Competition law and agentic AI: risks arising from shopping and pricing agents

What is agentic AI?

Agentic AI promises to reshape how businesses interact with consumers and compete with each other.

Unlike first generation AI systems, agentic AI systems are able to autonomously assess goals and plan end-to-end workflows, access data from other agents, databases and other services, execute actions autonomously and store memory of past interactions to improve output over time.

Agentic AI is fundamentally changing how businesses compete for consumer attention and how consumers make purchasing decisions. This has implications for how businesses engage with consumers and may also change how businesses make core strategic decisions, including in relation to pricing and terms.

As AI agents take on a more active role in searching, selecting, pricing and transacting, businesses may need to consider competition law risks arising not only from their own conduct, but from the decisions and interactions of the agentic products they deploy.

Agentic shopping

One of the most significant developments in AI is the emergence of agentic shopping, where AI agents autonomously or semi-autonomously search for products, compare offers and make purchases on behalf of consumers.

While agentic shopping will benefit consumers by increasing convenience and reducing search costs, it could potentially raise a number of competition law issues.

Self-preferencing and transparency risks

First, agentic shopping agents can potentially give rise to self-preferencing risks, where a vertically integrated agent makes purchasing decisions in favour of its own products compared to the products of competitors. Such risks may be exacerbated in circumstances where:

• high levels of concentration develop in the markets in which agentic systems operate, or in the underlying AI models; and
• there is limited ability to identify the reasoning underpinning why agentic systems make purchasing decisions.

A new basis for competition

Second, agentic shopping will change the basis on which firms compete and potentially increase incumbency advantages and raise barriers to entry.

In an agentic environment, to acquire consumers, businesses will need to optimise for algorithmic visibility and favourable ranking within AI systems. In such an environment, traditional forms of marketing and brand recognition may be less relevant. However, access to data used by agentic shopping products may create a barrier to new suppliers seeking to expand their consumer reach.

Barriers to entry and expansion may also be increased if agents, or the underlying models used by the agents, have in-built brand preferences, making it more difficult for new and expanding brands to win market share. Preferences may arise inadvertently due to biases in training data, or intentionally as a result of commercial arrangements between the suppliers of products and operators of agents.

Consumer law issues may also arise in relation to representations by suppliers of shopping agents regarding the process by which shopping agents make purchasing decisions. For instance, operators of shopping agents may represent to consumers that the agent will exercise certain preferences in making purchasing decisions. For instance, representations that the agent will select "healthy" or "good value" products. Operators of shopping agents will need to develop adequate safeguards to ensure that such representations are not false or misleading.

Collusion risks with agentic pricing products

The use of pricing algorithms, including those with some autonomous AI features, is already common and can facilitate more efficient pricing outcomes, benefiting both consumers and businesses.

However, pricing algorithms can lead to coordinated pricing between competitors. Both the UK Competition and Market Authority (CMA) and Australian AI Safety Institute (AI Safety Institute) have published papers regarding competition issues that may arise in the context of agentic pricing services.

How AI can facilitate collusion

The CMA has identified that agentic collusion risks can arise in a number of ways including:

• the use of pricing algorithms and tools to monitor and enforce anti-competitive agreements that have been established between competitors (i.e. agentic implementation or enforcement of an anti-competitive agreement);
• "hub-and-spoke" algorithmic collusion, where a single algorithm or product is used to share information and/or set prices for multiple suppliers, rather than each supplier setting its prices independently of others.

The US Department of Justice has engaged in enforcement action in respect of allegations of "hub-and-spoke" algorithmic collusion. In 2024, the DOJ commenced proceedings against RealPage alleging that its apartment pricing software facilitated unlawful coordination among competing landlords in breach of the Sherman Act. The DOJ alleged that RealPage collected non-public, competitively sensitive rental and lease data from rival landlords and used that information to generate pricing recommendations that aligned rents and reduced normal competitive rivalry. The case settled in 2025, with RealPage agreeing to amend the way in which its software used rival data in setting rental prices.

The CMA has also identified a number of areas where algorithmic or agentic pricing tools may give rise to the risk of coordinated pricing, without the need for human communication or explicit agreement. Examples include:

• "predictable agents", where algorithms react predictably to market events, the use of such agents will remove the uncertainty of competition as businesses and other agents will be able to anticipate the pricing decisions of such "predictable agents"; and
• advanced AI systems, including where agentic AI "may learn to reach coordinated outcomes, even without human intent to collude".

The CMA considers that businesses "remain responsible for the outcomes of pricing and commercial decisions shaped by AI systems, and must take proactive steps to understand, test and govern the technologies they deploy".

Explicit, tacit and covert coordination

In respect of agentic collusion in advanced AI systems, the AI Safety Institute has described different types of collusive conduct that may arise where multiple organisations deploy agents into a shared environment in which multiple agents can interact.

In respect of agentic collusion in advanced AI systems, the AI Safety Institute has described different types of collusive conduct that may arise where multiple organisations deploy agents into a shared environment in which multiple agents can interact.

In this environment, agents of different organisations may be able to engage in algorithm collusion by developing a coordinated strategy to further their individual agent-level goals. The AI Safety Institute identified three different variants of algorithmic collusion:

• explicit collusion, where agents communicate openly about their strategy and discuss how to maximise common outcomes.
• tacit collusion, where agents adapt by observing each other's actions, such as pricing and bid outcomes, and react accordingly (collusion between "predictable agents", identified by the CMA, may be a form of tacit collusion).
• steganographic collusion, where agents communicate covertly through channels to evade detection, using signals or codes to convey certain information.

The AI Safety Institute has suggested a range of potential controls to protect against algorithmic collusion. These controls include restrictions on the design of agents, including restrictions on agent communications, information-sharing and decision-making processes, as well as monitoring of patterns of pricing collusion, and mandated testing of agents.

Can Australia's competition laws deal with agentic collusion?

Section 84 of the Competition and Consumer Act 2010 (Cth) states that conduct engaged in by an agent is deemed to be conduct of the principal where the agent is acting in the scope of its actual or apparent authority.

However, genuine challenges arise in applying Australia's competition law prohibitions to such agentic collusion. For instance, the foundation of Australia's cartel laws is the establishment of a contract, arrangement or understanding between competitors.

It is yet to be tested how Australia's cartel laws, or other prohibitions requiring the establishment of a concerted practice, will apply where businesses are deploying agentic AI tools, potentially without the ability to understand the precise workings of such tools. Additional complexity arises where the agentic AI pricing tools remove the requirement for explicit communication or coordination between businesses.

Businesses should closely monitor the regulatory environment in which agentic pricing agents operate and ensure that products being adopted or developed are compliant with current regulatory settings. Where businesses are operating across jurisdictions, a multi-jurisdictional compliance approach will be required.

What should businesses do when deploying AI shopping or pricing agents?

• Take steps to understand how the agent will operate, including the data that it can access and share and how the agent will interact with other agents and online environments?
• Identify potential competition law risk areas and develop mitigants and strategies to address these risks
• Ensure that adequate controls are in place to monitor for ongoing compliance with Australia's competition laws