# The EU just wrote AI shopping agents into law

> The Digital Fairness Act proposal gives shoppers control over how much an agent can buy alone, forces agents to disclose who pays them, and lets offers be made in machine-readable form. Here is what it means for brands that sell.

By [Yotam Gabi](https://bazak.ai/writing/authors/yotam-gabi), Co-founder & CEO · Published 2026-11-11 · Market research, Agentic commerce

On 11 November 2026, the European Commission adopted its proposal for a Digital Fairness Act. Most of the coverage will focus on dark patterns, subscription traps and influencer disclosure. For anyone building or selling through AI agents, the more important text sits in two articles near the end of Chapter III.

Articles 26 and 27 are the first time EU consumer law describes an AI shopping agent in operational detail: who controls it, what it has to disclose, and what the seller owes it. The proposal still has to pass the European Parliament and the Council, and it may change on the way. But the shape of the rules is now on paper, and it is worth reading closely.

## The shopper sets the leash

The proposal defines an automated system as software that can take decisions with legal effects for its user, including concluding contracts, without human intervention. Article 26 then sets out what the provider of such a system has to build.

On first use, the shopper has to choose one of three settings. Confirm every contract the agent proposes. Confirm only certain contracts, based on criteria the shopper sets, including a maximum price and a time limit. Or confirm none at all. The provider sends the choice back on a durable medium, and the shopper can change it at any time.

Two more lines in the same article matter. A seller cannot design its site so that shoppers are forced to use an agent to buy. And the article does not apply to financial services.

None of this limits what agents can do. It makes autonomy something the shopper grants explicitly, in steps, and can take back.

## The agent has to say who pays it

Article 26(4)(b) is the clause I expect to shape business models. The agent has to disclose any payment its provider receives for proposing a contract, and any conflict of interest with the seller involved. When there is such a payment or conflict, the shopper has to confirm that specific contract, whatever setting they chose.

Read that from the seller's side. A product the agent recommends on its merits can flow straight through under the shopper's autonomy setting. A product the agent recommends because someone paid for the placement stops and waits for a human confirmation, with the payment named.

For twenty years, paying for placement was the fastest way to the top of a list. Under this proposal, inside an agent, paid placement is the slower path to the sale.

## The offer has to be readable by an agent

Article 27 covers contracts concluded solely through automated systems on both sides: the shopper's agent buying from the seller's agent. In that case, the seller may provide the required pre-contractual information only in a commonly used machine-readable format, in place of the usual human-facing pages. The shopper still receives a confirmation of the contract on a durable medium, in plain and intelligible language.

Strictly, this is a permission, not a new obligation. In practice it changes the question a brand should ask. If a seller's terms, prices, delivery conditions and withdrawal rights exist only as web pages written for people, they cannot be handed to an agent in the form the law now recognizes. If I ran e-commerce for a brand selling into Europe, I would read Article 27 twice. A Commission text now treats a structured, machine-readable offer as a lawful way to inform a consumer.

## The rest of the act lands on the seller's agent

We build agents that sell, so the other provisions land on our desk too.

Article 10 prohibits designing or operating an online interface in a way that is likely to deceive or manipulate consumers, and lists dark patterns that are banned in all circumstances. Commissioner McGrath singled out false claims that a product is nearly out of stock. A fake "only 2 left" line generated by an agent is still a fake "only 2 left" line.

Article 17 requires traders to give adult consumers tools to switch off features that encourage compulsive use. One of the listed features is persistent prompts to continue engaging with an automated system. A sales agent that keeps pulling a shopper back into the conversation is now inside the scope of addictive design.

Article 25 gives consumers the right to reach a human being in post-sale customer service. Fully automated support after the purchase is off the table.

Article 34 may be the one with the longest reach. Where a claimant shows enough evidence, or an authority has a reasonable suspicion, courts and authorities can require a trader to disclose evidence about the design, operation and impact on consumer behaviour of an online commercial practice. Failing to disclose it can be treated as a relevant element in finding the practice unfair. For an AI agent, the practice is the agent itself: its prompts, its nudges, the experiments that shaped it.

