AI Agent Governance UK: A Consumer Law Control Room

The Rule Has Not Changed: The Business Owns the Outcome

As of 31 July 2026, the UK does not have a separate consumer-law exemption for decisions made by an AI agent. The Competition and Markets Authority’s March 2026 guidance on using AI agents is direct: the same consumer rules apply whether a business deals with customers through people or automation, and the business is ultimately responsible if its agent acts unlawfully.

That matters because an agent can do more than draft text. It may rank products, recommend a deal, refuse a refund, change a booking, place an order or cancel a subscription. Each action can affect a customer’s rights or economic choices.

The CMA’s accompanying research on agentic AI and consumers identifies the central risk: an agent may optimise for conversion, engagement or commercial value rather than the customer’s actual interests. Greater autonomy can magnify errors, hidden incentives, manipulation, bias and loss of consumer control.

The compliance question is therefore not simply, “Is the model accurate?” It is: can the business show that the agent was authorised, used the right information, respected the customer’s rights and could be stopped or corrected?

This article is operational guidance, not legal advice. Businesses should confirm the legal position for their products, customers and sector.

Why Consumer Law Belongs in the System Design

The unfair-commercial-practices provisions of the Digital Markets, Competition and Consumers Act 2024 apply to commercial practices from 6 April 2025. Official CMA guidance on unfair commercial practices explains that businesses must not mislead consumers, omit information they need, behave aggressively or fall below professional diligence.

Those rules reach the whole customer journey, including:

• product descriptions and recommendations;
• prices and unavoidable charges;
• rankings, comparisons and commercial affiliations;
• cancellation and refund information;
• marketing claims and endorsements;
• complaint handling and after-sales care.

The CMA can now determine certain consumer-law infringements directly. Its direct-enforcement guide states that a penalty can reach 10% of worldwide turnover or £300,000, whichever is greater.

An agent must therefore operate within the same legal and contractual rules as a trained employee. A prompt saying “be helpful and fair” is not an adequate control.

Disclosure: Use the Correct UK Test

A blanket statement that every automated interaction must always be labelled would overstate current UK consumer law.

The CMA says businesses should consider disclosure where a failure to explain that a customer is dealing with AI could mislead them or affect their decision. It also warns businesses not to exaggerate either the human role or the AI’s capabilities.

A defensible disclosure answers four questions:

• Could a reasonable customer believe they are dealing with a person?
• Would knowing that automation is involved affect what they disclose or decide?
• Is the agent making recommendations or taking consequential actions?
• Are its limitations, coverage or commercial incentives material to the customer?

Where the answer indicates a risk of deception, identify the agent clearly and early. Avoid names, avatars or typing behaviours designed to create a false impression of a human adviser.

Other laws may create separate transparency duties. For example, data-protection rules can apply when personal data is used in significant solely automated decisions. The disclosure analysis should therefore consider the complete legal stack rather than consumer law alone.

Start With a Decision-Rights Map

A useful control room begins with authority, not with a dashboard. For each tool and action, decide how much control the agent receives.

Where an action also makes a significant decision about a person, map these authority levels to the safeguards in our UK automated decision-making guide.

These levels are an engineering pattern, not four statutory categories. Their value is that they turn a vague promise of “human oversight” into testable rules.

Limits should cover:

• maximum payment, credit or refund value;
• permitted products, regions and customer types;
• approved data sources;
• actions requiring fresh customer confirmation;
• situations involving vulnerability or distress;
• disputed facts or missing records;
• legal rights that the agent must never contract out of;
• actions that cannot be reversed safely.

Keep the map in version-controlled configuration or policy-as-code. Record who approved each rule and when it changed.

What the Control Room Must Show

A control room should help operators understand and repair real customer outcomes. Six views are enough for most implementations.

The decision record should be understandable without reconstructing the case from raw model traces. Preserve the facts that mattered: source records, relevant policy, tool calls, customer confirmation, final system-of-record state and any subsequent reversal.

Logs containing personal data need access controls, retention limits and a legitimate purpose. “Log everything forever” is not a safe governance strategy.

For a reusable claim-and-evidence structure around those records, see the AI assurance evidence-pack guide.

Translate Consumer Rights Into Executable Rules

The CMA gives practical examples involving refunds, marketing, support and comparison services. Those examples should become regression tests.

