Competition and agentic AI

Executive summary

Agentic AI systems extend the range of decisions and actions that consumers and businesses can delegate. They combine generative AI models with software, information and tools to carry out linked tasks autonomously, adapting to act with varying degrees of human oversight. This paper examines competition in the supply of these systems and the effects of their use in other markets. It identifies mechanisms and conditions that may strengthen or weaken competition, recognising that outcomes will depend on how the technology develops, how widely it is adopted and which decisions users choose to delegate.

Key findings

• Competition in the supply of agentic AI systems depends on more than the capabilities of underlying models. Access to existing generative AI models and software allows new suppliers to build systems without developing the models themselves. However, their ability to compete also relies on having access to these models and other key inputs, such as compute, data, and distribution. Integrated offerings and partnerships can bring complementary capabilities together, but control over important inputs, interfaces or distribution channels may also enable suppliers to disadvantage rivals.

• Personalisation and integration can improve services while making switching between agentic AI systems more difficult. Systems may accumulate data that help them serve a particular user or organisation. Where this cannot readily be transferred or reused, changing providers may involve losing functionality or rebuilding integrations. Feedback effects may also reinforce established positions, although their importance will differ across applications.

• The use of agentic AI systems could strengthen competition in other markets by reducing costs and making consumers more effective participants in markets. Consumers may be able to search more widely, compare complex offers, or negotiate better alternatives, increasing pressure on suppliers to improve price and quality. Businesses may also automate linked administrative and operational tasks, reducing the resources required to enter or expand. These gains depend on access to sufficiently capable and connected systems, and on whether savings exceed the costs of integration, operation and oversight.

• Delegation could, however, also weaken competition if it exhibits biased selection or restrictions on suppliers’ access to customers. Where providers receive payments from sellers, or have a vested interest in steering purchasing decisions, their incentives may differ from the objectives of users. Inaccurate comparisons, systematic biases or manipulation may also divert demand from offers that better match users’ preferences, affecting suppliers’ incentives to compete. Agentic AI systems could reduce dependence on existing intermediaries but may themselves become important routes to customers.

• Agentic AI systems may make collusion between firms easier but also create compliance opportunities. Systems may implement human agreements, align competitors’ decisions through common providers or sustain co-ordination autonomously if delegated pricing decisions. Controls on information sharing and actions, human approval requirements and activity records may support compliance, although their effectiveness in preventing collusion remains uncertain.

• Competition authorities are likely to be closely monitoring developments in agentic AI and have tools available to act if needed. Monitoring, market studies and engagement with suppliers can help distinguish emerging risks from hypothetical concerns, with timely enforcement available where needed, including the use of interim measures. Domestic and international co-operation can support these efforts, including where competition issues intersect with other policy areas.