EU directive holds companies selling AI agents liable for strict liability

Summary

The EU Product Liability Directive (2024/2853), which must be transposed into national legislation by December 9, 2026, classifies AI agents as 'products.' Sellers are held strictly liable for bodily harm, psychological damage, and the loss of non-professional data resulting from defects. Pure economic losses and privacy breaches are explicitly excluded, and only natural persons are eligible to claim compensation.

Why it matters

The directive removes the 'proof of negligence' barrier. This means any company that brands or significantly modifies an AI agent bears full liability to end-users, regardless of contractual exemption clauses. This poses a direct financial risk to banks, insurers, and retailers that offer third-party agents under their own brands. It also creates a gap: model providers' standard terms (typically limited to 12 months of fees) may not cover the integrator's downstream liability. The directive also retroactively covers agents launched before 2026 that undergo significant updates or continuous learning processes after December 9, 2026.

Highlights

• Strict liability without contractual shields: Article 15 prohibits any clause that limits the producer's liability to the injured party.
• Narrow scope: Only death, bodily/psychological harm, damage to private property, and loss of non-professional data are compensable. Pure economic losses and discrimination are excluded (Recital 24).
• The 'significant modification' trigger (Recital 40 + Article 7) implies that agents launched before 2026 but continuing to learn or receiving major updates will be retroactively covered.

Critical take

The article's most critical flaw is the exclusion of pure economic loss. This issue is brushed aside by framing the matter as a 'contract and payment law' problem. In practice, however, most real-world agent errors (incorrect transfers, erroneous purchases) are exactly this type. This leaves the majority of consumer harm to the patchwork of national contract regimes. Additionally, the text relies heavily on 'ActuIA's reading' for certain interpretive claims, such as model providers being considered component manufacturers and employees being included among covered persons. This indicates that these statements are editorial inferences rather than established legal positions.