Field note · CMA agentic AI framework · May 2026

You can buy the agent.
You cannot outsource the liability.

On 9 March 2026, the UK Competition and Markets Authority published the first guidance from any major consumer-protection authority anywhere in the world specifically addressing AI agents in consumer-facing roles. It is not new law. It is the Consumer Rights Act 2015, the Consumer Protection from Unfair Trading Regulations 2008, and the DMCCA 2024 applied to a new operational context. The deployer was already on the hook. The guidance made it impossible to pretend otherwise.

Run the self-test Jump to research
You are responsible for what an AI agent does in the same way you are responsible for what an employee does. This is true even if someone else designed or provides the AI agent on your behalf.
CMA · gov.uk · 9 March 2026
§ Orientation · The 30-second self-test

Three things to know before anything else.

Most coverage of the CMA's framework has framed it as "rules for AI agents." It isn't. It is a pre-existing standard for what businesses owe their customers, applied to a new way of meeting them.

§ Yes · The guidance is real and in force

Published 9 March 2026 on gov.uk. Five action sections, four worked examples, enforced under existing UK consumer law and the Digital Markets, Competition and Consumers Act 2024. Not a consultation. Not draft. Already operative.

§ But · It isn't new law

The CMA didn't legislate. It applied the Consumer Rights Act 2015, the Consumer Protection from Unfair Trading Regulations 2008, the Consumer Contracts (Information, Cancellation and Additional Charges) Regulations 2013, and the DMCCA. Every obligation in the guidance was binding before the guidance existed.

§ And · The deployer carries everything

Not the model provider. Not the agent platform. The business deploying the agent is responsible for what it does "in the same way you are responsible for what an employee does" — even where a third party designed or supplied it. Up to 10% of worldwide turnover. Imposed by the CMA directly, no court required.

§ 01 · Anatomy

Inputs from many hands. Liability in one.

An agent product is built from at least three contributions — the model, the runtime platform, and the deployer's own configuration. Each of those can be a different organisation. Under the CMA's framework, the legal exposure ignores that diversity entirely. It concentrates on the deployer.

§ Input · Third party Model provider "We supplied the AI." § Input · Third party Agent platform "We supplied the runtime." § Input · You Your configuration Prompts, data, guardrails, deployment context. § Single point of legal exposure The deployer Responsible for what the agent does "in the same way you are responsible for what an employee does" — CMA, 9 March 2026 Up to 10% of global turnover § Output The consumer Refund. Marketing. Query. Recommendation. Service.
§ The two-document structure
The CMA published two things on 9 March 2026, not one. Complying with consumer law when using AI agents is the operational guidance — five action sections plus four worked examples. The companion policy paper covers the wider terrain: dark patterns at scale, agentic collusion, vulnerable consumers, lock-in, the "many hands" accountability problem. Most coverage treats them as one document. They aren't. The guidance is what you do tomorrow. The policy paper is what you watch for over the next two years.
§ 02 · The four operational requirements

What the guidance actually tells you to do.

The gov.uk page reads as a calm five-section walkthrough — one framing principle ("the same rules apply") plus four action sections. Each is short. None is optional. They are reproduced and expanded below.