## What it means for AI companies

Most of Article 26 is addressed to the companies that put agents in front of consumers: the assistant providers, and vendors like us that build agents for brands. Several obligations land directly on them.

**The assistant becomes the information desk.** When a shopper has chosen to confirm purchases, the agent itself has to present the core pre-contractual information before asking: the main characteristics of the product, who the seller is, the total price, delivery and the right of withdrawal. Today an assistant can hand that job to the merchant's checkout page. Under the proposal, it cannot.

**The business model sets the autonomy.** Any payment for proposing a contract and any conflict of interest, including favouring a seller the provider is tied to, triggers a disclosure and a mandatory confirmation. An assistant that earns affiliate fees, sells placement or sells its own products will find its most profitable recommendations are the ones that can never run on autopilot.

**Ads in the chat inherit the targeting rules.** The recitals extend the ban on profiling-based advertising to minors, and on ads built from sensitive personal data, to product rankings and recommendations in natural language systems. As assistants start selling ads, the answer itself falls inside advertising law.

**The customer cannot be hidden from the seller.** When an agent concludes a contract, it has to pass the consumer's contact details needed for performance to the trader. An assistant cannot sit between the buyer and the brand and keep the relationship to itself.

**Engagement has a switch.** Article 17 names persistent prompts to keep engaging with an automated system as a feature adults must be able to turn off. Assistants tuned for session length will need an off switch.

**Liability flows back up the chain.** Under Article 36, a brand held liable for an unfair practice can pursue the business partner whose act or omission caused it. A brand whose agent vendor generated a fake scarcity line will have a legal route back to that vendor, and agent contracts will start to reflect it.

The Commission can also investigate large cross-border cases directly. For the largest AI companies, that means one regulator, one file and a fine ceiling of 6% of worldwide turnover.

## What I would start today

A log of every urgency line, every nudge and every A/B test your agent runs, with the reason each one is there. Under Article 34, a brand that cannot explain why its agent said something will find that silence counted against it.

I would add one more list: every recommendation your agent makes where money changes hands behind it. Under Article 26, those are the recommendations that will need a confirmation step and a disclosure.

## The timeline

This is a proposal. The Parliament and the Council have to agree on a final text, and some observers do not expect the rules to apply before the end of 2028. Once the regulation enters into force, it applies twelve months later. It covers any trader targeting consumers in the EU, including traders established outside it. In cross-border cases the Commission investigates directly, fines can reach 6% of worldwide annual turnover.

## What we do at Bazak

We build sales agents that run inside ChatGPT, Claude and Gemini and on brands' own websites. We read this proposal as a description of where serious agents were heading anyway: the shopper sets the limits, the agent is honest about who pays it, and the brand can explain every nudge.

## Where this leaves brands

The EU has written down what a trustworthy shopping agent looks like before most brands have deployed one. That is an unusual position to be in. A sales agent that can explain its own nudges has nothing to change. One that cannot will have a year from the day this becomes law.

## Sources

- [European Commission, Proposal for a Digital Fairness Act, COM(2026) 1100](https://commission.europa.eu/document/download/0575d59f-554f-4a9e-b54b-e6b717175809_en?filename=Proposal+for+a+Digital+Fairness+Act.pdf)
- [European Commission, Digital Fairness Act](https://commission.europa.eu/law/law-topic/consumer-protection-law/digital-fairness-act_en)
- [Digital Fairness Act, Full text of the proposal](https://digitalfairnessact.com/digital-fairness-act-full-text)
- [Quartz, E.U. proposes Digital Fairness Act to ban dark patterns online](https://qz.com/eu-digital-fairness-act-dark-patterns-addictive-features-111126)
- [Trending Topics, Digital Fairness Act: Start-Ups Warn of Costly New E.U. Obligations](https://www.trendingtopics.eu/digital-fairness-act-iab-criticism/)
- [Carmelon Digital Marketing, Hebrew summary of the proposal](https://eu-dfa-site.vercel.app/)

---

[Bazak](https://bazak.ai) · [Get a demo](https://bazak.ai/demo) · [All pages](https://bazak.ai/llms.txt)
