Artificial intelligence in business: legal risks and basic obligations

Organisations most often introduce artificial intelligence in two ways: by deliberately procuring a dedicated AI system or through the gradual addition of AI functions to tools already in use, such as office suites, document-management systems, video-conferencing platforms, monitoring systems, accounting software or customer-support tools.

At the same time, employees may independently use publicly available generative tools to draft text, summarise documents, translate, analyse data or create code. Such use can speed up work, but may also expose the organisation to risk if personal data, trade secrets, confidential documents or third-party protected content are entered into the tool.

The initial assessment should therefore cover systems, licences and suppliers as well as actual working practices. This makes it possible to determine which uses should be permitted, restricted, subject to additional review or governed by specific procedures.

The EU Artificial Intelligence Act distinguishes prohibited practices, high-risk systems, systems subject to specific transparency obligations and other uses carrying a lower regulatory burden. Classification depends on the system's intended purpose and the circumstances of use, not merely on the product name or general description.

The rules apply in phases. AI-literacy obligations and most of the original prohibitions have applied since 2 February 2025, while most other provisions have applied since 2 August 2026. Following the July 2026 amendments, the requirements and obligations for Annex III high-risk systems apply from 2 December 2027, and those for Annex I product-related high-risk systems from 2 August 2028. Separate transitional dates apply to certain new prohibitions and marking obligations, including 2 December 2026. The applicable date must be checked against the type of system and the organisation's role.

The role of the organisation

An organisation may be the deployer of a system, but in certain circumstances it may also assume the obligations of a provider. This may occur when it places the system on the market under its own name, substantially modifies it or changes its intended purpose.

Before implementation, the organisation should determine the system's purpose, the persons affected, the data processed, its role in decision-making and the respective legal positions of the organisation and supplier. A supplier's compliance statement is only one element of the assessment.

Generative tools can create text, images, sound, code and other content. Their output may contain inaccuracies, fabricated facts, bias or passages resembling protected works. It must therefore be reviewed before being used in legal, business or public communications.

Confidential and personal data should be entered only into approved tools with known processing settings and an appropriate contractual framework. In particular, the organisation should verify whether the supplier uses submitted data for model training, how long it is retained, where it is processed, who can access it and how it is deleted.

Where personal data is processed, data-protection law applies regardless of the system's classification under the AI Act. The organisation should determine the purpose and legal basis of the processing, the scope and retention period of the data, data-subject rights, security measures and any transfer outside the European Economic Area.

Special care is required where an AI system evaluates, ranks or profiles individuals or affects employment, access to education, credit, insurance, public services or the exercise of rights. In addition to the AI Act, data-protection, employment, equal-treatment, administrative-procedure and consumer-protection rules may apply.

For decisions producing legal or similarly significant effects, the actual role of automation and the possibility of human intervention should be established. Human oversight is effective only when the responsible person understands the system's limitations, can verify the relevant data and has the authority to change or stop the outcome and give reasons for doing so.

Certain systems must inform individuals that they are interacting with artificial intelligence, while labelling duties may apply to synthetic or manipulated content. The precise obligation depends on the type of system and content, the method of publication and the organisation's role.

Most organisations acquire an AI system as a service or as part of a broader information solution. Before contracting, they should obtain a clear description of its purpose and limitations, data sources and flows, security measures, subcontractors, oversight mechanisms and the documentation required to meet their own obligations.

The agreement should address permitted use, service levels, confidentiality, data processing and storage, intellectual property rights, incident reporting, cooperation with oversight, system changes, audit rights, liability and the return or deletion of data when the service ends.

Responsibilities are distributed between suppliers and users according to their actual roles. The organisation retains responsibility for its own choice of purpose, data, users and ways of including the system in the business process.

Documentation should reflect actual procedures. A register without an assigned owner, a rule without an approval process or training without clear instructions has limited value. In practice, employees should know which tools they may use, what data they may enter, when further review is required and to whom a problem should be reported.

• an AI-system register recording purposes, suppliers, users and key data
• rules of permitted and restricted use, especially for generative tools
• procedure for checking and approving a new system or significant changes to an existing system
• clearly defined responsibilities for procurement, legal assessment, IT security, data protection and business use
• rules for human oversight, reviewing outputs and informing people affected by the system
• a procedure for reporting errors, incidents, unreliable outputs or suspected unauthorised use
• training adapted to the role of employees and the risks of the tools they actually use
• regular checking of suppliers, settings, uses and regulatory changes

A more detailed legal, organisational and technical assessment is particularly important where the system processes sensitive or extensive personal data or biometric data, monitors behaviour, profiles individuals or contributes to decisions affecting their rights or opportunities.

Situations requiring additional review

The appearance of one of these elements is a signal for a deeper analysis. The final classification and necessary measures depend on the actual function of the system, the data, the legal effect and the context of use.

• systems related to employment, performance evaluation or termination of employment
• biometric identification, categorisation of persons or advanced analysis of video surveillance
• systems that affect access to public services, education, credit, insurance or other important rights
• automated decision-making with legal or comparable significant effect
• training or adapting the model on internal, confidential or personal data of the organisation
• public posting of synthetic content and systems that interact directly with users
• connecting AI systems with key business, security or infrastructure processes
• transfer of data to third parties or to countries outside the European Economic Area