Which data does the task actually need?
For a wording edit, AI often needs only a text excerpt. Names, email addresses, customer numbers and internal calculations are usually unnecessary for that task. Start by identifying what you can omit or replace with invented details.
- Public content: for example, a published product description. Use it to test the quality of a draft without uploading a customer file.
- Confidential business information: for example, purchase prices or draft contracts. Establish internally whether the specific service may process it, even if no person is named.
- Personal data: for example, contact details, applications or meeting notes. Establish the purpose, necessary information and data protection requirements before use.
ChatGPT, Claude and Copilot: which account and features?
Record the exact service and plan, together with the account being used. The following vendor information helps you assess the setup; it does not approve your use case.
- ChatGPT: Check that the work takes place in the intended company workspace. OpenAI states that business data in Business, Enterprise and Edu is not used for training by default. Also record which files and connected apps the feature can access and what retention applies.
- Claude: Distinguish the specific commercial offering from a personal account. Anthropic does not use commercial product inputs and outputs for training by default; explicitly submitted feedback or permission to use data can create exceptions. Include the feedback feature in your rules.
- Copilot: Identify the Copilot offering and organisational account being used. For Copilot in Microsoft 365, Microsoft describes both exclusion of organisational data from model training and storage of interactions. Review existing document permissions, web search and additional agents separately.
Four separate questions about data flows
Map the actual route: text or file → AI service → connected application, if any → storage of the answer. Include generated summaries: they may contain confidential information from the input.
- Training: May content be used to improve models, including through feedback or voluntary permission?
- Storage: What remains in chats, files or logs? Which retention periods and deletion options apply?
- Access: Which employees, administrators or provider staff can see the content?
- Transfer: Does another service receive data through a connection or web search? Which terms apply there?
A worksheet for your own decision
Follow the example’s structure for one task. Replace assumptions with a link to the contract, vendor documentation or an internally verified setting. Mark unknown details explicitly as unresolved.
- Task and data: desired output, necessary content and excluded data.
- Account and features: service, plan, workspace, permitted users and enabled connections.
- Evidence: contractual basis including processing arrangements, recipients and processing locations, training, retention, deletion and access, each with a review date.
- Decision: permitted, permitted subject to stated conditions, or deferred pending clarification; record the responsible person and what should trigger another review.
Turn the decision into everyday practice
Give the team a short instruction: which account, content and features to use, and whom to ask about exceptions. Walk through an example using that exact account. A new connection, data category or destination for answers should trigger another look at the record.
Have someone check answers before sharing them. Resolving data protection questions does not establish that a quotation is complete or that AI has copied its figures correctly.
What the Swiss FADP means for this decision
The Swiss Federal Act on Data Protection also applies to AI-assisted processing of personal data. The FDPIC highlights transparency and requires a data protection impact assessment in high-risk cases. A plan name or Swiss server alone does not address these requirements.
If you intend to use real personal data, have the responsible person or team resolve open questions about purpose, information for affected individuals, processing locations and any confidentiality obligations. The completed record makes that assessment specific; it does not replace it.