Copy article

CMA is right to worry about AI shopping agents. The risk is hidden steering, not bad answers

ended 12. March 2026

The next wave of consumer AI will not be chatbots. It will be agents that shop, switch providers, cancel subscriptions, and negotiate deals.

That sounds like convenience. It also creates a new kind of dark pattern: automated steering that looks like help. An agent can nudge a consumer towards a pricier tariff, a longer contract, or a partner product, then explain it away as a "best match". The user hears confidence, not caveats.

The hard part is that harm will not look like a dramatic failure. It will look like slightly worse choices, repeated millions of times, with no easy way to compare the counterfactual. Over time, competition becomes theatre because the "comparison" is happening inside a black box.

UK guidance on using AI agents under consumer law is a useful signal, but enforcement needs something more practical than policy promises. If an agent is acting for a consumer, it should have to prove three things: what it was optimising for, what options it considered, and who paid for any influence.

We'd like your views:

  • Should AI agents be required to disclose commercial incentives in plain English before acting?
  • What is the minimum audit trail for an AI agent choosing products or switching services?
  • Should regulators treat hidden steering by agents as mis-selling, even if the output is "accurate"?
  • Who is liable when an agent makes a harmful choice: the developer, the deployer, or the retailer?
  • How should consumers challenge an agent decision when the reasoning is proprietary?

1 responses from the Newspage community

Copy all

Star Quote
Copy

Rohit Parmar-Mistry
Founder at Pattrn Data
AI shopping agents will not scam you with obviously wrong answers. They will quietly steer you with perfectly plausible ones. If an agent is paid to 'optimise', the real question is who it is optimising for: the consumer, the platform, or the partner with the commission.

The CMA is right to frame this as consumer law, not a tech debate. If an agent is comparing tariffs or switching services, it should have to show a basic audit trail: what it searched, what it excluded, any commercial incentives, and what the user actually consented to (CMA 'Complying with consumer law when using AI agents', GOV.UK). Otherwise mis-selling just gets a new UI.

In our AI audits, the repeat failure is governance theatre. Firms ship the 'helpful' agent but cannot answer a complaint with evidence. Regulators should treat hidden steering as a breach even when the output is 'accurate', because accuracy is not fairness.