AI in the federal administration: What is permitted – and who is responsible?

AI is expressly encouraged within the federal administration

Artificial intelligence is transforming public administration. Language models assist with the analysis of documents, the drafting of texts and the processing of large volumes of information. AI-based applications promise more efficient administrative processes, shorter processing times and a reduction in the workload on staff.

The Swiss Federal Administration is taking a fundamentally technology-neutral approach in this regard. AI should not be prevented, but rather used responsibly.

With the report ‘Legal Framework for the Use of AI in the Federal Administration’, published on 25 September 2026, the Federal Chancellery has provided a tool to assist with the legal assessment of such projects. The report is supplemented by annexes and a checklist. Furthermore, an implementation plan for the AI strategy, including a roadmap for the years 2026 to 2028, was published on 18 September 2026.

The Federal Administration’s strategy is based on three areas of action: 1) building expertise, 2) earning trust and 3) increasing efficiency. This makes it clear that the use of AI is not only a technical task, but also a legal and organisational one.

The key message is: AI is not a legal vacuum. Existing legal requirements still apply even when administrative tasks are supported by AI.

No general authorisation for AI – and no general ban on AI

The new report is not a standalone AI law. In particular, it does not create a new general legal basis for the processing of personal data or the transfer of public authority tasks to AI systems.

Rather, the admissibility of a specific AI application depends on which administrative task is to be performed, which data is processed, and what impact the use of AI may have on the individuals concerned.

A distinction must be made between different usage scenarios.

The use of an AI system to edit the wording of a non-confidential internal text typically raises different legal issues to the automated analysis of citizens’ data or the AI-assisted preparation of an administrative decision.

The more an AI system intervenes in the actual performance of a task and the more significant its impact on the legal position of the individuals concerned, the more carefully the legal basis, procedural framework and control mechanisms must be examined.

Of particular relevance is the principle of legality under Article 5(1) of the Federal Constitution. State action requires a statutory basis. An internal administrative AI strategy cannot replace this.

However, this does not mean that every use of an AI system requires a specific, AI-related legal basis. If AI is used merely as a technical aid within the scope of an existing statutory task, the existing legal basis may suffice. This must be assessed on the basis of the specific intended use.

Data protection: Responsibility remains with the public authority

When federal bodies use AI systems, particular attention must be paid to the provisions of the Federal Act on Data Protection (FADP).

Under Article 34 of the FADP, federal bodies require a legal basis for the processing of personal data. Stricter requirements apply to personal data requiring special protection.

Generative AI in particular raises additional questions. If personal data is fed into an externally operated language model, it must be assessed, amongst other things, whether this data is used for training or improving the model, which subcontractors are granted access, and whether the data is transferred abroad.

The data protection assessment must therefore not be limited to the question of whether a provider offers a data processing agreement.

Rather, it must be clarified what data the provider actually processes, for what purposes this is done, and whether the provider pursues its own processing purposes. Article 9 of the Data Protection Act (DSG) on data processing on behalf of a controller and Articles 16 et seq. of the DSG on the transfer of personal data abroad are particularly relevant in this context.

In the case of AI applications that may pose a high risk to the privacy or fundamental rights of data subjects, the obligation to carry out a data protection impact assessment under Article 22 of the Data Protection Act must also be taken into account.

The decisive factor is not whether an application is designated as AI, but rather which processing operations and risks are associated with its specific use.

Information security: Not every AI system may process every type of information

Alongside data protection, information security is a key prerequisite for the use of AI in the Federal Administration.

The Information Security Act (ISG) governs the protection of the Confederation’s information and IT resources. The requirements are based in particular on the protection needs of the information and systems in question.

For AI applications, this means that it is not solely the technical performance of a model that is decisive. Equally relevant are the security architecture, the provider’s access rights, the processing and storage of input data, and the integration of other technical components.

In the case of external AI services, for example, questions arise regarding access to prompts, the storage of conversation histories and the integration of third-party providers. Added to this are specific AI security risks such as prompt injection, the unintended disclosure of confidential information, or the manipulation of AI-generated results.

A particular challenge arises when AI systems are connected to internal documents, databases or specialist applications. Such integrations can significantly increase the benefits of the technology, but at the same time create new points of vulnerability.

From a legal perspective, a risk-based security assessment is therefore required, taking into account both the information used and the technical functions and authorisations of the AI system.

The Federal Chancellery expressly states that existing information security requirements must be complied with at all times when using AI.

Official secrecy: Confidential information must not be fed into AI systems without proper control

Another key aspect concerns official secrecy.

Article 320 of the Swiss Criminal Code (StGB) protects secrets entrusted to members of public authorities and public officials in the course of their official duties, or which they have become aware of in their official capacity.

If such information is entered into an external AI system, this may constitute a legally relevant disclosure. Whether a breach of official secrecy relevant under criminal law has actually occurred depends on the specific circumstances, in particular on the confidential nature of the information, the access rights of third parties and any authorisation for disclosure.

The mere fact that an AI provider guarantees contractual confidentiality does not conclusively answer these questions.

Rather, the assessment must take into account the actual technical and organisational means of access. These include, in particular, support access, administrative rights, subcontractors and potential access by foreign authorities.

This does not mean that the processing of confidential information by external AI providers is generally ruled out. However, it does require a thorough legal and technical assessment.

