AI (Bayiates, 2026[4]). These indicators suggest rapid experimentation and adoption alongside continued limits on delegation and gaps in governance.3

2.2.1. Supply arrangements and key inputs to agentic AI systems

Agentic AI systems build on the wider AI value chain and can combine inputs supplied at several different layers. These include generative AI models, computing resources, data, orchestration and application software, access to external tools and services, technical and integration expertise, and distribution to users. The underlying infrastructure is in turn supported by specialised chips, data centres, cloud services, telecommunications and energy (OECD, 2024[10]; 2025[11]). Their ability to pursue objectives and take actions can, however, add further dependencies. Agentic AI systems require software to orchestrate actions across models and tasks and may need connections to external tools, applications, data and services through which they retrieve information or act. Their performance can therefore depend on access to upstream AI inputs, as well as their ability to interact with complementary services and existing digital and organisational systems.

Figure 3 provides a simplified representation of the supply inputs. The availability of suitable alternatives may differ across these layers, and some inputs may become bottlenecks where access is limited or concentrated among a relatively small number of suppliers.

Figure 3. Illustrative high-level view of the AI Stack

Note: Simplified representation of supply functions, not defined markets or necessarily separate firms. Inputs may support several activities, model adaptation is not always required, and deployment arrangements may be combined.

Source: Based on OECD (2026[25]), Benefits of AI Openness, https://doi.org/10.1787/746e8c9a-en.

The combination and sourcing of these inputs vary according to the task and deployment. Many of the underlying inputs are also required for generative AI systems more generally. What is more specific to agentic AI is the need to combine these inputs with components that enable systems to plan and carry out actions, interact with tools and services, and operate across multiple steps.

Many individual users are likely to access agentic capabilities through services in which most underlying inputs are assembled by the provider. Developers and organisations building their own systems have a wider range of sourcing choices. They may access third-party generative AI models through APIs, deploy open-weight models on infrastructure they operate or rent, or combine these approaches. Organisations with sufficient technical capabilities may also combine open-weight models and open-source software with their own data, computing infrastructure and existing software systems.

The additional components and connections used to enable agentic capabilities may likewise be developed internally or obtained from specialist suppliers, while integration with existing systems and workflows may require additional technical expertise. Computing resources may be obtained through cloud services or infrastructure operated by the developer or user (OECD, 2025[11]). At the same time, AI-assisted and increasingly agentic software-development tools may reduce the time and expertise required to develop, test and integrate software components, potentially lowering the cost of assembling new applications. The ability to assemble systems in different ways does not necessarily mean that suitable alternatives are available for every input. Providers may still depend on particular suppliers of models, computing resources or connected services. Section 3 examines the competitive significance of these dependencies.

Data are another important input, both in developing and operating agentic AI systems. Their role can extend beyond information used to train or adapt the underlying model, including through techniques such as reinforcement learning.4 Agentic AI systems may require current, organisation or sector-specific information during operation, as well as access to external sources and services. A procurement system, for example, may combine supplier offers with inventory records and internal purchasing rules. Over time, systems may also accumulate histories, preferences, configurations or other information associated with their use. Restrictions on the ability to access, transfer and reuse such information with another provider may increase switching costs and weaken competition (OECD, 2024[10]; Autorité de la concurrence, 2026[14]).

The firms involved in the supply of agentic AI systems may participate at one or several vertical layers of this value chain, and the inputs and capabilities required to provide agentic AI systems may be brought together through different organisational and commercial arrangements. Firms may obtain inputs through contractual relationships with specialist suppliers or enter partnerships combining model access, computing capacity, software, integration expertise and distribution. Vertically integrated firms may instead provide several of these capabilities within the same corporate group or ecosystem. For example, integration into an existing productivity suite, browser or operating system can combine a route to users with access to applications and data used to carry out tasks. In platform-based ecosystems, the operator’s participation rules and technical interfaces shape how external providers can connect their services (OECD, 2021[26]).

