AI Law, Risk & Governance

Liability for AI errors: What managing directors should know

Marin Niehues3 March 202512 min read

Two men discussing a document

A company is liable to its customers for AI errors as it is for any other service, and senior management owes an organization that detects those errors. A language model that gives a customer incorrect information or produces a contract draft with an invented clause creates no special legal status for machines. The liability question is settled by the organization: who approved the use, who checks the results, and who can show that the check took place? This article sets out the liability scenarios including the new product liability rules, describes a review framework for senior management and shows how it becomes a dependable operating routine. This article provides an overview and does not constitute legal advice.

Who is liable for AI errors?

For AI errors, the company that deploys the AI is liable: towards customers under the contract and towards third parties under general law. Liability scenarios differ by counterparty. Towards customers and contractual partners, a company answers for the agreed service, whether an employee or a model produced it. Towards third parties, claims arise when an AI result infringes personality rights or copyright. Product liability comes on top for companies that build AI into their own products or offer it as software.

A further level concerns senior management itself. Under German company law, Section 43 of the GmbH Act and Section 93 of the Stock Corporation Act, management owes an organization that identifies and steers risks; if it breaches this duty, it is personally liable to the company. In practice this leads to a simple test question: did the organization decide on the use deliberately and monitor it continuously?

Chatbots in customer service: who answers for wrong information

A chatbot's answer is attributed to the company in the same way as an employee's; a notice that the bot may make mistakes does not relieve the company of responsibility for the answer. Under German law, automatically generated declarations count as declarations of the operator that set up the system. In 2024 a Canadian civil resolution tribunal held an airline to the incorrect refund information given by its chatbot and rejected the argument that the bot was responsible for its own statements.

Three design rules follow for customer-facing applications. The chatbot answers only from approved sources, such as current terms and price lists. It does not make commitments on prices, deadlines or refunds itself; it hands over to a person. Every answer is logged so that, in a dispute, the company can show what the customer read.

Does product liability apply to AI?

Yes, under the new EU Product Liability Directive, software including AI systems is expressly a product, and the manufacturer is liable for its defects regardless of fault. Under the previous product liability regime it was disputed whether pure software was covered; the European legislator has now settled that question. The member states are transposing the directive into national law.

Four points of the new directive matter for companies:

  • Software is a product, whether embedded, installed locally or provided as a service.
  • Anyone who substantially modifies an AI system, for example through their own retraining, becomes a manufacturer themselves.
  • Defects also arise from missing security updates and from behavior that a learning system develops only after being placed on the market.
  • Injured parties benefit from easier proof: courts can order the disclosure of evidence, and for technically complex products a defect is presumed under certain conditions.

The boundary matters just as much. Product liability covers personal injury and damage to health as well as damage to property used privately; this includes the loss of data used privately. Pure financial loss and damage to property used exclusively for professional purposes are not covered. For the typical B2B error, a flawed analysis causing financial loss, contractual liability remains the basis.

Four error scenarios that occur in operation

Four error scenarios can be distinguished from project work, each requiring its own control.

  • Incorrect information: a model phrases something plausibly but inaccurately. This affects customer communication and internal research. Control: grounding in sources and approval before sending.
  • Flawed decision support: an analysis or summary produces a distorted picture on which an investment or a personnel decision is based. Control: documented plausibility check by the responsible specialist.
  • Data leakage: employees enter confidential or personal information into a tool that has not been approved. Control: approved tools and a clear policy.
  • Biased results: a system treats groups systematically differently, for example in pre-screening applications or in credit questions. Control: bias testing before deployment and during operation.

The first scenario occurs most frequently and is most often underestimated. The control must be a fixed step in the process so that it takes place independently of individual people. For infringements through generated text and images, the article Copyright and AI-generated content sets out the position.

Is a managing director personally liable for AI errors?

A managing director is personally liable when breaching the duty of care under Section 43 of the GmbH Act, as is a management board member under Section 93 of the Stock Corporation Act, for example through a risky use of AI without controls. The liability runs towards the company, which reclaims the loss. Protection applies to anyone who takes a business decision on an adequate information basis in the interest of the company. For AI this means that a documented risk classification, a deliberate approval and ongoing monitoring are the evidence that management has fulfilled its duty.

The obligations under the AI Act raise this standard. The regulation requires sufficient AI literacy among staff and, for high-risk systems, human oversight and monitoring in operation. Anyone who fails to organize these obligations has weak arguments for diligent management in a dispute.

Five questions for every AI application

A review framework for senior management has to work with a small number of questions that are asked in the same way for every application. Five have proven effective.