§ Requirement 01
Tell your customers.
If the fact that a customer is dealing with AI rather than a person might affect their decisions, label it. Do not overstate what the agent can or cannot do. Build trust by being open about how the agent is being used.
"Being clear and open about how you use AI agents is a good way to build trust, especially if it might be a surprise."
§ Requirement 02
Train for compliance.
Configure the agent with the relevant statutory and contractual obligations baked in — refund rights, cancellation rights, pricing transparency, no misleading actions or omissions, lawful consent. Test rigorously before deployment (A/B testing, unit testing are both named).
"If someone else is providing the AI, and you don't check that they have done so, you may break the law."
§ Requirement 03
Monitor performance.
A human must be in the loop, actively checking the agent's decisions and outputs. Models can hallucinate — the CMA's word. Regular human oversight is required to catch mistakes and ensure outputs stay legally compliant. Document the process.
"Make sure there is a human in the loop, actively checking that the AI agent is making correct decisions."
§ Requirement 04
Refine quickly.
When a problem is detected, act immediately — especially where the agent interacts with large numbers of people or vulnerable customers. Forward-looking remediation does not cure prior breaches. Speed is itself a compliance variable.
"If you do not act quickly to address problems, you may end up breaking the law."
§ Peer-reviewed: principal-agent theory underneath
Kolt (2025, Notre Dame Law Review Vol. 101 forthcoming) anchors the deployer's exposure in the common-law doctrine of agency and the economics of principal-agent problems. Both frameworks make the same prediction: where one party (the principal) relies on another (the agent) to act on their behalf, the principal's information asymmetry over what the agent actually does is what creates the legal exposure. The CMA's four requirements are not new policy choices — they are the standard remedies for principal-agent problems, written for software.
§ 04 · The four use cases

Where the requirements actually bite.

The guidance includes four worked examples. The four scenarios were not chosen at random — each maps onto a specific cluster of consumer-protection obligations and a specific class of common operational failure.

01
§ Use case · DMCCA + advertising rules
An AI agent running your marketing campaign.
Marketing materials must give accurate price information including all unavoidable charges. Paid endorsements must be properly labelled. Offers and price reductions must be genuine. Someone with appropriate experience must review the agent's output regularly.
§ Where it bites
The CMA names this scenario specifically: the AI agent is not consistently requiring social media influencers to label their endorsements. Influencer-side disclosure compliance gets pushed onto the deployer through their agent.
02
§ Use case · CRA 2015 + CCAR Regs 2013
An AI agent processing refund requests.
Train the agent against statutory rights under the CRA 2015 (faulty goods) and the 14-day cancellation right under the CCAR Regs 2013 (change of mind). Layer in your own contractual terms — for example, extended returns policies — and ensure the agent doesn't fall back on a narrower interpretation.
§ Where it bites
The CMA's example: the AI agent is not always taking into account your extended returns policy. The deployer's own consumer-friendly term, ignored by the agent, becomes the deployer's own breach.
03
§ Use case · CPRs 2008 · misleading omissions
An AI agent answering customer service queries.
Respond accurately to queries about prices, products and rights. Give consumers all the information they need to take informed decisions. Do not make it difficult for them to exercise their rights.
§ Where it bites
The CMA's example: the AI agent keeps giving people the wrong information about the features of a product or their cancellation rights. Hallucinations and incomplete answers are both treated as misleading omissions under the CPRs.
04
§ Use case · Ranking transparency + market coverage
An AI agent providing a deal-comparison service.
Results must be accurate. Important information must be clearly and prominently disclosed — including how much of the market has been looked at, what data has been searched, and any particular limitations. Coverage, ranking, and supplier links must all be transparently disclosed.
§ Where it bites
The CMA's example: the AI agent keeps including deals which do not match the customer's search criteria. Mismatched recommendations are not a UX quirk under this framework — they are a transparency breach.
§ 05 · The numbers

The framework, in figures.

A handful of statistics that the rest of the field note rests on. The first four are taken from the guidance and its enforcement architecture. The fifth is a peer-reviewed anchor for the dark-patterns concern the policy paper raises.

Maximum fine
10% global turnover

Maximum penalty under the DMCCA 2024 for consumer-protection breaches. Aligns the UK with the GDPR fine ceiling. The CMA can impose this directly, without court proceedings.

DMCCA 2024 · gov.uk
Action sections in guidance
5steps

"Same rules apply" + four operational requirements (Tell, Train, Monitor, Refine). The most compact regulator framework of 2026 — short enough to fit a single A4 page.

CMA, 9 Mar 2026
Worked examples
4scenarios

Marketing campaigns, refund processing, customer-service queries, deal-comparison services. Each maps to a specific cluster of statutes and a specific operational failure mode the CMA expects to see.

