Key takeaways
|
Pricing agents under enforcement spotlight
In March 2026, the CMA stated that businesses are responsible for an AI agent’s conduct in the same way as they are for an employee, including where the agent was designed or supplied by a third party. Businesses should therefore understand the law and how the technical system works, set clear boundaries, monitor outputs and intervene quickly if the agent risks distorting competition, for example through price coordination or exchanges of competitively sensitive information.
Risk is highest where an AI agent can recommend or execute decisions affecting core parameters of competition: price, output, quality, choice, innovation, access, ranking, discounts or customer allocation. Enforcement to date has focused mainly on pricing and price recommendations, but competition law concerns can arise wherever an agent reduces uncertainty between competitors, deploys a foreclosure strategy or reinforces an existing dominant position.
Key competition law risks and recent enforcement trends
In its recent blog on AI and collusion, the CMA identified three main risks involving AI systems and pricing:
‘Classic’ collusion – businesses agree to collude and use software to implement or monitor a cartel. In the UK, the CMA fined Trod GBP 163,371 and pursued director disqualification proceedings after finding that Trod and GB eye had agreed not to undercut each other on poster and frame sales on Amazon Marketplace. In the US, the DOJ brought criminal enforcement action against David Topkins, who was fined USD 20,000 after allegedly agreeing with competitors to adopt pricing algorithms and instructing algorithm-based software to set prices in line with the conspiracy. In both cases, the illegal agreement was made by people, but automated pricing software helped to implement the conduct by monitoring and aligning prices.
Hub-and-spoke collusion – businesses may use the same algorithm or data hub to exchange competitively sensitive information indirectly. This may involve delegating pricing decisions to the hub or receiving pricing or other conduct recommendations from it based on co-mingled data. In the US, rental-pricing software was alleged to have enabled competing landlords to share non-public, competitively sensitive rental data through a common platform, which then generated rent recommendations and allegedly aligned pricing across competing properties. The case was settled with RealPage giving a number of undertakings on future conduct. The case illustrates that agentic or algorithmic pricing tools can create hub-and-spoke coordination concerns even where the tool is operated by a third-party software provider. Similarly, in May this year, the Danish competition authority issued a formal warning to an AI accounting assistant that benchmarked users’ fees. In its findings, the authority cited an example of the agent suggesting that carpenters in Copenhagen raise prices where they charged below the local average.
Competition law risks in predictable-agent and autonomous AI systems
Businesses may use algorithms that react predictably to market events, potentially softening competition. Such algorithms may follow price leadership and punish deviations, achieving collusive outcomes without human communication or explicit agreement. Similarly, an AI system tasked with maximising profits may learn to reach coordinated outcomes that align or increase prices, even without any human intent to collude.
As agentic AI increasingly makes autonomous pricing decisions, competition authorities are increasingly focused on whether it was foreseeable that AI systems could produce anti-competitive effects, such as the price collusion and information exchange discussed above. Singapore’s Competition and Consumer Commission takes a compliance-by-design approach and has launched an AI Markets Toolkit, which references the CMA’s Foundation Model report, to help businesses check whether their models are, by design, compliant with Singaporean competition and consumer laws. The CMA does not have a similar tool, but it has issued helpful guidance and blogs to help businesses navigate compliance by design and monitoring.
What should businesses do? Compliance safeguards for the AI lifecycle from design to deployment
Similar to compliance programmes for employees, businesses should build competition law into the AI lifecycle from the outset. As mentioned in our latest article in the series, given the speed of AI systems and how quickly decisions are made, it is more important than ever to ensure the system is competition and consumer law compliant by design and that these risks are stringently tested and documented before deployment. In practice, this means:
Compliance by design: define what the AI system is allowed to optimise for and explicitly prohibit objectives that reward collusion, price alignment or avoiding competitive undercutting.
Define permission boundaries: decide whether the system can recommend, approve or execute decisions, and apply stricter controls where it can influence pricing, discounts, access, ranking or customer allocation.
Human oversight for pricing decisions: require appropriate human oversight, especially where the system recommends price increases or other market-wide commercial changes.
Stress test for competition risks: use linguistic stress-testing prompts and other pre-deployment testing to check whether the system suggests collusion, information exchange, price alignment or suppression of discounts.
Build in anti-collusion constraints: include explicit constraints in prompts, system instructions and model governance, particularly for pricing tools and LLM-enabled pricing solutions.
Maintain audit trails and escalation routes: document prompts, data inputs, model changes, approvals and interventions, and escalate unusual price recommendations or evidence of market alignment quickly.
/Passle/5f3d6e345354880e28b1fb63/MediaLibrary/Images/2025-09-29-13-48-10-128-68da8e1af6347a2c4b96de4e.png)
/Passle/5f3d6e345354880e28b1fb63/SearchServiceImages/2026-08-10-14-02-25-111-6a79d9f1f45ff5f0bdb36e63.jpg)
/Passle/5f3d6e345354880e28b1fb63/SearchServiceImages/2026-08-10-11-01-28-507-6a79af88f45ff5f0bdb2d619.jpg)
/Passle/5f3d6e345354880e28b1fb63/MediaLibrary/Images/2024-08-23-11-31-07-354-66c872fb971eecc249d83d40.png)