AI Liability in Switzerland: Who Pays When Your AI Agent Errs?

When an agent that reviews invoices, answers customer emails or drafts contracts makes a mistake, the first question in the boardroom is usually: who pays for this? The short answer: not the AI. It has no legal personality, no assets and no liability a court could assign to it. Under current Swiss law, responsibility rests with the same company that deployed the agent — just as it would for a human employee.

The short answer: responsibility stays with the company

Swiss contract and tort law has no special provision for artificial intelligence. It does not need one: the existing liability rules are technology-neutral and treat an AI agent like any other tool or auxiliary person acting on a company's behalf. For boards and executive teams, this means the question is not whether liability exists, but how it is allocated — and whether it is insured.

Three articles in the Code of Obligations every board should know

Three provisions of the Swiss Code of Obligations (CO) form the foundation of any discussion about AI liability in Switzerland:

1. Art. 41 CO (unlawful harm): Anyone who unlawfully causes damage to another is liable for it — regardless of whether a human or an AI agent triggered the error.
2. Art. 55 CO (vicarious liability): A company is liable for damage caused by its auxiliary persons while carrying out their assigned tasks. The prevailing interpretation increasingly extends this to digital auxiliaries such as AI agents acting on the company's behalf.
3. Art. 97 CO (contractual liability): Anyone who defectively performs a contractual obligation — for example toward a customer — is liable for the resulting damage, even if an AI system executed the process.

When AI is embedded in a physical product: the Product Liability Act

Once an AI system does not merely advise digitally but is embedded in a physical product — a machine control unit or a medical device, for instance — the Swiss Product Liability Act also applies. It holds the manufacturer liable regardless of fault when a defective product causes harm. For companies integrating AI components into their own products, this raises the bar for due diligence in model selection, testing and documentation.

The EU shifts the burden of proof: Directive 2024/2853

In November 2024, the EU adopted Directive 2024/2853, updating product liability for the AI era. The central change: in cases of damage caused by defective AI systems, the burden of proof is partially reversed. Claimants no longer need to prove in technical detail how an algorithm failed — manufacturers and operators increasingly need to demonstrate that their system was safe. Switzerland is not directly bound by this directive, but it becomes practically relevant for companies with customers or subsidiaries in the EU — and it signals the direction Swiss practice is likely to follow.

Why the terms of service of ChatGPT, Claude and Gemini won't protect you

OpenAI, Anthropic and Google word their terms of service carefully: all three disclaim liability for errors, misinterpretations or damage arising from the use of their models. This is legally understandable — and it also means the liability risk does not stay with the model provider, but shifts to whoever deploys the model within a business process. If you build an AI agent on top of these models, you effectively inherit the responsibility the provider contractually disclaims.

The case that shows the direction: Nippon Life v. OpenAI

A case analysed by Stanford in 2026 — Nippon Life v. OpenAI — frames AI liability explicitly as a matter of product liability, not as an unprecedented grey zone. The reasoning follows a logic that is also relevant for Swiss companies: an AI system is treated like a product that can have a defect, for which whoever places it in commerce or deploys it is held accountable. For Swiss SMEs, this signals that international legal developments are clearly moving toward more, not less, responsibility for deploying companies.

The insurance gap: what your policies probably don't cover

Traditional general liability and professional indemnity policies were designed for human action — for mistakes an employee makes, not for autonomous decisions an agent makes across hundreds of transactions within seconds. Many policies contain exclusions or ambiguous wording when it comes to damage arising from automated or AI-supported decisions. This gap only becomes visible once a claim occurs — by which point it is too late to close it.

What this means for the board and executive management

Under Swiss corporate law, the board of directors bears overall responsibility for adequate risk control within the company. Deploying AI agents does not change this — it only shifts the nature of the risk. Automating processes today without clarifying liability, insurance coverage and governance structure means operating in a legal grey zone that neither the Code of Obligations nor the Product Liability Act will resolve for you.

The good news: none of this prevents the sensible use of AI agents in your own operations. It simply changes how carefully the rollout needs to be prepared — from model selection through contractual structuring to insurance coverage. That preparation is what decides whether AI liability remains a theoretical risk for your company or becomes a very real one.