Notes · AI at work

What can go into Claude or ChatGPT?

Ask around your company who uses AI. Usually more hands go up than management expects. A quote gets reworded with ChatGPT, minutes get summarised with Claude, a customer email gets translated. Often on a private account.

That is no reason to panic. It shows the tools are useful. The problem is elsewhere. Nobody knows which data has left the building, and nobody has decided what is allowed.

The private account is the real risk

With a private account, each person decides on settings the company should be deciding. One example: Anthropic trains new models on chats from the private Free, Pro and Max plans if the person allows it in their settings. In that case, Anthropic says, the data is kept for up to five years, otherwise for 30 days. For its business offerings (Team, Enterprise and the API), Anthropic does not use inputs and outputs for training by default.

ChatGPT, too, has different terms for private accounts and business offerings. OpenAI says it does not use data from ChatGPT Business, Enterprise and Edu or from the API for training by default. I check the current version of the contract before every rollout.

And when someone leaves the company, their chats leave with them. You do not know what is in them and cannot delete anything.

A one-page policy is enough to start

You do not need a handbook. One page that everyone reads protects more than twenty that nobody reads. It answers four questions:

1. Which tools and accounts are allowed? For example Claude on the company account, no private accounts for company work.
2. Which data may go in, and which never? General and public information, yes. Passwords, health data and anything under professional secrecy, never. Customer data only in a tool approved for it.
3. Who checks the results? AI writes convincingly even when it is wrong. Anything that goes to customers is read by a person first.
4. Who do you report a mistake to? Anyone who pasted customer data by accident should be able to say so without fear. Only then can you react.

You can download a template for this here. It has blanks for your company name and the person in charge. It is not legal advice and has to be adapted to your company.

Download the template (Word)View (PDF)

Company accounts instead of bans

A ban pushes AI onto private phones. A company account that people like to use works better. For Claude that means a Team or Enterprise plan. You hand out accounts centrally, close them when someone leaves and decide what is allowed. If an employee rates an answer with thumbs up or down, Anthropic says it may use that chat for training. On the Team and Enterprise plans you can switch rating off for everyone.

What Swiss data protection law requires

The revised Federal Act on Data Protection (FADP) has applied since September 2023. For AI tools, three points matter.

• Disclosure abroad. Personal data may go to a country whose data protection the Federal Council recognises as adequate. For the US, since 15 September 2024 this only applies to companies certified under the Swiss-U.S. Data Privacy Framework. Whether your provider is one of them is shown in the US Department of Commerce list. Otherwise you need safeguards such as standard contractual clauses (Art. 16 FADP). I check this in the provider's contract.
• Record of processing activities. Using AI with personal data belongs in it as a processing activity (Art. 12 FADP). Companies with fewer than 250 employees are exempt, unless they process sensitive personal data on a large scale or carry out high-risk profiling (Art. 24 DPO). I recommend the entry anyway. It forces you to answer the questions properly once.
• Reporting incidents. A data security breach likely to result in a high risk for the people concerned must be reported to the FDPIC as soon as possible (Art. 24 FADP). That is why the policy names a contact for reports.

Where Claude processes your data

If you use Claude directly from Anthropic, including the Team or Enterprise plan, Anthropic says it stores the data in the US. Requests may be processed in data centres in the US, Europe, Asia or Australia. For many tasks that does not matter, for some it does. So we decide together which data stays out. Where processing should stay in the EU area, there are other routes, such as Claude through a cloud provider in an EU region, or GPT models on Azure with the right deployment type. What that means on Azure is in the note Azure OpenAI: where your data is really processed.

What I do

• A workshop with your management. Afterwards you have a draft of your policy and know which accounts you need.
• I set up the company accounts and switch off what you do not want.
• I review data protection and write the entry for your record of processing activities.
• Training on your own tasks: how to replace names before pasting, and how to check results.