All articles
12 min read

LLM Data Privacy for Swiss Businesses: What to Check Before You Deploy

A 1980s steel counter holding an open lockable document box, a paper shredder with a folder half-fed into it and a small amber indicator lamp, a brass key, and a capped pneumatic tube canister.

LLM data privacy comes down to three questions: what goes into the model, where that data is stored and who can reuse it, and what comes back out. In Switzerland you answer them under the Federal Data Protection Act (FADP), which applies to AI-supported data processing like any other processing. The safest setup is rarely a single tool choice. It is a workflow built so that sensitive data never reaches places it should not, with contracts, access rules and monitoring around it. That is software design work, and it is easiest to get right when senior engineers shape it from the start.

TL;DR

  • LLM privacy risk sits in two places: the model itself (memorization of training data) and the system around it (prompts, logs, retrieval from your documents, integrations).
  • Many AI tools can use your inputs to train future models unless this is contractually excluded, and data fed in cannot practically be taken back.
  • The FADP is technology-neutral: users have a legal right to know whether they are talking to a machine and whether their input is used to train it.
  • A “private LLM” limits exposure by keeping data inside a controlled environment, but it still needs access rules, logging and review.
  • Before launch, map the data flows, classify what the model may see, fix the contract terms and set up monitoring. Then repeat the review on a schedule.

Where LLMs actually expose data

Most teams picture one risk: an employee pastes a customer list into a chatbot. That risk is real, but it is only one of several, and they need different fixes.

The Council of Europe’s research on LLM privacy separates model-level risks from system-level risks. Model-level risks include personal data ingested during pre-training, memorization and regurgitation. System-level risks include persistent user profiling through memory features, unclear processing, leakage across contexts and weak user controls. Different safeguards apply to each category, which is why a single “we use a secure model” statement rarely holds up.

Risks that come from the model

LLMs can memorize personal details such as names, phone numbers, email addresses and email signatures. The same Council of Europe report notes that memorized sequences are very likely to show up in outputs, either repeated verbatim by accident or pulled out deliberately through training data extraction attacks.

For most Swiss businesses using a hosted model, you do not control this layer. What you control is what you add on top: the data you send, and whether the provider may learn from it.

Risks that come from your own workflow

This is where most business exposure happens. Walder Wyss’s data protection team, in a report on LLM risks and measures, lists application-level problems: leakage, reconstructability of inputs, and disclosure of sensitive content through retrieval-augmented generation (RAG). RAG means the model looks up passages from your own documents before answering. If the document index contains HR files or contracts and access rules are loose, any user of the assistant may be able to coax those passages out.

Prompt injection adds another path. Swisscom describes how a cleverly worded input can bypass a model’s safety guardrails and be used to extract confidential data, including the sources the model was given. An assistant connected to your CRM or shop is only as private as the permissions you gave it.

The reuse problem

The Swiss Federal Chancellery’s cyber security guidance describes the greatest hidden risk plainly. Many AI tools can use input data to train future models unless this has been contractually excluded, and once information is in the system it cannot in practice be retrieved and may surface in responses to other users. KPMG Switzerland makes the same point: unless you explicitly restrict the provider from using your data, it may be used to answer someone else’s prompt, and heavier use raises the chance that others reach your confidential information.

A well-known example is Samsung, where employees submitted proprietary data to ChatGPT, which led to a breach and forced the company to restrict internal use.

What Swiss data protection law expects

You do not need a new AI law to be on the hook. The Federal Data Protection and Information Commissioner (FDPIC) states that the FADP is technology-neutral and covers AI-supported processing, and that manufacturers, providers and users must give affected people the greatest possible control over their data (FDPIC, AI and data protection).