Partnerships, commercial agreements and acquisitions provide different ways of bringing together the capabilities required to supply agentic AI systems. Agreements may cover model licensing, computing capacity, development and distribution, and may be combined with minority investments. Such arrangements can secure access to complementary resources without bringing the participating firms under common ownership. Acquisitions may instead bring previously separate capabilities within the same corporate group, while other transactions may combine technology licensing and recruitment of key personnel without acquiring an entire company (OECD, 2025[11]).5 These arrangements can overlap; their implications for access to inputs, distribution and competition are considered in Section 3.

The components combined in agentic AI systems may also differ in their degree of openness and interoperability. Models may be accessed through proprietary APIs or made available as downloadable open-weight models, while orchestration software, runtimes and tool connectors may be open-source or proprietary. APIs and communication protocols may follow open technical specifications while still requiring commercial access or provider approval. Access to software source code and access to model parameters should therefore be distinguished, and openness at one layer does not imply openness across the system as a whole (OECD, 2025[27]). Common communication protocols can nevertheless facilitate interaction between components supplied by different providers.6 Box 1 illustrates these distinctions.

Box 1. Openness and interoperability at different layers of an agentic AI system

Agentic AI systems combine components with different degrees of openness. The relevant distinctions concern both what is made available and the conditions under which others can inspect, use, modify and redistribute it. Openness should also be distinguished from interoperability: the ability of components supplied by different providers to work together.

• Models: Trained weights may be available for download, while other models are accessible only through provider-hosted services, including APIs or particular platforms. Downloadable models may also be offered through hosted services.

• Software: Open-source orchestration frameworks, agent runtimes and connectors can be inspected, modified and redistributed under their licences. Making code publicly visible is not equivalent to granting those rights. Moreover, open-source software may still depend on proprietary models, libraries or services.

• Tools, interfaces and protocols: Common specifications such as the Model Context Protocol (MCP), which provides a standard way for AI applications to connect to external tools and data sources, can enable independently developed components to communicate without requiring their underlying software or models to be open. However, technical compatibility does not itself grant access to a connected service. Authentication, permissions and provider access conditions may still apply, including where an open protocol is used.

• Data and memory: Users may have varying rights and practical abilities to export and reuse histories, preferences, workflows and configurations. Retaining these benefits when switching providers depends on whether the receiving system can use the transferred information. The relevant issue is authorised portability and reuse, rather than public disclosure of confidential information.

• Infrastructure and deployment: Open models and software still require computing resources and supporting tools. Implementations may depend on particular hardware, libraries or cloud services, and switching may require changes to code or infrastructure. Open availability therefore does not necessarily imply that a system can be deployed equally readily with alternative suppliers.

Governance is a separate dimension. Publication of a specification should be distinguished from the arrangements for proposing changes, approving revisions and maintaining implementations. For example, MCP’s governance documentation specifies contribution processes and decision making roles separately from the licences governing its code and specifications.

Note: Openness at one layer does not imply openness throughout the system. An agent may, for example, use open-source orchestration software while relying on a proprietary model API, or deploy an open-weight model while connecting to closed commercial data and payment services.

Source: OECD Secretariat, drawing on OECD (2025[27]), AI openness: A primer for policymakers, https://doi.org/10.1787/02f73362-en; OECD (2025[11]), Competition in artificial intelligence infrastructure, https://doi.org/10.1787/623d1874-en; Model Context Protocol documentation on authorisation and governance, https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization.

2.2.2. Business models and illustrative applications

The business models of agentic AI suppliers are still developing and are likely to differ across applications, both in how charges are calculated and in who pays. Enterprise and industrial deployments currently tend to involve direct payments from organisations through subscriptions,7 licences, usage-based charges, integration services or, in some cases, payments linked to defined outcomes. Consumer-facing services may also rely on subscriptions or bundling. Advertising, sponsored offers and merchant transaction fees have featured in some consumer-facing AI services, although providers’ approaches differ and continue to evolve. Their importance to agentic AI systems remains uncertain.8 Table 1 illustrates some of these differences.

