Before a model reads a single line, we clarify which data it may see and where it runs. There are three routes. The direct one: the frontier models from Anthropic, OpenAI or Mistral via their business offerings, which do not use inputs for training. The shielded one: large models such as Claude via Amazon Bedrock, in your own AWS account and in European data centres. There, the model provider has no access to your prompts and responses. And the local one: open models on your own server or with a Swiss provider, where nothing leaves the building.
Which route fits depends on the data. For drafts, summaries and information without sensitive content, the direct route is usually enough. For customer data and contracts we often recommend Bedrock, because you get the performance of the large models and still determine where processing takes place. For personnel files or patient data, the local model is the safest choice.
That includes clear permissions and a human who approves. With us, an AI may prepare, sort and draft. What gets sent, booked or deleted is your decision, until you trust the process enough to let it run on its own.