What changes with agents?
A chat assistant returns text; an AI agent carries out steps: it reads data from several systems, makes intermediate decisions and writes results back. Data protection law does not change so much as the intensity of processing: more data, more systems, fewer human steps in between.
As of October 2026. This article is orientation, not legal advice.
The five points that matter for agents
1. Automated individual decisions (Art. 21 revFADP)
If an agent makes decisions without human involvement that have legal effects on a person or significantly affect them (e.g. rejecting an application), you must inform the person concerned. They can ask for a human to review the decision. In practice: agents propose, humans decide.
2. Data protection impact assessment (Art. 22 revFADP)
Where processing is likely to pose a high risk to the people concerned, an impact assessment is required beforehand. With agents that is often the case when personal data is processed at scale or sensitive data is involved. More in AI and Swiss Data Protection.
3. Data processors (Art. 9 revFADP)
If you use a third-party model or platform, that provider processes personal data on your behalf. This requires a contract covering purpose, data security and sub-processors.
4. Cross-border disclosure (Art. 16 and 17 revFADP)
If the agent runs through a provider outside Switzerland, you need a country with adequate data protection or suitable safeguards (e.g. standard contractual clauses). Check where model calls, logs and caches actually run.
5. Data security and traceability (Art. 8 revFADP)
An agent with write access needs a permission model and logs. Every action should be traceable so that access requests and notification duties can be met. Technically we implement this through clean tool interfaces, for example with MCP.
Rule of thumb: The less personal data and the more human approval, the simpler the compliance. Start with use cases that involve no personal data.
Checklist to get started
- Which personal data does the agent see, and does it really need it?
- Does it make decisions affecting people — or only propose them?
- Where do the model, logs and caches run (country, provider, sub-processors)?
- Are there processor agreements and, if needed, safeguards for cross-border disclosure?
- Is an impact assessment needed and documented?
- Are permissions and per-action logging defined?
Agent planned, data protection unclear?
In a free discovery call we place your use case and show which architecture makes compliance easier.
Request a Discovery Call