CMA worked examples · 2026
Peer-reviewed dark patterns
1,818instances

Dark-pattern instances found across 11,000 shopping websites in Mathur et al's CSCW 2019 paper. Predates the agent era — the baseline before agents make the same patterns operate at greater scale and lower visibility.

Mathur et al · ACM CSCW 2019
§ 06 · Risks the policy paper flags

What the companion document sees coming.

The operational guidance tells deployers what to do. The companion policy paper, published the same day, tells regulators what to watch for. Five themes recur. None of them is fully addressable inside the four requirements. Each one is where the next round of enforcement will be drawn.

§ Risk 01 · Crit
Dark patterns at scale.
Agents personalise. Personalisation at scale becomes a vector for steering consumers against their own interests in ways no human operator could match. Decisions optimised against the consumer become invisible inside an aggregate of helpful suggestions. The CPRs 2008 doctrine of misleading omissions reaches this — but only if regulators can detect it.
§ Risk 02 · Warn
Agentic collusion.
Multiple businesses deploy autonomous pricing or commercial-strategy agents. The agents independently optimise. Without anyone deciding to collude, prices align. The competition risk is real even where no human ever signed a cartel agreement. Pricing-agent buyers should expect the CMA to be looking for this pattern across sectors.
§ Risk 03 · Warn
Vulnerable consumers.
Expectations are higher where the agent is interacting with vulnerable consumers or large numbers of consumers. The guidance is explicit about this — and explicit that sector context can raise the bar further. Financial services, health, energy, telecoms: the CMA has flagged these as zones where the standard ceiling rises.
§ Risk 04 · Structural
"Many hands" accountability.
When the model is by one vendor, the platform by a second, the deployment configuration by a third, and the interaction by a fourth — the audit trail breaks down. The CMA's response is to concentrate liability at the deployer: you commissioned the agent; you carry the outcome. The supplier mitigation strategy is contractual, not legal.
§ Risk 05 · Market
Lock-in and data mobility.
The policy paper expects interoperability and data mobility to be designed in from the start. Agent products that make it hard for consumers to leave — or that hoard data so that switching agents costs the consumer their history — concentrate market power and raise switching costs. A consumer-protection issue and a competition issue at the same time.
§ Peer-reviewed: the dark-patterns baseline
Mathur, Acar, Friedman, Lucherini, Mayer, Chetty and Narayanan documented 1,818 dark-pattern instances across 11,000 shopping websites in their CSCW 2019 paper (Princeton/Chicago), grouping them into 15 types and 7 categories. That study predates the agent era. The point: none of these patterns needed an AI to exist. What agents change is the scale at which they can be applied, the personalisation that targets each one to the individual consumer, and the invisibility of the steering inside what looks like helpful conversation. The legal exposure was there before agents arrived. Agents make it bigger.
§ 07 · Enforcement philosophies

Three regulators. Three theories of how to hold deployers to account.

The UK CMA is the first major consumer-protection authority anywhere in the world to publish AI-agent-specific guidance. It will not be the last. How each jurisdiction enforces its position on deployer liability tells you what posture you need to take depending on where you ship.

