EU Digital Fairness Act First to Regulate AI Shopping Agents: Three Authorization Tiers and 6% Fines Draw the Legal Boundaries of Agent Autonomy

On November 11, 2026, the European Commission formally proposed the Digital Fairness Act (DFA). One clause specifically targeting AI shopping agents quickly became a focus of the technology world: any company providing an AI shopping assistant to EU consumers must, when a user first uses it, require the user to choose among three authorization modes. Violations could result in fines of up to 6% of global annual turnover. This is the first time in history that formal legislation explicitly defines “to what extent an AI agent can make decisions on a person’s behalf.”

The DFA’s AI agent provisions appear in Articles 26 and 27. The logic behind the three authorization tiers is as follows: the first tier is “confirm everything”—before each purchase or contract the AI is to make on the user’s behalf, it must obtain the user’s explicit consent; the second tier is “conditional automation”—the user pre-sets conditions, including a spending cap and validity period, and the AI may act autonomously within those conditions, requiring confirmation only when they are exceeded; the third tier is “full delegation”—the user waives prior approval and authorizes the AI to act entirely on their behalf. According to legal analysis firm Lewis Silkin, the proposal requires platforms to communicate the user’s choice through a “durable medium” (email or an archivable file) to ensure it has legal effect.

But even if a user chooses the third tier, “full delegation,” the law still sets a rigid exception to that authorization: when the provider of the AI shopping assistant receives remuneration from the transaction counterparty, or has any conflict of interest with the seller, the transaction must return to the user for confirmation regardless of the authorization settings the user previously chose. This exception directly targets one of the most hidden forms of loss of control in current AI recommendation systems—the agent is deciding on the user’s behalf while in fact deciding on behalf of another paying principal.

From an engineering implementation perspective, the three-tier authorization design is not complex; what is difficult is the compliance infrastructure behind it. First, the agent system needs to query the user’s current authorization status before initiating each transaction, and that query must occur before the order is actually submitted, not after. Second, the system needs to detect in real time whether the current transaction triggers the “conflict of interest” exception—which recommendations are paid placements and which are not—and, when triggered, interrupt the automated flow and push a manual confirmation request. Third, all such authorization records must be persisted in a form the user can retain, which means pure frontend state management is insufficient; the backend must maintain a complete authorization event log.

The DFA also notes a deliberately left-open design: the regulation does not require platforms to proactively inform sellers of the spending caps set by users. This leaves an enforcement gap in cross-system transaction scenarios—whether the user-side limit constraint can truly take effect downstream in the supply chain depends on the buyer platform’s technical implementation, not on a legal mandate. For AI shopping assistants operating on third-party seller ecosystems (such as AI agents embedded in e-commerce platforms), this gap means the boundary of compliance responsibility remains unclear.

Overall, the DFA is a systematic replacement for the EU’s 2005 Unfair Commercial Practices Directive (UCPD). The legislative framework from two decades ago could not foresee that AI agents would sign contracts on users’ behalf, much less that recommendation algorithms would implant commercial interests without consumers’ awareness. The DFA explicitly lists several “dark patterns” as illegal, including false low-stock warnings, shame-based unsubscribe design (such as requiring users to click “I don’t like good deals” in order to unsubscribe), and pre-checked default options. Beyond AI shopping agents, the proposal also sets out a simple exit mechanism for subscription services, full disclosure requirements for dynamic pricing, and default-off rules for virtual currency for minors. The Commission estimates that if these measures are fully implemented, EU consumers could save about €13.7 billion annually.

For developers of global AI agent products, the impact of this proposal is not limited to the EU market. The technical architecture of AI shopping assistants is usually deployed globally. Adding three authorization choices for EU users means the product must support an authorization-state management mechanism worldwide, or otherwise maintain a separate parallel architecture for the EU. Historically, companies have been far more likely to choose the former—after the GDPR was introduced in 2018, many technology companies extended EU-standard privacy settings to users worldwide, and the Digital Fairness Act could trigger a similar “Brussels effect.”

For consumer organizations, the proposal provides a legal handle that has never existed before: when an AI agent completes a purchase on a user’s behalf without explicit authorization, or automatically places an order involving paid placement without a prompt, consumers will for the first time have a clear legal basis. Before this, such disputes could only rely on platform terms of service, and those terms were usually defined by the platforms themselves.

For the industry, the issue of compliance costs is unavoidable. The three tiers of authorization are not complex to implement on the frontend, but embedding conflict-of-interest detection logic into the decision chain of a shopping agent involves deep changes to the recommendation system architecture—especially when the recommendation model itself does not distinguish between “paid priority” and “relevance priority,” this detection line almost requires system-level redesign. If the proposal is ultimately adopted, platforms whose AI agent capabilities are still at an early stage and whose recommendation systems are deeply coupled with paid logic will face far greater compliance retrofit costs than systems that were clearly layered at the architecture design stage.

In terms of the legislative process, the proposal is now entering review by the European Parliament and the Council. According to Lewis Silkin’s analysis, the text “is likely to evolve in negotiations with the European Parliament and the Council,” and the specific effective date and details have not yet been determined. Considering that the GDPR took four years from proposal to formal entry into force, the DFA at this stage is more like a design specification than an imminent compliance requirement.