In practice, three points matter for a business using an LLM:

  1. Transparency. Users of language models that communicate directly have a legal right to know whether they are dealing with a machine, and whether what they enter is processed to improve self-learning programs or for other purposes. If you put a chatbot on your website, say so, and say what happens to the input.
  2. Purpose and sources. The FDPIC and the National Cyber Security Centre (NCSC) stress that providers and users must be open about the purpose, functioning and data sources of their AI applications.
  3. Training data. The FDPIC, with other international authorities, has published guidance in 2023 and 2024 on data scraping for training language models, obliging companies to follow data protection rules and set safeguards against illegal scraping. This matters most if you fine-tune a model on scraped or third-party data.

In October 2026 the FDPIC and NCSC ran a joint Cyber Security Month campaign on generative AI at work, with questions and answers on using AI tools consistently with the FADP. It is the most current official starting point for a Swiss team.

If your workflow touches health data, employee files or other sensitive personal data, have a Swiss data protection lawyer review the design before launch. The checks below prepare that conversation, they do not replace it.

What to check with an LLM provider

Provider terms decide whether your inputs stay yours. PwC Switzerland’s legal guidance on procuring AI tools says agreements should guarantee the company ownership of its data, include data protection provisions that reduce the risk of unauthorized access, and bar the supplier from using your data except for improvements that solely benefit you. It also recommends a warranty of compliance with AI regulation, backed by an indemnity.

Turn that into a checklist you can put in front of any vendor:

Question What a good answer looks like
Is my input used to train models? Contractually excluded, not just a settings toggle
How long are prompts and outputs retained? A stated period, with a zero-retention option if the data is sensitive
Who at the provider can read my data? Named roles and a logged, limited access process
Where is data processed and stored? Stated regions and subprocessors, matching your own disclosure duties
Who owns outputs and inputs? You, in writing
What happens on a breach? A notification duty and an indemnity

Zero data retention options exist from the large providers, but they typically need approval or a sales arrangement, and not every model is eligible. Treat retention as something you negotiate and document, not a default you can assume.

What a private LLM for business means

A “private LLM” is a language model that runs in an environment you control, such as your own cloud tenancy or servers, instead of a shared public service. The main reason enterprises choose it, according to InformationWeek, is data security and control over what the model sees.

What it limits:

  • Prompts and documents do not leave your environment for a third party’s training pipeline.
  • You decide retention, logging and who may query the model.
  • You can keep the model separate from systems it does not need.

What it does not remove: a private model still memorizes whatever you fine-tune it on, still follows prompt injection if you let it, and still needs access control on the documents behind RAG. “Private” describes where the model runs. Privacy is the result of how the whole workflow is designed.

For many teams the right answer is not a fully self-hosted model but a hosted enterprise model with a no-training contract, plus a thin layer of your own software that strips or masks sensitive fields before anything is sent. That layer is ordinary application engineering, and it is where most of the privacy value is created.

Safeguards that work in practice

The Federal Chancellery’s guidance names four moves: clear internal guidelines and policies, staff training, contractual guarantees that enterprise solutions will not train on your data, and technical controls such as data loss prevention (DLP) tools, which detect and block sensitive data leaving your systems. PwC adds a governance structure with a multi-disciplinary team of data scientists, ethicists and legal experts.

Here is how to make each of those concrete.

Classify before you connect

Decide three tiers of data and write them down:

  • Never goes to an LLM: health records, payroll, credentials, client secrets under professional confidentiality.
  • Only to an approved, contract-protected model: customer correspondence, internal documents, source code.
  • Free to use: public marketing copy, published documentation.

Staff follow a three-line rule far better than a twenty-page policy.

Minimize what the model sees

Send the model the smallest slice of data that does the job. Replace names and account numbers with placeholders before the call and map them back afterwards. Give a RAG assistant access to only the documents that the asking user is already allowed to read. Every extra field is an extra thing that can leak.

Control the integrations

An assistant wired to your shop, CRM or ERP inherits whatever those systems expose. Give it a service account with read-only, narrow permissions. Do not let model output trigger actions such as refunds or data exports without a check. This is where prompt injection does damage, because the attacker steers the model into using access you handed it.

