Hybrid AI: external models such as Claude for selected groups, sensitive data kept protected
Hybrid here means: everyone works with an AI model that runs in a protected setting, and selected groups also get access to external models, for example from Anthropic, OpenAI or Google. This works if the external model is tied to separate apps that only selected groups can see, and if each of these apps only gets the documents and connections that belong there.
What does “using AI in a hybrid way” mean?
Many organisations want both: for patient, citizen or personnel data, the AI should run in a protected setting, in the EU, without training on the data. At the same time, individual teams would like to use the large models of the well-known providers, such as Claude from Anthropic, ChatGPT from OpenAI or Gemini from Google, for example for programming, texts for the public or research.
“Hybrid” in this case means: one AI model for everyone in the protected setting and, in addition, external models for selected groups, on one platform, with one sign-in and one log. It shares only the name with hybrid operation, where the model for everyone runs somewhere else.
How do you separate external models and sensitive data?
The separation works through three levers that act together:
| Lever | What it controls |
|---|---|
| Model per app | The external model is tied to a separate app, not to the general chat. |
| Groups per app | Only selected people and groups see and use this app. |
| Data per app | The app only gets the documents and connectors intended for its task. |
An example: the press office gets an app “Texts for the website” with an external model and the guidelines on language and design. The HR department continues to work with apps that use the model in the protected setting and are allowed to access the personnel files. The press office does not see these apps, and the HR department does not see the press office’s app if it does not need it.
What should you bear in mind?
- Inputs go to the provider: whatever people enter or upload in an app with an external model is processed by the external provider. Even with clear rules, something that does not belong there will end up there now and then in everyday work, such as a paragraph copied from a file. A short note in the app, a rule in the AI policy for employees and a careful choice of groups help.
- Storage and training: providers store inputs for different lengths of time, some not at all. Check for each provider whether it stores or logs inputs or uses them for training; more on this in Zero Data Retention.
- Location and law: if a provider based in the USA processes the data, the US CLOUD Act may apply, even with data centres in the EU. The provider belongs in the list of sub-processors and in the record of processing activities.
- Costs: external models are charged by consumption. Those who work with them a lot cause higher costs; a budget per group or an agreed volume keeps this predictable.
- Network: if the platform runs on your premises, the firewall must allow the connection to the provider.
How does this work with basebox?
Everyone works with the AI model in basebox, in our cloud on our servers in the Munich-East data centre or on your server. Your data does not train any model.
In addition, we set up apps with an external model, for example from Anthropic, OpenAI or Google; for this we use Google Vertex in Frankfurt. Your administrator decides which people and groups use such an app and gives each app its own documents and connectors. The rule is: sharing an app shares its documents, never your own access. Each person only sees what they are allowed to see anyway, and every call of a connector is recorded in the audit log.
This works in the cloud and with the server, provided your firewall allows the connection to the provider. You can order the use of external models as part of your licence; the costs are then included up to the agreed volume, and the terms are in your quote. What changes for data protection is shown in the comparison under Security.
Not legal advice: whether you may pass certain data to an external provider depends on the type of data, the contract and national or state law. Clarify this with your data protection officer. Information as of 15 October 2026.
Frequently asked questions
Is this the same as hybrid operation?
No. In hybrid operation, the platform runs in your own network and the AI model for everyone runs somewhere else. This is something different: everyone works with the model in the protected setting, and only selected groups also use an external model in separate apps.
Can a group with the external model access all documents?
Only if it is set up that way. The opposite makes sense: each app with an external model only gets the documents and connectors intended for its task. Sensitive collections stay in apps with the model in the protected setting.
What does using external models cost?
External providers charge by consumption, that is by the amount of text processed. At basebox you can order this usage as part of your licence; the costs are then included up to the agreed volume.
Are our inputs stored by the external provider?
That depends on the provider and the contract. With Google Vertex, for example, inputs are not used for training, but they are logged for abuse detection for up to 30 days. Check storage, training and place of processing for each provider separately.
Sources
This article was created with the help of AI, based on the sources listed. Responsible for the content: basebox GmbH. Not legal advice, no liability. Research as of 15 October 2026.