Table 1. Illustrative applications and business models for agentic AI systems

Illustrative application	Main use and source of value	Illustrative supplier business models
Consumer-facing services and agentic commerce	Assistance with search, comparison and tasks; potentially negotiating and completing transactions on users’ behalf	Subscriptions or bundling; advertising and transaction-related payments in some services
Enterprise applications	Automation of internal workflows, customer interactions and software development	Licences, subscriptions, usage- or outcome-based charges
Industrial applications	Automation and optimisation of procurement, production and other operational processes	Licences, subscriptions, usage-based charges and integration or support services

Note: Examples are illustrative and not exhaustive. Applications and business models may overlap, and business models continue to develop.

Source: OECD Secretariat, drawing on Autorité de la concurrence (2026[14]), Opinion No. 26-A-05 on the Competitive Functioning of the Artificial Intelligence Agents Sector, https://www.autoritedelaconcurrence.fr/sites/default/files/commitments/2026-09/26a05_eng.pdf; CMA (2026[15]), Agentic AI and Consumers, https://www.gov.uk/government/publications/agentic-ai-and-consumers and OECD (2026[13]) Agentic AI in organisations: Early insights from practitioner interviews, https://doi.org/10.1787/1257a26f-en.

The cost implications of agentic AI differ depending on whether they are considered from the supplier or user perspective. Multi-step tasks may require repeated model calls, computing resources and access to external tools or services, so the variable cost of supplying agentic AI may remain material. While quality-adjusted model prices have fallen rapidly, agentic systems may use substantially more model computation (inference) per task, making the effect on the overall cost of providing a service uncertain (OECD, 2026[28]). For users, however, agentic AI systems may reduce the overall cost of completing a task by lowering search, labour, administrative and co-ordination costs, or reducing reliance on intermediaries. Whether these savings outweigh the costs of operating, integrating and supervising the system will vary across applications (OECD, 2025[2]).

Agentic commerce provides a useful illustration of how these emerging business models can interact with the technical characteristics of agentic AI systems.9 A system acting on behalf of a consumer may search and compare offers across suppliers, negotiate aspects of a transaction and, ultimately, place orders or make payments subject to the authority delegated by the user. In doing so, it may interact with product information, ordering and payment services and potentially with businesses that generate revenues for the agent provider.10 Reliability, user confidence, and arrangements for authentication, permissions, errors and disputes will affect the delegation of such tasks.

2.2.3. Further development and adoption

Further adoption will depend on whether agentic AI systems can perform useful tasks reliably, securely and at an acceptable overall cost, including the costs of human oversight. Resource requirements vary across tasks and system configurations, making future operating costs difficult to predict. The sensitivity of the information involved, and the consequences of errors may also influence which models, tools and deployment arrangements are suitable. Together with users’ willingness to delegate and applicable regulatory requirements, these factors will shape the range of uses and the extent of autonomy.

The pace and direction of technological development remain uncertain. Further progress may involve advances in algorithms and computing resources, as well as improvements in memory, learning and the ability to complete tasks reliably. Agentic AI systems may also increasingly be combined with robotics and other physical systems, extending their use beyond digital environments to tasks involving perception and action in the physical world.11 However, the development of useful applications does not necessarily depend on further advances in frontier capabilities. It may also reflect more effective ways of incorporating existing capabilities into products and services (Hobbs et al., 2026[29]).

Notes

← 1.  Runtime software provides the supporting functions and environment that a computer programme needs while it is running.

← 2.  While the technical aspects of agentic AI safeguards are not within the scope of the paper, the role of auditability and liability in the context of agentic AI conduct is discussed further below.

← 3.  These surveys capture respondents’ reported practices and expectations rather than independently observed deployment or oversight. Results may be affected by reporting bias. Differences in survey populations and methods also limit direct comparisons.