Train people and set the rules

Short, specific training beats a general warning: which tools are approved, which data tier each may touch, and what to do after a mistake. Marketing and content teams have their own questions about ownership and disclosure, covered in this overview of AI content legal issues. Name an owner for the policy so it gets updated.

How to assess and monitor an LLM workflow

A privacy review is not a one-time gate. Models, provider terms and your own integrations change. A workable routine has a before-launch pass and a recurring pass.

Before launch

  1. Map the data flow. Draw every point where data enters, is stored, is sent to a provider and comes back: user input, prompt templates, retrieval indexes, logs, analytics, backups. Most surprises are in the logs and the indexes, not the model call.
  2. List the personal data involved. For each flow, note what personal or confidential data is present and the legal basis for processing it. Check that your privacy notice describes the use, including the transparency duties above.
  3. Fix the contract. Walk the provider checklist from earlier and keep the answers on file.
  4. Test with adversarial inputs. Try prompt injection, requests for other users’ data and attempts to extract the system prompt or document sources. Fix permissions, not just prompt wording, when something gets through.
  5. Define a stop condition. Decide in advance what incident means switching the feature off, and who can do it.

After launch

  • Log carefully. Keep enough to investigate an incident, but treat logs as personal data too: restrict access and set a retention period.
  • Watch for drift. Review a sample of outputs regularly for leaked identifiers, odd retrieved sources and refused-then-allowed requests.
  • Re-check on change. A new model version, a new connected system, a provider terms update or a new document source each reopen the data map.
  • Review on a schedule. A quarterly review of access rights, retention and vendor terms is a reasonable rhythm for most small teams.
  • Keep your incident path ready. Know who decides, who informs the affected people and who talks to the authorities if personal data was exposed.

Building it so privacy is designed in

The checklist above is mostly engineering: data minimization layers, permissioned retrieval, narrow integrations and monitoring. These are the parts that go wrong when a project is handed from a salesperson to a junior team and the requirements are rewritten halfway.

At Ampersand, a small Zurich studio, senior engineers are involved from the first consultation through to handover, so the people who hear your privacy requirements are the people who build against them. We build AI automation for business processes and system integrations for shops, CRM and ERP, and our clients include UBS and the City of Lugano. If you want an outside review of an existing setup, we also offer audits and senior-led technical consulting, and AI masterclasses for teams who need to learn the rules of safe use.

For ongoing work, two plans are published: Run & monitor at 750 CHF per month and Guidance at 799 CHF per month. Project scopes for automations and integrations are quoted per project. You can see the full list on the Ampersand pricing page.

FAQ

Is it safe to put customer data into ChatGPT or similar tools?

Only if the tool and plan you use contractually exclude training on your inputs, and the data tier allows it. Swiss federal guidance warns that inputs to many tools can be used to train future models and cannot practically be retrieved afterwards. Keep special-category data such as health records out of general-purpose tools altogether.

Does the Swiss FADP apply to AI and LLMs?

Yes. The FDPIC states that the FADP is technology-neutral and applies to AI-supported data processing. Users also have a right to know whether they are communicating with a machine and whether their input is used to train it, so your chatbot disclosures and privacy notice should say so.

What is a private LLM for business?

It is a language model run in an environment you control, rather than a shared public service, so prompts and documents stay under your retention and access rules. It limits exposure to third-party training and reuse, and it still needs permissioned document access and monitoring to keep data private.

Can an LLM leak the data it was trained on?

Yes. The Council of Europe reports that models can memorize names, phone numbers, email addresses and signatures, and that these can appear in outputs through accidental regurgitation or deliberate extraction attacks. This is why you should limit what you fine-tune on and what you send.

How often should I review an LLM workflow for privacy?

Review before launch, whenever the model, provider terms, connected systems or document sources change, and on a fixed schedule such as quarterly. Each review checks the data map, access rights, retention periods and a sample of outputs.

Sources

Talk to us

Have a project this touches on?

A free 10-minute call is the fastest way to find out whether we are the right studio for it.

Book a free 10-min call