Small companies usually get a safe AI rollout from three kinds of help: an AI consultancy that sets up the tools and trains staff, a data protection lawyer for contracts and legal duties, and the IT provider for accounts and access. The AI policy itself rests on four rules: banned data, approved tools, a human review step and client confidentiality.
This article is general guidance for companies of 10 to 100 people in Switzerland and the EU, not legal advice. For a binding answer about your client contracts or a regulated activity such as health or financial services, ask a data protection lawyer.
What goes in an AI policy for a company of 10 to 100 people?
An AI policy for a company of 10 to 100 people fits on two pages plus a one-page tool list. It has to answer the two questions staff face at their desk: which tool may I open, and what may I paste into it? A twenty-page document gets skimmed once and forgotten, so keep each part of the outline below to two to five sentences in the real policy, written in the words your team uses.
- Scope: everyone who works for the company, freelancers included, on company devices and on personal phones used for work.
- Approved tools: a named list with the account type for each, for example the company's ChatGPT Business or Microsoft 365 Copilot workspace, and a rule that free or personal accounts are never used for company work.
- Data that never goes in: the red list in the next section, on every account.
- Data allowed only in approved business accounts: client names, internal documents, draft proposals and customer emails.
- Human review: AI output is a draft, and a named person checks it before it reaches a client or feeds a decision about a person. The policy also says what happens when AI makes a mistake with a client.
- Client confidentiality: check client contracts and NDAs before client material goes into any AI tool, and state how the company tells clients it uses AI.
- Connecting AI to files, email or meeting notes: allowed only after someone has checked who can see which folders.
- Owner: one named person approves new tools and receives reports of mistakes. Questions go to the same person, and reporting a mistake carries no penalty.
- Review date: the tool list is checked every quarter and the whole policy once a year, because vendor terms and the law both change.
What data should employees never put into ChatGPT or similar tools?
Employees should never put passwords, access keys or banking logins into ChatGPT or any similar tool, on any account. The same applies to sensitive personal data as defined in Article 5 of the Swiss data protection act, which includes health, religious or political views, genetic and biometric data, criminal proceedings and social assistance, unless a lawyer has approved that specific use. HR files on pay and performance, board papers, unpublished financial results and client documents under an NDA or professional secrecy belong on the red list too, unless the client contract and the policy explicitly allow an approved business tool.
Say a project lead pastes a client's 30-page supply contract into a personal ChatGPT account to get a summary before a 2 pm call. The summary helps, but the contract now sits with a vendor the company has no data processing agreement with, under consumer terms the client never agreed to. On a business account set up by the company, the same request can be fine if the client contract allows it. Business terms from the large vendors also exclude training on your inputs: Anthropic's commercial terms state that Anthropic "may not train models on Customer Content from Services." The account type changes the answer for client documents, while the red list stays red on every account. [Anthropic]
Does Swiss data protection law (revFADP) apply when staff use ChatGPT?
Swiss data protection law applies to AI tools the same way it applies to any other software. The Federal Data Protection and Information Commissioner (FDPIC) states that the revised Federal Act on Data Protection (revFADP, also called nFADP), in force since 1 September 2023, "is directly applicable to AI-supported data processing." When staff paste customer or employee data into an AI tool, these articles of the act carry most of the weight: [FDPIC] [Fedlex]
- Article 9: the AI vendor acts as a processor, so you need a contract with it, you must satisfy yourself that it can keep the data secure, and a statutory or contractual duty of confidentiality can forbid handing the data over at all.
- Article 16: personal data may go abroad only to a country the Federal Council has found adequate, or with safeguards such as standard data protection clauses approved by the FDPIC. Check where each vendor processes data and which safeguard covers it.
- Article 19: your privacy notice must name the recipients, or categories of recipients, of personal data and, when data goes abroad, the country it goes to.
- Article 21: when AI alone makes a decision with a legal consequence or a considerable adverse effect for a person, you must tell that person, and they can ask for a human to review it.
- Article 22: processing likely to create a high risk for the people concerned needs a data protection impact assessment beforehand.
The FDPIC adds a transparency point that matters for any website chatbot: users have a legal right to know whether they are speaking or corresponding with a machine. A Swiss company running a customer-facing chatbot should label it as AI even when no EU customer ever uses it.
Does the EU AI Act apply to a small Swiss company that only uses ChatGPT or Copilot?
The EU AI Act reaches a Swiss company when the output of its AI system is used in the EU, under Article 2 of the act, so a Swiss firm that sends AI-drafted work to EU clients can be in scope. For a company that only uses off-the-shelf tools for office work, two duties usually matter. The first is AI literacy: Article 4, rewritten in 2026 by Regulation (EU) 2026/1744, requires providers and deployers to "take measures to support the development of AI literacy" of their staff, and it does not require any specific level of AI literacy for each person. A training session tied to the policy, with a record of who attended, is a sensible way to show those measures. The second duty is transparency for chatbots and AI-generated content, covered in our guide to what the EU AI Act requires of a small business. [AI Act Service Desk] [AI Act Service Desk]
How do you roll out AI tools safely in a small company?