Who is liable for AI errors?

Of particular interest is the question of responsibility for AI-generated results.

The Federal Chancellery expressly states that responsibility for the generated content remains with humans.

This statement should be understood as an organisational principle. It does not mean that legal responsibility lies exclusively with the individual operating the AI in every case.

Rather, a distinction must be made between different levels of responsibility.

The competent authority remains responsible for the lawfulness of its administrative actions. Its staff must ensure that AI-generated results are not adopted without verification, insofar as the task in question requires such checks. In addition, contractual or non-contractual liabilities may arise on the part of manufacturers and service providers.

In the event of erroneous administrative decisions, the focus is initially on the remedies available under administrative law. The Liability Act (VG) is particularly relevant to any claims for damages against the Confederation.

The use of AI therefore does not automatically alter the legal framework of liability.

In particular, a public authority cannot absolve itself of responsibility on the grounds that an erroneous result was generated by an external AI system.

Automated administrative decisions: where is the line drawn?

A particular legal challenge arises when AI systems do not merely process information, but prepare or automatically make administrative decisions.

Article 21 of the Federal Data Protection Act (DSG) contains specific provisions for automated individual decisions. In the case of federal bodies, the requirements under constitutional law and administrative procedural law must also be observed.

In particular, the procedural rights of the persons concerned must be safeguarded. Depending on the procedure, these include the right to be heard under Article 29(2) of the Federal Constitution, the provision of reasons for decisions, and the possibility of effective judicial review.

The use of AI must not result in the actual basis for decisions becoming unclear or in the competent authority effectively delegating its decision-making responsibility to a system whose workings cannot be traced.

A distinction must be made between mere decision support and a genuinely automated decision.

If an AI-generated proposal is independently reviewed and assessed by a competent person, this is to be classified differently in legal terms from a procedure in which the system makes the decision without substantial human involvement.

Formal human approval is not sufficient in every case to qualify a decision as having actually been made by a human.

What is decisive, rather, is whether an independent and proper review takes place.

Public procurement: AI compliance starts with the contract

The new legal framework is also relevant for companies that supply AI systems to the public sector.

When procuring such services, public authorities must not only comply with the general requirements of public procurement law. They must also ensure that the procured systems can be operated in compliance with the law within their intended context of use. This requires an analysis of the legal basis.

For federal procurement, the Federal Public Procurement Act (BöB) and the associated Ordinance (VöB) are particularly relevant.

In practice, the requirements relating to data protection, information security and the auditability of AI systems should be taken into account as early as the stage of defining the subject matter of the procurement and the technical specifications.

Of particular importance are contractual provisions concerning the use of input data, the further development and updating of models, subcontractors, security incidents, and rights of audit and participation.

The question of whether, and to what extent, a supplier can technically limit the processing of data by its models must also be clarified both contractually and in practice.

This has an important implication for providers of AI systems: a high-performance AI application alone is not enough. It is equally crucial whether its legal and technical characteristics allow for its use in the public sector.

What private companies can learn from the report

Although the Federal Chancellery’s report is addressed to the Federal Administration, its considerations are also relevant to private companies.

Admittedly, the public-law requirements cannot be applied to private companies in their current form. In particular, the legal frameworks governing the processing of personal data by federal bodies and private data controllers differ.

Nevertheless, numerous comparable questions arise.

Private companies, too, must assess what information may be processed in AI systems, what risks arise from the use of external providers, and how responsibility for AI-generated results should be organised.

The Federal Chancellery’s documents can therefore serve as a guide for developing a company’s own AI governance framework.

Such governance should not be limited to a general AI policy. Rather, specific use cases must be identified, their risks assessed, responsibilities defined and appropriate technical and organisational control mechanisms introduced.

It is particularly important to distinguish between the approval of an AI product and the approval of a specific use case.

An AI system may be suitable for generating general texts, whilst its use for processing confidential customer data or for preparing legally significant decisions requires additional measures.

Conclusion: The use of AI is a matter of governance

The legal framework published by the Federal Chancellery on 25 September 2026 highlights a fundamental approach to Swiss AI policy: the use of artificial intelligence should be facilitated without abandoning existing legal safeguards.

A standalone, comprehensive AI regulation is not necessary for every use case. Data protection law, information security law, administrative procedure law and other legislation already contain essential requirements for the legally compliant use of AI.

The real challenge therefore lies not so much in the absence of legal rules as in their application to new technical functions and operating models.

For public authorities, this means that AI projects must be supported from the outset in legal, organisational and technical terms. Corresponding detailed documentation is necessary. For AI providers, the requirements regarding transparency, security and contractual traceability are increasing. Private companies can use the tools developed by the Federal Government as a guide for their own AI governance, in particular the checklist in Annex C.

Anyone wishing to use AI responsibly must not only know what the technology is capable of, but also what data it processes, who controls it and who is accountable for its results.

Sources

• Federal Chancellery – Use of AI in the Federal Administration. Contains the report of 25 September 2026, including Annexes A and B, as well as the separate checklist (Annex C).
• Federal Data Protection Act (DSG). In particular, Articles 9, 16 et seq., 21, 22 and 34 of the DSG.
• Federal Administration’s AI Strategy and Implementation Plan 2026–2028.