  • Who is accountable for the subject matter? Every application has a named person in the business unit who answers for the results. Operating the technology sits separately with IT.
  • Which risk class applies to the use? An internal text summary is subject to different requirements from a customer response or a decision about individuals.
  • Where does a human check? For each class it is defined whether a person checks before the result takes effect, checks on a sample basis or intervenes only when anomalies appear.
  • What can be evidenced? Approval and review rules, together with test results, are available in a form that allows a third party to follow the process.
  • What happens when an error occurs? The reporting route and the shutdown criterion are set in advance, as is responsibility for communication.

The answers belong in a central inventory of all AI applications. This inventory is the basis for governance and at the same time the first document requested in a dispute. How committees and policies are derived from it is shown in the article Building AI governance. For assessing individual risks, AI risk management: a framework for decision-makers provides the method.

Contracts and insurance: where liability shifts

Providers of AI tools regularly limit their liability in their terms to a narrow scope. Towards its own customers, the deploying company is liable under the contract; whether it can take recourse against the provider is determined by the liability cap and the indemnity clauses in the provider contract. Between the two positions there is a gap that the organization closes through selection and contract design, supplemented by its own controls.

For procurement and the business unit, a fixed checklist pays off: how high is the provider's liability cap? Is there an indemnity for infringements caused by outputs? How does the provider use input data? Which audit and information rights does the company have, and how does the provider report incidents? The contractual assessment in an individual case belongs with your legal specialists.

Insurance covers part of the residual risk. D&O insurance protects directors and officers when the company or third parties hold them personally liable for a breach of duty. It does not cover the company's liability towards customers; general and professional liability policies as well as cyber policies serve that purpose. Before expanding the use of AI, it is worth checking whether the policies expressly include or exclude AI-related losses.

Niehues Consulting as a partner for the liability framework

A liability framework takes effect when it is anchored in processes and responsibilities. This is exactly where Niehues Consulting comes in. AI management produces business analysis and requirements as well as the steering of the AI portfolio with metrics, including the inventory and the risk classes. In the rollout, review points and approvals, together with enablement of the business units, are designed so that they are followed in day-to-day business. Organizational design defines roles and decision paths so that it is clear who acts when an error occurs.

The approach follows the Niehues Delivery Model with the phases Understand, Build and Roll out. For liability, three of its review questions matter. Who decides, who delivers, who uses? Which roles, decision paths and interfaces does operation require? Who is accountable for operation from day one? Legal requirements remain the task of the company's legal specialists; Niehues Consulting delivers the organization in which those requirements take effect in operation. How the liability framework fits into overall AI governance is shown in the article AI management: what executive leadership must steer. The entry point is a free initial consultation of 30 minutes via the contact page.

Frequently asked questions

Is senior management personally liable for AI errors?

Yes, if it breaches its duty of care under Section 43 of the GmbH Act or Section 93 of the Stock Corporation Act, for example through missing controls in a risky use. The liability runs towards the company. A documented risk classification and approval are the most important evidence of diligent conduct.

Who is liable if the AI provider causes the error?

Towards its own customers, the deploying company is liable under the contract. Whether it can take recourse against the provider is determined by the liability cap and the indemnity clauses in the provider contract. These points therefore belong in the review before every procurement.

Who is liable for wrong answers from a chatbot?

The company that operates the chatbot is liable for its wrong answers, as it would be for information given by an employee. A notice that the bot may make mistakes does not change this. Protection comes from grounding in sources, handing commitments over to people and complete logs.

What happened to the AI Liability Directive?

The European Commission has announced that it will withdraw its proposal for an AI Liability Directive because no agreement was foreseeable. AI damage is therefore governed by general contract and tort law. The revised Product Liability Directive, which covers software including AI, is the new addition.

Does D&O insurance cover AI errors?

D&O insurance helps directors when they are held personally liable for an AI error. It does not cover the company's liability towards customers. Cover lapses for knowing breaches of duty, so documented diligence counts here as well.

Does every AI result have to be checked by a person?

No, the extent of the review follows the risk class. For customer responses and decisions about individuals, a person checks before the result takes effect; for internal text drafts, samples are sufficient. The rule per class belongs in the inventory.

How do I document the use of AI so that it stands up as evidence?

A central inventory of all applications with the responsible person for each is the starting point. Approval records and review rules are added, along with test reports and an incident log. The records are created in day-to-day business so that they are complete when an incident occurs.

Let’s talk about your initiative.

In 30 minutes we clarify where your organization stands and which step will pay off most now.

Request a consultation

30 minutes · free · no obligation