A refund agent, for example, should not rely only on a generic 30-day rule. It must consider whether goods are faulty under the Consumer Rights Act 2015, whether cancellation rights apply under the Consumer Contracts Regulations 2013, and whether the business promised an extended return period.

If a third-party vendor supplies the model or agent, the responsibility does not disappear. The CMA warns that a business may still break the law if it fails to check whether a supplier has trained and controlled the system appropriately.

Test the Boundaries That Affect Customers

Average-case demonstrations are weak evidence. Build tests from disputed and inconvenient situations:

• the policy changed yesterday;
• the source systems disagree;
• the customer is just outside a threshold;
• a product description omits a limitation;
• a discount excludes an unavoidable charge;
• an extended-return promise is more generous than the statutory minimum;
• the customer corrects a material fact;
• the agent has an undisclosed commercial relationship;
• repeated failure indicates distress or vulnerability;
• the requested action is irreversible.

Test both decisions and behaviour. An agent may reach the right answer while presenting it in a misleading, pressuring or inaccessible way.

Useful operating measures include:

No single target is universally correct. Establish thresholds by action type and customer impact, then investigate changes rather than rewarding the agent merely for completing more tasks.

Example: When a Refund Agent Gets It Wrong

Consider an agent that refuses a return because the purchase is 35 days old.

A compliant workflow should ask:

• Is the product faulty?
• Was it bought at a distance?
• When was delivery completed?
• Has the business promised a longer return period?
• Is the customer asking for cancellation, repair, replacement or refund?
• Are the facts complete and consistent?
• Is the requested action inside the agent’s authority?

If the company advertises a 60-day return period, the agent must apply it. If the product is faulty, a generic change-of-mind policy is not enough. If records conflict, the agent should stop rather than invent certainty.

The control room should then show the policy used, evidence consulted, reason for refusal, customer correction and final human decision. If the same error appears elsewhere, the team can identify the affected cohort and repair it rather than waiting for each customer to complain.

Build Reversal and Redress Before Launch

Every write-capable tool should have a compensating process.

Before activation, document:

• how an action can be cancelled or reversed;
• who has authority to approve the correction;
• how the customer will be informed;
• which downstream systems must be reconciled;
• how related cases will be found;
• when the agent must be paused globally.

For consequential actions, use staged execution:

• prepare the proposed action;
• validate it against current policy and source data;
• obtain confirmation where required;
• execute once with an idempotency key;
• verify the system of record;
• record the outcome and reversal route.

A “human in the loop” is not meaningful if staff lack context, authority or time to change the result.

A 30-Day Implementation Checklist

Week 1: Scope

• Inventory customer-facing agents and tools.
• Name a business owner for every agent.
• Map statutory, contractual and sector rules.
• Classify actions by authority and impact.

Week 2: Controls

• Implement tool allowlists and transaction limits.
• Add customer disclosure where non-disclosure could mislead.
• Create stop conditions and escalation queues.
• Design reversal paths for every write action.

Week 3: Evidence and testing

• Build decision records around sources and policy versions.
• Test refunds, pricing, rankings, complaints and vulnerability cases.
• Verify writes against systems of record.
• Run tests after common model, prompt and policy changes.

Week 4: Live governance

• Set alert owners and response targets.
• Review complaints, reversals and repeated failures.
• Train staff to challenge agent outputs.
• Schedule a formal post-launch review.

Frequently Asked Questions

Who is liable if an AI agent breaks UK consumer law?

The business remains responsible for how it deals with customers. Buying the agent from a supplier or delegating operations does not transfer that obligation.

Must a UK business always tell customers that they are using AI?

The CMA’s consumer-law guidance uses a contextual test: disclose the agent where failing to do so could mislead customers or affect their decisions. Other laws may impose additional duties.

Is regular sampling enough human oversight?

Sampling is useful for monitoring, but it does not replace case-level intervention where a customer needs correction or applicable law requires human review.

What is the minimum evidence to retain?

Keep enough to reproduce the outcome: authority rule, material source data, policy version, action, customer confirmation where applicable, final system state and any reversal. Apply appropriate security and retention controls.

The best agent control room is not a wall of charts. It is a compact operating model that makes authority, evidence, exceptions and redress visible before automation is allowed to affect customers.