UK refund bots: your returns policy still counts

Governance Watch | Published | HaltState

Watch UK refund bots: your returns policy still counts on YouTube

The customer asks on day 45. Your shop promises 60-day returns. The bot says the window closed at 30.

That is a hypothetical, not a reported incident. It assumes a shop has offered an extended returns policy; eligibility still depends on the applicable policy terms and facts. But the underlying failure is one the Competition and Markets Authority expressly warns businesses to look for: an AI refund agent overlooking an extended returns policy.

What changed on 9 March 2026?

The CMA published guidance explaining how existing consumer law applies when businesses use AI agents. It is not a new UK AI Act. The regulator says businesses remain responsible for what their agents do, including when a third party supplies the AI. Using a bot does not remove duties owed to customers.

This matters to teams using agents for customer service, refunds, product recommendations or marketing. The useful question is not whether the bot sounds polite. It is whether its answers reflect the rights and terms that actually apply.

A 30-day rule is not a universal answer

Keep these questions separate:

The CMA's refund example tells businesses to account for both statutory rights and contractual terms, check exchanges and time limits, and quickly refine prompts or workflows when the AI overlooks an extended policy.

Tell customers what matters; check what the bot does

The guidance says to tell customers they are dealing with AI if that fact might affect their decisions. It does not state a universal requirement to label every AI use regardless of context.

The Digital Markets, Competition and Consumers Act 2024 prohibits unfair commercial practices. Sections 225 and 226 address misleading actions, including the relevant average-consumer and transactional-decision test. This article does not decide that a particular refund response breaches the Act.

Operationally, the CMA recommends testing before deployment, regular human oversight, monitoring outputs and acting quickly on problems. A disclaimer saying the answer came from AI is not a substitute for that work.

Security guidance is a separate layer

The National Cyber Security Centre's agentic-AI guidance recommends safeguards, sandboxing and active oversight. For higher-risk scenarios it recommends human oversight alongside technically enforced controls. These are security recommendations, not consumer legislation and not a guarantee that a deployment is safe.

What the isolated HaltState test actually showed

We evaluated a synthetic refund.deny proposal using the real policy engine from an immutable source archive. The in-memory policy matched authored test fields: region GB, review_required true, proposed_outcome deny and an allowed support role. It returned APPROVAL_REQUIRED.

The test then created an unsigned hash-only Proof Pack and independently recomputed its content digest. This records an isolated policy evaluation, not a customer transaction. The review flag was supplied by the test author: HaltState did not infer a legal entitlement or decide that the customer qualified for a refund.

No customer message was sent, no refund was issued and no human review was completed. The test did not exercise a production enforcement adapter, persistent storage or signing. A verified digest is not signature authenticity or legal compliance certification.

Before the next automated refusal

Check the policy version the bot uses against the current contract and statutory rights. Identify who reviews disputed decisions, how they get context and how mistakes are corrected. Test the specific workflow before allowing it to affect customers.

Start with one question: which rule is your bot actually using?

Official sources

General information, not legal advice. Obligations and remedies depend on the facts, contract, applicable law and exceptions. Hypothetical scenario and isolated product test are separately labelled. AI-generated narration and illustrative footage; authoritative text and test results are deterministic.