How Should a Bank Prepare Its AI Systems for the EU's New Product Liability Rules?

6 min readAbleneo AI transformation team

Short answer. From 9 December 2026, the revised EU Product Liability Directive (Directive (EU) 2024/2853) treats software and AI systems as products, so a bank can be held strictly liable when a defective AI system causes harm. The directive also lets a court order the bank to disclose how an AI system works and presume the system was defective when that evidence is missing. A bank prepares by documenting every production AI system, keeping an evidence trail an outsider can read, and fixing vendor contracts before the deadline.

1What This Means in Practice

The old product liability regime, Directive 85/374/EEC, covered physical goods. The revised directive extends “product” to software, firmware, applications, and AI systems, whether they run on a device or in the cloud. Developers and providers of that software are treated as manufacturers. For a bank, this pulls the credit-scoring model, the customer-facing AI agent, and the data pipeline that feeds them inside the liability net.

Liability is strict. A customer who suffers harm does not have to prove the bank was negligent. They prove the product was defective and that the defect caused the damage. That is a lower bar than most banks assume, and it applies to AI outputs that reach a customer.

Concrete situations a bank should map now:

2Why This Matters for Regulated Industries

Member states must have national law in place by 9 December 2026. Banks and insurers in the Czech Republic, Slovakia, and Austria already carry EU AI Act and DORA obligations, and the directive stacks a civil-liability layer on top of both. Annex III of the EU AI Act classifies credit scoring and insurance risk pricing as high-risk, with those obligations deferring to 2 December 2027, so the systems most exposed to a liability claim are also the ones under the heaviest compliance load.

The separate AI Liability Directive was withdrawn in 2025, which leaves the revised Product Liability Directive as the main EU route for harm caused by AI. For a regulated firm that means there is now one clear instrument to prepare for, not a moving target. The bank that can produce a clean record of how a system was built and run will spend the next three years defending claims on evidence rather than on hope.

From 9 December 2026, the revised EU Product Liability Directive treats AI systems and software as products a bank can be strictly liable for.

3What Changes for AI Software Under the New Directive?

The definition of a product now names software directly, including operating systems, applications, and AI systems, embedded in a device or delivered as a service. Non-commercial open-source software supplied outside a commercial activity is excluded. A software update that introduces a defect, or a failure to supply a security update the product needed, can itself trigger liability. For a bank this means the model is not the only thing in scope. The surrounding software, the integration code, and the data pipeline all count as part of the product, and a defect in any of them can land on the bank.

4How Does the Burden of Proof Shift to the Bank?

The directive adds two mechanisms that move weight onto the defendant. First, a court can order the bank to disclose relevant evidence about how a system was built and operated, once the claimant shows the claim is plausible. If the bank cannot or will not disclose, the court can presume the product was defective. Second, when technical or scientific complexity makes proof excessively difficult for the claimant, the court can presume defectiveness or a causal link, and the bank then has to rebut it.

The practical effect is blunt. A system with no logs, no version history, and no readable documentation hands the claimant both presumptions by default. A system with a complete evidence trail lets the bank answer a disclosure order and rebut the presumption. Trade secrets are protected, so disclosure is controlled rather than open-ended, but the protection only helps a bank that has the evidence in order in the first place.

5Who Is Liable, the Bank or the AI Vendor?

Both can be. The directive reaches any economic operator that makes a product available: the manufacturer of the AI component, the party that integrates it, and in defined cases the one that substantially modifies it. A bank that buys a model from a vendor and runs it in a customer-facing workflow can be named as a defendant even when the defect sits in the vendor’s code. The bank then recovers from the vendor, which is a contract question between them, not a defence against the customer. This is the reason to rewrite AI vendor contracts now so they carry evidence-sharing clauses, indemnities, and a defined liability split before the deadline.

6What Should a Bank Do Before 9 December 2026?

Four steps, in order. Inventory every AI system in production and name a single accountable owner for each one. For each system, keep a documentation file an outsider can read: its purpose, data sources, model version, test results, and the human controls around it. Keep immutable logs of inputs, outputs, and human overrides, so a disclosure request can be answered in days rather than months. Then rewrite AI vendor contracts to cover disclosure, indemnity, and the liability split.

These are the same controls DORA and the EU AI Act already require. A bank that builds documentation, logging, and human oversight once covers all three obligations with one program, which is far cheaper than running a separate liability project six weeks before the deadline.

7The Ableneo Perspective

Ableneo builds AI systems that are observable and accountable from the first pilot, which is the exact posture the new directive rewards. In 2025 Ableneo shipped 34 production AI projects across 7 industries and 4 countries, and roughly 4 of 5 of its projects reach production, so the evidence trail is designed in rather than reconstructed before an audit. For a Czech bank weighing its exposure, Ableneo’s work with T-Mobile Czech Republic cut document validation from 7 hours to 30 minutes at over 90% accuracy, with the logging and human checks a disclosure request would later rely on. For the related question of internal accountability, see Ableneo’s answer on who is liable when an AI system makes a wrong decision in a bank.

Key takeaways

Sources

Planning AI in a regulated business? Ableneo takes systems from classification to governed production.

Talk to Ableneo