11-18-2026 Article

AI in Everyday Work: The Legal Risks of Uncontrolled Use and How Companies Can Manage Them

In many companies, employees now use artificial intelligence to summarize management board presentations, review contracts, and draft communications. This often happens without the involvement of the data protection officer or the legal department, and not infrequently via personal accounts rather than the systems provided by the company. This is a classic case of “shadow IT” – that is, the use of digital tools that have not been approved by either the IT department or a compliance department. What may seem practical can have significant legal consequences.

At its core, there are two key questions: What information may be entered into which AI system, and who reviews the results before they are used? In the event of violations, the company faces fines and claims for damages; members of the management and supervisory board face personal liability toward the company; and in some cases, the individuals involved may also face fines and criminal liability. This article highlights the key risk areas and explains how they can be managed through a contract with the provider, an internal AI policy, and a standardized review process.

I. What Happens to Prompt Inputs and Uploaded Data at the Provider's End

Most AI services do not run within the company itself but on the provider's servers, often in the U.S., and in the case of Chinese providers, also in China. Therefore, whatever an employee enters there is transmitted to another company. What the provider is permitted to do with the inputs is governed by the contract entered into with the provider or by its general terms and conditions. In consumer versions – whether free or paid – the provider is often permitted to store the inputs, allow its own employees to view them, and use them to train its models. This can be excluded in specific B2B contracts between AI providers and companies, but it must be expressly agreed upon.

II. Erroneous AI Results Are the Company's Responsibility

AI systems generate convincing-sounding texts that may be factually incorrect or omit essential information without this being apparent. They misrepresent key figures from financial statements, omit crucial points when summarizing a management board presentation, overlook termination or change-of-control clauses in contracts, or invent sources and court decisions. Since mid-2025, German courts have repeatedly objected to legal briefs containing fabricated citations. The responsibility lies with whoever uses an AI-generated result. Little can be expected from the provider in the event of damages, as its standard terms and conditions assign the responsibility for verifying the results to the customer and limit the provider's own liability. The company, on the other hand, is held accountable for what it communicates externally with the help of AI. In May 2026, the Higher Regional Court of Hamm ruled that a company must bear responsibility under competition law for fabricated statements made by its chatbot.

III. The Legal Pitfalls

1. Data Protection When Entering Personal Data

If entries contain personal data, the General Data Protection Regulation (GDPR) applies. Even the name of a contact person in a contract is sufficient to trigger this requirement. The obligations under the GDPR are wide-ranging: ensuring a legal basis for processing, entering into data processing agreements, documentation requirements, safeguarding data subjects' rights, erasure obligations, and – depending on the use case – data protection impact assessments and the implementation of additional security measures, to name just a few of the GDPR's requirements. Without consulting the legal department or the data protection officer, these obligations – each of which is subject to fines – will not be met.

For providers based in the U.S., there are additional requirements regarding transfers to third countries (see our Data Protection Update No. 204). Because the European Court of Justice has yet to rule on the validity of the EU-U.S. Data Privacy Framework, standard contractual clauses should be agreed upon as a precautionary measure, even with certified providers. These standard contractual clauses are usually not included in the providers' terms and conditions for private or consumer accounts, and even in B2B contracts, their applicability must sometimes be requested separately from the provider. Violations can be punished with fines of up to EUR 20 million or 4% of global annual revenue.

Furthermore, the obligations and fines under the GDPR apply to the “controller.” If an employee, managing director, or board member disregards the clear policy prohibiting the use of personal AI accounts, there is a risk that the employee, managing director, or board member in question will be deemed the “controller” under the GDPR and become subject to the GDPR's sanctions.

2. Trade Secrets in Third-Party Systems

The Trade Secrets Act protects information only if the company has taken reasonable measures to keep it confidential. If the company allows patentable ideas, software source code, formulas, calculations, plans, and forecasts, or customer lists to enter systems unsecured – where the provider is permitted to use the data for its own purposes – this calls into question the adequacy of the company's trade secret protection measures. In the worst-case scenario, the information loses its legal protection entirely, and claims against competitors or former employees based on the trade secret come to nothing. The use of AI must therefore be addressed as a regulatory issue in the company’s confidentiality policy.

