Managed AI: what it covers

All

Managed AI takes model procurement off your hands: access, key and models come with the licence. That holds for chat, knowledge search, document OCR and image generation — not for web search. Both package variants, EU data residency and EU sovereign, cover the same functions. This article says where the line runs, when a key of your own beside it makes sense and how you control which models your users see.

What Managed AI provides

Chat — three quality levels plus Auto. The licence provisions the models, you pick the level. → Switch the model per task
Embeddings — one model for knowledge search, set up automatically; the setup wizard’s embeddings step drops out. → Step by step: configure a single workstation
Reranking — re-sorting search hits rides along on the licence’s chat models.
Document OCR — the cloud OCR converter used by import uses the package’s OCR model.
Image generation — the image generator renders with the package’s image model (currently Gemini Flash with EU data residency, Qwen Image with EU sovereign); the Enhance Prompt step runs through Managed AI as well while Only Managed-AI models → Images is on (see below).

Which image and OCR models your package contains is shown under Administration → Managed AI, below the level table, in the rows Image generation and Document OCR.

What Managed AI does not provide

Web search in chat — it needs a key of your own at Brave or Serper. Its key is a system setting rather than a model, entered under Administration; the licence cannot carry it at all. → Web search in chat

Silent, not loud: if web search is switched on but no key is stored, chat reports no error — the tool simply is not offered to the model. If answers come back without current web hits, check here first.

A key of your own beside it: recommended for images

The package’s image model covers simple motifs. For demanding images, frontier models such as Nano Banana Pro (Google Gemini) or GPT Image (OpenAI) deliver clearly better results. That is why we recommend a provider account of your own for image generation alongside Managed AI. For OCR it is optional. Step by step: → Set up a frontier image model with a key of your own

You conclude a DPA with such a provider yourself, and reference images must contain no personal data — PII protection does not check them. → Data protection with external image providers

Open Administration → Models and create a provider account of your own, quickest via Quick add. → Where do I get an LLM API key?
Under that account, enable the model with the image generation or OCR modality. → Admin tab: AI models / providers

The Managed AI row stays untouched: it is read-only and reconciled on every sync. Your own account sits beside it, not above it — your users pick which model to use in the image generator or the import dialog.

So it stays clear where a request goes, the model lists carry a label: Managed AI models are marked EU data residency or EU sovereign, models with a key of your own External provider (own key).

Allow Managed AI models only

Under Administration → Models, the Only Managed-AI models switch decides whether your users also see models with a key of your own — separately for Chat, Images and OCR. That way you can keep chat strictly on Managed AI and still offer a frontier model in the image generator.

With the switch on for a modality:

Model lists only show Managed AI models; providers with a key of your own stay configurable in the administration but are not offered.
Auxiliary steps stay on Managed AI. Prompt optimisation in the image generator (switch Images) and the vision pass on import (switch OCR) do not fall back to a model with a key of your own. If no suitable Managed AI model exists, prompt optimisation is visibly skipped and the image is generated from your original prompt; the vision pass is locked in the import dialog, with the reason shown.
For chat it is more than a list filter: no chat call of the installation goes to a provider with a key of your own — not from agents, workflows, research or rerank either.

Empty list: if no Managed AI model is usable for a modality, its list stays empty and the function does not work — the image generator then reports that there is no usable image model. In that case, check the switch and Administration → Managed AI.

Costs

Chat, embeddings, OCR and image generation via Managed AI draw on the same monthly budget. Once it is exhausted, they pause together until the first of the month. Models with a key of your own keep running: their usage does not draw on the Managed AI budget but goes on your bill with the respective provider. With an image account of your own, the image generator therefore keeps producing images even when the budget is exhausted.

For web search you pay Brave or Serper directly. But it is a tool of the chat and is only invoked while an answer is being produced: if chat pauses for lack of budget, web search rests with it; with a key of your own or a local model it stays available.

Data sovereignty

Managed AI models — for OCR and images too — follow your package’s routing: with EU sovereign EU-native providers only, with EU data residency processing in the EU.

Keys of your own sit outside that assurance. A frontier image model from a US provider leaves the sovereign path — which is exactly what the External provider (own key) label in the model list signals. If that should be ruled out for a function, switch on Only Managed-AI models for it.

Web search: every query goes to the search provider you picked — including whatever it contains. Before switching it on, check that this provider meets your sovereignty requirement; the assurances of your Managed AI package do not extend to it.