Approach · Direct regulatory
United Kingdom
CMA · DMCCA 2024
Enforcement mechanism
The CMA imposes fines directly. Up to 10% of worldwide turnover. No court proceedings required since the DMCCA commenced. Order can include compensation to affected consumers.
Speed
Fastest of the three. Investigation, finding, fine — all administrative. Judicial review available, but the CMA's finding is in force until overturned.
What the CMA wants
Compliance by design from launch day. The guidance is a roadmap — companies that "follow it now" will be in the strongest position when an investigation begins.
Approach · Litigation-driven
United States
CFAA · AB 316 · State UDAP
Enforcement mechanism
Plaintiffs sue. Federal anti-hacking law (CFAA) used by platforms against agent providers (Amazon v. Perplexity). State law used by consumers against deployers. California AB 316 removed "the AI acted autonomously" as a usable defence.
Speed
Slower at first; punishing at scale once case law sets. The first decisions take years; once the pattern is settled, class actions follow quickly.
What plaintiffs look for
Documented agent misbehaviour, communications showing the deployer knew and didn't act. Discovery is the operational risk — your training data, prompts, and monitoring logs are all in scope.
Approach · Layered framework
European Union
AI Act · PLD · GDPR
Enforcement mechanism
Multiple regulators, multiple frameworks. AI Act Article 50 from 2 August 2026 (transparency). Revised PLD 2024/2853 from 9 December 2026 (strict liability for defective software). GDPR Art. 22 already in force (automated decisions).
Speed
Variable. National DPAs enforce GDPR quickly; AI Act enforcement is still bedding in. PLD claims will be civil court matters.
What regulators want
Conformity assessments, documentation, downstream safety. The deployer's risk surface is broader than under the CMA framework — strict product liability for defectiveness adds an additional liability layer the UK doesn't have.

Sources: CMA guidance · DMCCA 2024 · California AB 316 · Amazon.com Services LLC v. Perplexity AI Inc., N.D. Cal., 9 March 2026 · EU Regulation 2024/1689 (AI Act) · EU Directive 2024/2853 (revised PLD).

§ 08 · Misconceptions

Three things the framework is not.

The CMA's guidance is short. That brevity has produced predictable misreadings. Below are the three most common claims the field note's analysis pushes back on.

01
It's only voluntary guidance.
It isn't. The guidance interprets the CMA's existing enforcement priorities under existing UK law — the Consumer Rights Act 2015, the Consumer Protection from Unfair Trading Regulations 2008, the Consumer Contracts Regulations 2013, and the DMCCA 2024. The CMA's enforcement authority is statutory. The guidance tells you what compliance looks like; the obligations are already binding.
02
It only applies to large companies.
It doesn't. The four requirements apply to any business deploying an AI agent in a consumer-facing role — regardless of size. The CMA's enforcement discretion may factor in proportionality and consumer impact, but the legal obligations are universal. A small e-commerce business that deploys a chatbot is in scope; a SaaS startup with an AI customer-service feature is in scope.
03
If a third party supplied the agent, they're liable.
They aren't, under this framework. The CMA is explicit: the deployer is responsible "in the same way you are responsible for what an employee does", even where someone else designed or provided the agent. The third-party supplier may be contractually liable to the deployer — but that's a separate question. To the consumer and to the CMA, the deployer is the responsible party. The supplier contract is a mitigation strategy, not a defence.
§ 09 · Self-test

Seven questions before you ship.

If you are deploying an AI agent into a UK consumer-facing role — chatbot, refund handler, recommender, marketing assistant — these are the questions the CMA's framework expects you to be able to answer. They map directly onto the four requirements plus the policy paper's structural concerns.

§ Deployer self-test · Pre-launch
For any business shipping an AI agent into a UK consumer-facing role
  1. Have you labelled the agent clearly wherever the fact of AI involvement would affect a customer's decision, and avoided overstating what it can do?
  2. Have you mapped each customer interaction the agent can have against the statutory rights that apply — CRA 2015, CPRs 2008, the Consumer Contracts Regs 2013 — and trained the agent to respect each one, including any of your own more generous contractual terms?
  3. Have you run A/B testing or unit testing of the agent's behaviour against compliant outcomes before deployment, with results reviewed by someone with relevant experience?
  4. Is there a documented human in the loop — with consumer-protection-law experience — monitoring the agent's decisions in production, with a defined escalation path?
  5. Can you refine prompts or workflows quickly when problems are detected — measured in hours, not weeks — especially where the agent reaches vulnerable consumers or large volumes?
  6. Where the agent was supplied by a third party, have you independently verified that they trained it for UK consumer-law compliance, and do your supplier contracts allocate the indemnity correctly (knowing that the indemnity does not transfer the CMA's enforcement authority)?
  7. If a problem is found, do you have a remediation plan you can execute at the scale at which the agent operates — and do you accept that forward-looking remediation does not cure prior breaches?
§ Research

What sits underneath this field note.