A safe rollout starts with the tools people already use. Staff may already run ChatGPT or Claude on personal accounts, sometimes for client work. Ask openly which tools people use and for what, and promise no penalty for honest answers. Then move that work onto company business accounts of the tools people already like, which is usually faster than introducing a new tool nobody asked for. Write the policy around those real uses, and train the team inside real weekly tasks, so the rules show up at the moment someone is about to paste a document.
Connecting AI to company files is where a quiet mistake hides. An assistant that reads your shared drive or meeting notes sees whatever the person using it can see, so old sharing settings suddenly matter. Microsoft states that Copilot "only surfaces organizational data to which individual users have at least view permissions", which also means it surfaces everything those permissions allow. So before you connect a meeting-notes tool or a shared drive to an AI assistant, check who can open each folder, and move recordings of conversations about pay, hiring or ownership into a folder only the owners can open. [Microsoft]
Who helps small companies roll out AI tools safely?
In a company of 10 to 100 people, an AI partner and a data protection lawyer do most of the work in a safe AI rollout. Anovis AI, headquartered in Basel, is one such partner: it joins the client's team, sets up the approved tools, drafts the internal AI usage policy and trains people to work with them. Anovis AI gives advisory, not legal counsel, so the legal sign-off on vendor contracts, data transfers abroad and client confidentiality goes to a data protection lawyer or an outsourced data protection officer (DPO) service. The IT provider handles accounts and access. In the table below, the middle column says what each kind of help is good for and the right column says what it usually does not cover.
| AI consultancy | Good for: choosing tools, setting up business accounts, writing the policy around real tasks and training staff inside their own work | Usually not: binding legal opinions, which still need a lawyer |
| Data protection lawyer | Good for: processor contracts, data transfers abroad, privacy notices, professional secrecy and regulated work such as health or finance | Usually not: tool setup, staff training and day-to-day habits |
| IT provider | Good for: managing accounts and access rights, multi-factor login, blocking unapproved sites and checking security settings | Usually not: deciding which tasks to use AI for or writing the policy |
| Free policy template | Good for: a fast first draft to react to | Usually not: your own tools and client contracts, so someone still has to adapt it |
In Switzerland, the law firm VISCHER publishes a free one-page AI policy for employees, a useful first draft for any company. Treat it as the free template row in the table: someone still has to adapt it to your tools and your client contracts before staff sign it. [VISCHER]
A simple rule for choosing: if your main risk is legal, such as health data, financial services, professional secrecy or personal data of EU customers, start with the lawyer and add a consultancy for the rollout. If your main risk is habits, such as staff on personal accounts and nobody checking output, start with a consultancy or a trained internal owner and bring the lawyer in to review the vendor contracts. Ask any provider for a sample policy and a training plan before you sign, and read how to vet an AI consultant first.
Frequently asked questions
Should a small company simply ban ChatGPT at work?
A ban rarely holds, because the same tools sit one tap away on a personal phone, where the company sees nothing and has no contract with the vendor. A business account with a data processing agreement gives staff a safe default, and it gives the company a record of which tools are in use, which a ban never does.
Do we have to tell clients that AI is used on their work?
Check each client contract first. Some contracts and NDAs restrict sharing material with outside services, and under Article 9 of the Swiss data protection act a contractual duty of confidentiality can forbid passing data to a processor at all. Where personal data goes to an AI vendor, your privacy notice must name the recipients or categories of recipients and any country abroad. One sentence on AI use in your terms of engagement settles the question for new clients.
Does Microsoft 365 Copilot need different rules from ChatGPT?
Copilot works inside your files and email, so the extra rule is about access: before switching it on, clean up shared folders and old company-wide links, especially those holding HR and finance files. Microsoft states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, so the risk to manage is oversharing inside the company.
How often should a small company update its AI policy?
Review the tool list every quarter and the whole policy once a year. Rules move faster than most internal documents: the AI literacy article of the EU AI Act first applied on 2 February 2025 and was rewritten in 2026 by Regulation (EU) 2026/1744.
Who should own the AI policy in a 20-person company?
One named person, often the operations lead or the managing director, who approves new tools and keeps the tool list. The owner needs no legal training, only a direct line to a lawyer or consultancy for questions the policy does not answer, such as a new client contract with strict confidentiality terms.
If your team already uses AI on client work without a written policy, send us the tools people use today and we will tell you which rule to put in place first.