3. Confidentiality, Liability, and Criminal Liability of the Management Board, Supervisory Board, and Management

Members of the management board and supervisory board of a stock corporation are legally obligated to maintain confidentiality regarding the company's confidential information and trade secrets. They may only entrust such information to a service provider if it is in the company's interest and the service provider is, in turn, bound by a confidentiality obligation. It is therefore problematic to enter a supervisory board proposal, a term sheet, or transaction documents into a system whose provider is permitted to use the information for its own purposes. The unauthorized disclosure of corporate secrets is a criminal offense. The penalty is up to one year of imprisonment; for publicly traded companies, it is up to two years. Intent is required, but it is sufficient that the perpetrator considers the disclosure possible and accepts the risk of it occurring. Anyone who is aware of the provider’s terms and conditions and nevertheless enters the documents will hardly be able to claim that they did not accept the risk of disclosure. For managing directors of a GmbH, there is a corresponding criminal offense. It is also a criminal offense to intentionally misrepresent the company's financial position, the compensation report, or information presented at the shareholders' meeting, punishable by imprisonment for up to three years. Anyone who adopts AI-generated figures for this purpose without any verification is, in any case, in breach of their duty of care.

In addition, the management board of members and managing directors are liable for damages resulting from inadequate organization of AI use. According to a ruling by the Higher Regional Court of Frankfurt am Main, this may also include fines imposed on the company. According to the case law of the Federal Court of Justice, management may rely on legal advice only if an independent and expert advisor was fully informed and management has reviewed the advice for plausibility. In our assessment, an AI result alone does not meet these requirements. However, AI can be used as a tool within the context of expert advice provided by a qualified professional – such as the legal department or an external law firm – if the responsible legal counsel has reviewed and personally approved the AI result.

Furthermore, there is a significant risk that a D&O insurance policy will invoke the exclusion for knowing breaches of duty and will not provide coverage if trade secrets were uploaded to and processed in private AI accounts in violation of a known internal policy.

4. Disclosures at the Annual Shareholders' Meeting

The information provided by the Management Board at the shareholders' meeting must be complete and accurate. If it is based on an erroneous AI result, the resolutions passed may be subject to change – according to prevailing opinion, even if the Management Board is not at fault. The same applies if a shareholder's question remains unanswered due to an input error. Every AI-supported response must therefore be reviewed by experts before it is presented.

5. Inside Information and Ad Hoc Disclosures by Issuers

For companies subject to the Market Abuse Regulation, the use of AI poses a risk, particularly under insider trading laws. The prohibitions apply to every employee who handles inside information, not just members of the management board and supervisory board. A term sheet, a management board presentation on a planned transaction, or preliminary quarterly figures often contain inside information. Anyone who enters such documents into an AI system regularly discloses the inside information to the provider. This is permitted only if the disclosure is necessary for the task at hand. The fact that an AI system makes the work easier is not sufficient justification. If the provider is permitted to use the data for its own purposes, the disclosure cannot be justified. Intentional violations are criminal offenses punishable by imprisonment of up to five years. Reckless violations are administrative offenses punishable by fines of up to EUR 5 million for the individual involved and up to EUR 15 million or 15% of annual revenue for the company. Careless use of the consumer version of an AI system strongly suggests recklessness.

Added to this is the requirement for ad hoc disclosure. An issuer may delay the publication of inside information only as long as confidentiality is ensured. If the information enters a system without contractually guaranteed confidentiality, the delay ends, and the information must be published immediately. Anyone who fails to do so risks a fine of up to EUR 2.5 million for the company or, depending on its revenue, 2% of annual revenue. As of 5 June 2026, interim steps in longer-term processes – such as negotiations regarding a corporate transaction – no longer need to be disclosed separately. However, they must still be kept confidential, and a violation is subject to a fine. If the ad hoc announcement itself is drafted using AI and contains inaccurate information, the issuer is liable to investors for damages unless it can prove that it was neither aware of the inaccuracy nor grossly negligent in failing to recognize it.