Five categories of source: the primary regulator document, the legal-academic and computer-science peer-reviewed scholarship, the supporting regulator publications, the underlying UK statute, and the practitioner analyses. The CMA's own guidance is at the top because it is the document everything else interprets.

⬢ Primary source · The CMA's framework
CMA · gov.uk · Guidance

Complying with consumer law when using AI agents

The document this whole field note interprets. Published 9 March 2026. Five sections, four worked examples. Authoritative statement of how the CMA reads existing UK consumer-protection law as it applies to AI agents in consumer-facing roles. Open Government Licence v3.0 — quotable in full, but the operational reading is what this page adds.

gov.uk · 9 March 2026
Read →
⬡ Peer-reviewed scholarship
ACM CSCW · Peer-reviewed · 2019

Dark Patterns at Scale · Findings from a Crawl of 11K Shopping Websites

Mathur, Acar, Friedman, Lucherini, Mayer, Chetty and Narayanan (Princeton/Chicago). Automated detection identified 1,818 dark-pattern instances across 11,000 shopping websites, sorted into 15 types and 7 categories. The empirical baseline for the dark-patterns concern in the CMA's companion policy paper. Honored by the Future of Privacy Forum's Privacy Papers for Policymakers award.

Proc. ACM Hum.-Comput. Interact., Vol. 3 No. CSCW · DOI 10.1145/3359183
Read →
Notre Dame Law Review · Forthcoming

Governing AI Agents

Noam Kolt, forthcoming in Notre Dame Law Review Vol. 101. Brings two analytic frameworks to agent governance: the economic theory of principal-agent problems and the common-law doctrine of agency. Predicts the structural logic underneath the CMA's "responsible like for an employee" rule. The most-cited legal-academic anchor for any analysis of deployer liability in 2026.

Notre Dame L. Rev. 101 · SSRN 4772956 · Feb 2025
Read →
Science · Peer-reviewed · 2024

Regulating advanced artificial agents

Cohen, Kolt, Bengio, Hadfield & Russell. The argument for why governance frameworks cannot rely on empirical testing alone for sufficiently capable agentic systems. Long-term planning agents recognise test environments and behave differently in production — the structural insight underneath the CMA's "human in the loop" and "monitor performance" requirements.

Science 384(6691):36–38 · DOI 10.1126/science.adl0625
Read →
Comp. Law & Security Review · 2026

Transparency in human-AI interaction — an analysis of Article 50(1) AI Act

Doctrinal and functional analysis of how EU AI Act transparency obligations compare to consumer-protection-law transparency duties. The CMA's "tell your customers" requirement and Article 50's "interaction with an AI" obligation overlap but do not align. Useful for any deployer operating across UK and EU markets.

ScienceDirect · April 2026
Read →
⬣ Regulator publications
DRCF · Foresight paper

The Future of Agentic AI

Cross-regulatory paper from the Digital Regulation Cooperation Forum (CMA, FCA, ICO and Ofcom together) on 31 March 2026. Defines a five-level autonomy spectrum, catalogues seven categories of compliance risk, and distinguishes "amplified" from "novel" risks. Carries a polite disclaimer that it should not be read as policy. Should be read as policy.

DRCF · 31 March 2026
Read →
CMA · Enforcement guidance

How the CMA uses its consumer protection enforcement powers (CMA58)

The CMA's general framework for consumer-protection enforcement under the DMCCA. Sets out investigation procedures, undertakings, directions, and the scale of fines. Reads alongside the agent-specific guidance — anything found to be a breach via the agent route is enforced by the same processes that apply to any other consumer-protection breach.

