Before You Roll Out AI in Your Office: Your Data Protection Adviser's Questions
A practical checklist under the Swiss FADP for AI rollouts: impact assessment, processors, transfers abroad, logs, retention and access rights.
Privacy & Law8 min read
Most AI projects in small firms do not fail on technology. They stall when the data protection adviser, the external auditor or a large client asks: where is the data, who can access it, and where is that written down?
Answering those questions before the rollout saves you from piecing it together afterwards. The checklist below follows the revised Swiss Federal Act on Data Protection (FADP) and the Data Protection Ordinance (DPO). It is general information, not legal advice.
The example throughout is an AI assistant that answers questions from the firm's own files: engagement letters, file notes, correspondence, tax documents. Technically this is often called RAG (retrieval-augmented generation): the system first finds relevant passages, then a language model writes an answer from them.
1. What exactly is processed, and why
Your adviser will first want to know which data goes into the system. That sounds trivial and is not. An assistant that indexes the whole file server also processes medical certificates in HR files, criminal record extracts in client files or details of trade union membership. These are sensitive personal data under Art. 5(c) FADP.
Write down:
- which folders and systems are connected and which are deliberately left out (HR files, for example),
- what the assistant is for, such as "research in the firm's own files for staff", and what it is not for,
- whether data about clients, employees or third parties is involved.
This belongs in the record of processing activities under Art. 12 FADP. Under Art. 24 DPO, companies with fewer than 250 employees are exempt unless they process sensitive personal data on a large scale or carry out high-risk profiling. Many law and fiduciary firms keep the record anyway, because it collects the answers to every other question on this list.
2. Is a data protection impact assessment needed
Art. 22 FADP requires a data protection impact assessment (DPIA) where processing may lead to a high risk to the personality or fundamental rights of the people concerned. The Act gives as examples the use of new technologies, extensive processing of sensitive personal data and systematic monitoring of public areas.
An AI assistant over client files can cross that threshold, but does not have to. What matters is the volume and type of data: an HR folder weighs more than a template library. If the DPIA shows a high residual risk despite the measures taken, Art. 23 FADP requires you to consult the FDPIC beforehand.
In practice: even if you conclude that no DPIA is needed, write the reasoning down on one page. It is the first document anyone will ask for.
3. Who processes data on your behalf
Every external party with access to personal data is a processor under Art. 9 FADP. That includes the server host, the language model provider (if you use one in the cloud) and possibly the IT partner who maintains the system.
Questions you should be able to answer:
- Is there a data processing agreement with every party?
- Does the provider only process the data as you yourself would be allowed to, which rules out training its own models on it?
- Does a statutory or contractual duty of confidentiality prevent the disclosure? For professional secrecy holders this is the central question; see AI for Swiss law and fiduciary firms.
- May the provider use sub-processors, and only with your prior authorisation, as Art. 9(3) FADP requires?
If the whole system runs on your own server without a cloud language model, most of these questions fall away.
4. Does data go abroad
Under Art. 16 FADP, personal data may be disclosed abroad if the Federal Council has found that the destination provides adequate protection (list in Annex 1 DPO), or if suitable safeguards such as standard contractual clauses are in place. Art. 17 FADP sets out exceptions, such as explicit consent in the individual case.
Since 15 September 2024, US companies certified under the Swiss-US Data Privacy Framework are treated as providing adequate protection. That settles the data protection side of the transfer, but not professional secrecy and not access by US authorities under the CLOUD Act.
With a cloud language model, ask concretely: in which region is the request processed? How long does the provider keep inputs and outputs? And name the destination countries in your privacy notice, as Art. 19(4) FADP requires.
5. What is logged, and who sees the logs
An AI assistant produces logs: questions asked, passages found, answers generated. These logs often hold more sensitive information than individual documents, because they show who was interested in what.
Where sensitive personal data is processed automatically on a large scale, Art. 4 DPO requires logging of storage, modification, reading, disclosure, deletion and destruction. The logs must be kept separately from the production system for at least one year. The FDPIC has published a recommendation on this.
Whether or not that applies, decide:
- who may read the logs (not everyone with admin rights),
- that logs are not used to monitor staff performance,
- where the logs are stored. With some cloud services they sit with the provider.
6. When is data deleted
Personal data must be destroyed or anonymised as soon as it is no longer needed for its purpose (Art. 6(4) FADP). At the same time, retention duties apply, such as ten years for accounting records and vouchers under Art. 958f of the Code of Obligations.
For an AI assistant this means: when a document is deleted from the file server, it must also disappear from the search index. Ask the provider how quickly that happens and whether copies remain in caches or backups. Also decide how long questions and answers are kept.
7. Who may see what
This is where AI projects most often need rework. Art. 8 FADP requires appropriate technical and organisational security measures. An assistant that gives every employee answers from every file undermines the permissions on your file server.
Check:
- Does the system inherit existing file permissions, or are rights maintained separately?
- Are changes (someone leaves, someone changes team) picked up automatically?
- Can a staff member learn content from files they are not allowed to open by asking clever questions?
8. How do you handle access requests and incidents
Under Art. 25 FADP, people can ask whether and which data about them is processed. You should be able to say whether someone appears in the AI system, in the index and in the logs.
And if something goes wrong, for example an answer shows data to the wrong person: data security breaches likely to result in a high risk must be reported to the FDPIC as soon as possible under Art. 24 FADP. Decide who assesses that internally.
The checklist at a glance
| Topic | Legal basis | Document to have ready |
|---|---|---|
| Data categories and purpose | Art. 6, 12 FADP | Entry in the record of processing |
| Impact assessment | Art. 22, 23 FADP | DPIA or documented reasoning for not doing one |
| Processors | Art. 9 FADP | Agreements with host, model provider, IT partner |
| Transfers abroad | Art. 16, 17, 19 FADP | List of destination countries, safeguards |
| Logs | Art. 4 DPO | Logging concept, access rules |
| Deletion | Art. 6(4) FADP, Art. 958f CO | Deletion concept including the index |
| Access rights | Art. 8 FADP | Permissions concept |
| Requests and incidents | Art. 24, 25 FADP | Process for requests and notifications |
How KADARAG simplifies these questions
KADARAG runs on your own server. In local mode, no documents, questions or logs leave your network, so the points on transfers abroad and cloud processors largely fall away. In hybrid mode, only the question and a few short excerpts go to the chosen language model; documents, index and logs stay with you. Your file server's permissions also apply to the answers.
Next step
Go through the eight points with your data protection adviser before you choose a provider, and get the provider's answers in writing. If you would like KADARAG's answers, book an intro call.