Data flow sheet: where does the content go?
As of 4 October 2026 · every detail with original source and retrieval date · to print and file
Anyone assessing the use of an AI tool needs a verifiable answer to two questions: Where do prompts and the content shown go? And: Who decides — the user or the vendor? This sheet answers both questions for Biwak and for the four services we are asked about most often. Every detail comes from the linked document of the respective provider, with a date.
Biwak
| Where does the content go? | Via the signed-in Biwak model access to Microsoft Azure in the EU Data Zone, contracting party Microsoft Ireland; if Azure declines, to Prem AI: the contracting party is PREM SA in Switzerland, and according to Prem, processing takes place with partners in the EU.What is transmitted: input, the conversation context needed and the file content used. If an image fails with both, Scaleway in Paris reads it as a fallback. Otherwise only app versions up to 0.2.24 and a previously saved model choice, which the slider shows as “Custom”, still run on Scaleway. |
|---|---|
| Who manages access? | BiwakModel access is provided for your Biwak account and managed on the server side. |
| Training on your content? | No Biwak does not use your content for training (Terms § 2(2)). With Microsoft Azure, the data processing agreement (version of 22 May 2026) ties processing to providing the service; according to the documentation, neither the model makers nor other customers have access, and nothing is used for training; suspicious requests may be stored for abuse monitoring. Prem contractually obliges its partners neither to store input nor to use it for training, without technical safeguards. |
| Verifiable | Local tools and server-side model access Account and usage data for access and billing — see Review by your IT. |
| Source | Privacy policy, section 14 · Terms § 2 · DPA As of 23 September 2026. |
ChatGPT in the browser (Free, Plus, Pro)
| Where does the content go? | To OpenAI; conversations and uploaded files are stored on OpenAI servers. |
|---|---|
| Who manages access? | The vendor A custom or European model endpoint is not provided for. |
| Training on your content? | On by default The “Improve the model for everyone” switch is on for personal accounts and has to be switched off actively. Team, Enterprise and Edu are excluded by default. |
| Source | OpenAI Help: “Disable model training” Retrieved on 11 August 2026. |
Claude in the browser & Claude Code (personal accounts)
| Where does the content go? | To Anthropic — explicitly including Claude Code sessions via personal accounts. |
|---|---|
| Who manages access? | The vendor In corporate set-ups, detours via cloud platforms are possible; in normal operation, the endpoint is fixed. |
| Training on your content? | A choice, set to on by default Since September 2025, personal users choose via the “Help improve Claude” switch; anyone who leaves the default accepts retention of up to five years. Business and API accounts are excluded. |
| Source | Anthropic: “Updates to Consumer Terms” Retrieved on 11 August 2026. |
Microsoft 365 Copilot
| Where does the content go? | Into the Microsoft 365 service boundary; according to Microsoft, EU traffic stays within the EU Data Boundary. Current exception: Anthropic models as subprocessors run outside the EU Data Boundary. |
|---|---|
| Who manages access? | The vendor The admin can allow or block additional models (OpenAI, Anthropic); a custom or European model operator is not provided for. |
| Training on your content? | No According to Microsoft, prompts, responses and Graph data are not used to train the foundation models. |
| Source | Microsoft Learn: “Data, Privacy, and Security” Document as of 9 July 2026. |
OpenAI Codex (cloud agent)
| Where does the content go? | For every task, the repository is loaded into a sandbox on OpenAI infrastructure (“preloaded with your repository”). |
|---|---|
| Who manages access? | The vendor |
| Training on your content? | The data rules of the OpenAI account in use apply See the ChatGPT row above: on by default for personal accounts, business accounts excluded. |
| Source | OpenAI docs: “Cloud environments” Retrieved on 11 August 2026. |
What this sheet does not claim
This sheet does not claim that the services listed are insecure. Enterprise contracts can exclude training, and Microsoft documents its data boundary carefully. The question of this sheet is a different one: Who decides where your content goes? For the four services above, the answer is the vendor — together with its contracts, its jurisdiction and its changes. With Biwak, model access is managed by Biwak. The model providers involved and their processing therefore belong in your assessment too.
Providers change their rules. If you find a detail that is no longer accurate, write to us — we will correct it and note the new date. We give answers you can pass on in writing.
This version as of: 4 October 2026