CMA · 2024–2026
Read →
DRCF · AI & Digital Hub

DRCF AI and Digital Hub · case studies

The CMA's own guidance explicitly directs deployers here for case-study examples. The Hub aggregates worked scenarios across the four DRCF regulators — useful for deployers operating across multiple regulatory frameworks at once (FCA-regulated firms deploying agents, ICO/data-protection touchpoints, Ofcom-regulated platforms).

DRCF · ongoing
Read →
⬨ Primary law & statutes
UK · Statute

Consumer Rights Act 2015

Statutory rights on goods, services, and digital content. Unfair contract terms (sections 62–65). The framework the CMA's refund-request worked example references directly. The CRA 2015 is what an agent must respect when handling returns, statutory faults, and contract performance.

UK Public General Act
Read →
UK · Statutory Instrument

Consumer Protection from Unfair Trading Regulations 2008

The UK's general unfair-commercial-practices framework. Misleading actions, misleading omissions, aggressive practices, the average-consumer test. The most likely route for enforcing against an agent that steers, pressures, or misleads consumers at scale. The CMA's policy paper specifically connects dark-pattern personalisation to this regulation.

UK SI 2008/1277
Read →
UK · Statute

Digital Markets, Competition and Consumers Act 2024

The act that transformed UK consumer-protection enforcement. The CMA can investigate, find, and fine for breaches of consumer protection law — up to 10% of worldwide annual turnover — directly, without court proceedings. The teeth that turn the CMA's guidance into a live operational risk.

DMCCA · 2024
Read →
UK · Statutory Instrument

Consumer Contracts (Information, Cancellation and Additional Charges) Regulations 2013

The 14-day cancellation right for distance and off-premises contracts; specific information requirements. The framework the CMA's refund-request worked example cites for change-of-mind returns. An agent that doesn't recognise or apply the cancellation right breaches these regulations.

UK SI 2013/3134
Read →
⬩ Practitioner analyses
Cooley · Legal analysis

AI Agents and Consumer Law: What Businesses Need to Know

The cleanest practitioner walk-through of the CMA guidance's four operational principles — transparency, compliance by design, human oversight, swift remediation — and what each implies in practice. Includes the operational point most internal compliance teams miss: forward-looking remediation is not a cure for, or a defence against, previous breaches.

Cooley · 26 March 2026
Read →
TLT · Legal analysis

Agentic AI · CMA publishes guidance on consumer law and DMCCA risks

TLT's reading of the two-document structure (guidance + policy paper) and the new DMCCA enforcement architecture. Strong on the operational expectations: pre-launch due diligence, supplier-contract considerations, the higher bar that applies for vulnerable consumers and high-volume agents.

TLT · April 2026
Read →
Lewis Silkin · Analysis

Agentic AI and consumer law · The CMA's guidance for businesses

Compact summary of the four CMA requirements with the specific use-case implications spelled out: marketing campaigns (Digital Markets Act + ASA rules + influencer disclosure), refund processing (CRA 2015 + Consumer Contracts Regs 2013), customer service (CPRs 2008), and agentic collusion as a competition-law concern.

Lewis Silkin · 13 March 2026
Read →
Ashurst · Privacy update

Data Bytes 65 · CMA agentic AI guidance in context

Ashurst's privacy team places the CMA framework alongside the parallel ICO consultation on automated decision-making under the Data (Use and Access) Act 2025. The intersection of the CMA's consumer-protection authority and the ICO's data-protection authority is where multi-track regulatory risk concentrates for any agent that processes personal data.

Ashurst · April 2026
Read →
§ Companion field note
The Two-Consent Problem · the other side of the agent
Where this field note sits inside the deployer's frame, the companion sits outside it — the gap between user authorization and platform authorization in agentic AI, after Amazon v. Perplexity on the same 9 March 2026 day the CMA published this guidance. Read them together.
Open →