6. Client Confidentiality in the Legal Department

In-house counsel are subject to the same duty of confidentiality as attorneys in a law firm. They may disclose client confidential information to an AI provider only to the extent necessary for the provision of the service, and only if the provider has been carefully selected and has been bound to confidentiality text form (e.g., by email). Otherwise, they violate their professional duties and may be subject to criminal prosecution. When private accounts provided by AI providers are used, such a confidentiality obligation typically does not apply. In fact, the opposite is true, because AI providers reserve rights to the uploaded and prompted content in their terms of service.

7. Obligations of the Company as a Deployer Under the AI Act

Companies that use AI have specific obligations as so-called “deployers” under the EU AI Act. They must take measures to promote AI literacy among their employees. The easiest way to demonstrate this is through a documented training program. Since August 2026, transparency requirements have applied, for example, to chatbots and to AI-generated images, audio, or video recordings that appear realistic. Starting 2 December 2027, strict obligations will apply to high-risk systems – for example, AI that plays a role in hiring, promotion, or termination decisions, or that assesses customers' creditworthiness. Even if a company primarily uses AI from external providers for these purposes, the company may be considered a “provider” under the AI Act. However, routine contract and document work involving AI is, at the very least, not considered a high-risk case.

8. Shadow AI via Personal Accounts

The most common problem in practice is of an organizational nature. Employees use personal AI accounts for business purposes because the company does not provide its own system or because personal access is more convenient. Even a paid personal account remains a consumer version, and its terms and conditions then apply to the data entered. The company usually only finds out about this once damage has occurred. An AI policy that does not explicitly prohibit such services therefore fails to serve its purpose. It is equally important that the approved systems actually help employees.

IV. What Companies Should Do Now

In practice, the following points are regularly a priority:

  • Review provider contracts. Use only corporate or API contracts that exclude training with user input, strictly limit storage and access by the provider's employees, regulate confidentiality and data processing, and do not limit liability for confidentiality and data protection violations to a minimum. Retrospectively review contracts concluded with a single click. For each approved system, document who approved it, when, and on what basis.
  • Implement an AI policy. Specify which systems may be used and what information may be entered into them, for example, categorized as non-critical, confidential, or personal and strictly confidential. Explicitly prohibit the use of unauthorized services, including via personal accounts and personal devices. If there is a works council, the policy must be coordinated with it. Incorporate AI use into the confidentiality policy and review your own confidentiality agreements for non-disclosure provisions.
  • Secure inside information. For issuers, incorporate the use of AI into insider compliance procedures. Enter inside information only after individual approval and only into systems that do not allow access by the provider's employees and do not store data beyond the scope of processing.
  • Establish a review routine. Have AI results reviewed by knowledgeable employees on a risk-based scale; for critical documents, apply the dual-control principle, and document the review. Review chatbots that have direct customer contact before deployment, as the company is liable for their incorrect information just as it is for its own statements.
  • Implement data protection and AI regulations. Include AI incidents in the data protection incident response procedure; before approving a system, determine whether a data protection impact assessment is required; document a training plan; and classify potential high-risk cases by December 2027.

V. Conclusion and Outlook

The use of AI becomes critical where information flows unchecked to a provider or where AI results are used externally without verification. Those who have properly established the provider agreement, the company's internal AI policy, and the process for verifying results have addressed the key risks. For board members and issuers, this is not only a matter of due diligence and liability but also of avoiding criminal liability and substantial fines.

The legal framework continues to evolve, including with the Digital Omnibus on data law and the pending decision by the European Court of Justice on the Data Privacy Framework. Companies should therefore regularly review their provider agreements, internal policies, and review processes. We are happy to assist with the review and negotiation of vendor agreements, the drafting of AI policies, and the issues under stock corporation and capital markets law that arise for management boards, supervisory boards, and issuers when deploying AI.

Download as PDF

Contact persons

You are currently using an outdated and no longer supported browser (Internet Explorer). To ensure the best user experience and save you from possible problems, we recommend that you use a more modern browser.