← 4.  Reinforcement learning (RL), in which systems learn from feedback or rewards, is already used in the training and adaptation of AI models. Its role in agentic AI may evolve as systems become increasingly able to learn from the outcomes of their actions and interactions. The extent to which deployed agentic systems will themselves learn and improve over time, rather than relying principally on pre-trained models, context and memory, remains uncertain, but further developments could enable increasingly adaptive or self-improving systems.

← 5.  Acquisitions across the AI value chain may also involve businesses operating platforms for developing, distributing and deploying open models and open-source software. In September 2026, NVIDIA announced an agreement to acquire Hugging Face, which operates such a platform (Nvidia, 2026[86]). Its regulatory filing described completion as subject to regulatory approvals and other closing conditions. This paper does not assess competition implications of any transaction but simply uses it as an example to illustrate the distinction between the openness of models and software and the ownership of the platforms through which they are distributed and deployed.

← 6.  Examples include the Model Context Protocol (MCP), which connects AI applications with tools and data, and the Agent2Agent Protocol (A2A), which supports communication between AI agents supplied by different providers or built using different frameworks. (Linux foundation, 2025[94]; Google Cloud, 2025[71])

← 7.  For the purposes of this paper, enterprise applications refer broadly to agentic AI used in internal business workflows, customer-facing processes and software development, while industrial applications refer to its use in production, logistics, maintenance and other operational processes. The categories are illustrative and may overlap.

← 8.  For example, Google announced Direct Offers in January 2026, a Google Ads pilot allowing advertisers to display sponsored offers in AI Mode in Search (Google, 2026[80]). OpenAI’s September 2025 launch of Instant Checkout provided for fees charged to merchants on completed purchases, with users required to confirm transactions (OpenAI, 2025[76]). In March 2026, it announced a shift towards product discovery and merchants’ own checkout experiences (OpenAI, 2026[79]; 2026[78]). By contrast, Anthropic stated in February 2026 that Claude would remain ad-free, identifying enterprise contracts and paid subscriptions as its revenue sources while also expressing an intention to support user-initiated purchases and bookings (Anthropic, 2026[81]). These examples illustrate differing and evolving commercial approaches, rather than an established revenue model for autonomous purchasing agents.

← 9.  Availability differs by service, jurisdiction and degree of autonomy. In its July 2026 opinion, the French Competition Authority reported that the transactional services it described were not yet available in France. OpenAI introduced Instant Checkout for purchases from US Etsy sellers in September 2025, with user confirmation and fees charged to merchants on completed purchases. In March 2026, it announced a shift towards product discovery and merchant-operated checkout experiences, citing limitations in the flexibility of the initial implementation. In June 2026, Worldline, ING and Mastercard reported a live agent-initiated payment in the Netherlands, also subject to explicit consumer approval (Autorité de la concurrence, 2026[14]; OpenAI, 2025[76]; Mastercard, 2026[77]; OpenAI, 2026[78]; OpenAI, 2026[79]).

← 10.  The Agent may also thereby gain access to confidential data (payment, health and other data) which raises different issues in relation to data protection beyond the scope of this paper.

← 11.  Robotics illustrates a possible direction for the development of agentic AI systems. Google DeepMind’s Gemini Robotics 2 combines high-level task planning and progress monitoring with models that control physical actions (Google Deepmind, 2026[83]). Hugging Face’s LeRobot ecosystem includes SmolVLA, which translates visual observations and language instructions into robotic actions and is available with model weights and training code (Hugging Face, 2025[84])). NVIDIA’s September 2026 Isaac ROS 5.0 release introduces workflows for AI-assisted robotics development. Its announcement also describes AgenticROS, a separate open-source project connecting AI agents to ROS-based robots, and NVIDIA’s contribution of a common data-handling interface to ROS (Nvidia, 2026[85])). These developments illustrate both the use of agentic AI systems to develop robotic applications and their integration into systems capable of physical action, as well as the extent to which major suppliers are operating at different levels of the value chain.