Basebox AI

Every reply carries a citation back to the document and passage it used. A clerk can open the source and read the exact sentence. That single feature decides whether an answer is usable in a hospital or an authority, where an unsourced claim is worse than no answer.

Staff pick a task and get a result. The library covers discharge letters, care reports, guideline lookups, audio transcription, translation, job adverts and summaries, plus hundreds of documented use cases across hospitals, municipalities, insurers and cancer registries. The empty chat window is still there, but nobody has to start from it.

An app is an instruction, a set of documents and a set of connections. Authors build one in minutes without writing code, then decide whether a group can only use it or also edit it. Sharing an app never shares the underlying data.

Basebox AI runs in a German datacenter on shared infrastructure, or on a server reserved for one customer. On the server route you can use basebox hardware or your own, place it in your room or in a datacenter, and run it with no internet connection at all. Features and security are the same on both paths, which makes the choice mostly about size and procurement rules. A self-hosted AI install with no outside connection is the version security authorities tend to ask for.

Connections to hospital and administration systems are gated by the user's own account. Someone opening a patient file through an app sees what they would see in the source system, nothing more. Identity comes from your directory, so leavers lose access when the account closes.

Developers get an OpenAI compatible endpoint with a personal key, which means existing tools and research scripts work with little change. Models behave like connectors: an administrator enables them, then an app picks one. A special model trained on your own tasks can be added on request.