On-premise AI for your M‑Files vault
When documents may not leave the building, an AI assistant is still possible. CtrlScout runs inside your vault, and the AI model behind it can be one your organisation runs on its own servers. It works on every M‑Files edition, in M‑Files Cloud or on-premise.

Where the model runs decides the answer
Every question an AI assistant answers goes to a language model, together with the document text it needs. Whether an assistant can live with a rule that nothing leaves the building comes down to where that model runs.
CtrlScout itself runs inside your vault, so with M‑Files on your own servers, it runs on those servers. The model is a separate choice your administrator makes for each vault: a European service, OpenAI, or a model your organisation runs itself.
Three places the model can run
What each choice means for where your content goes, and for what your own IT department looks after.
| Feature | Scaleway, the preset | OpenAI | A model you run |
|---|---|---|---|
| Who runs the model | Scaleway, hosted in the EU | OpenAI | Your organisation, on its own servers |
| Content is sent to | Scaleway's Generative APIs | OpenAI's own service | Your own server |
| What you provide | A key for the service | A key for the service | The server, the model and their upkeep |
The choice is made for each vault, and another OpenAI-compatible service can be chosen in the same way.

Answers from your documents, with sources
People ask in plain words, in their own language, from a panel beside their documents in M‑Files. CtrlScout finds the objects, reads the documents and answers with the sources it used, each one a click away.
- Summaries and comparisons — A long agreement summarised from start to finish, or two versions compared on what changed in the wording.
- Counts and charts — The vault does the counting, and the answer shows what was counted.
- The work that follows — A status email drafted from what was found, or a new object filled in for someone to approve.
Where the documents must stay in-house
Three situations in which the rules allow an assistant only with a model kept in-house.
A municipality's case files
Case files hold citizens' personal data, and the rule is that they stay in the municipality's own data centre. With a model running there, a caseworker can ask what has been decided on a case so far and read the answer with its sources.
Drawings under export control
A manufacturer's drawings and specifications fall under export control. An engineer asks which revision of a part's specification is approved and what changed from the one before, and the text goes to a model in the manufacturer's own server room.
Client files at a law firm
Each lawyer sees only the matters their M‑Files rights open, and CtrlScout reads the same way, so an answer draws only on matters the person asking may open. The vault of client files can use the firm's own model while another vault stays on the European service.

Inside the permissions you already keep
Every search, every document CtrlScout reads and every change it makes runs as the person asking, under their own M‑Files permissions, wherever the model runs.
- Changes wait for a yes — Where you allow changes, each one is shown first and made only when the person approves it, in their own name. A change can be undone.
- Conversations kept in the vault — A conversation is stored in the vault, and only the person who had it can open it again.
- Files from people's own computers — A file added to a question goes to the model with it. You decide whether to allow that.
What running your own model takes
Your organisation provides the model service itself: a model on a server you run, answering through the OpenAI-compatible interface that CtrlScout speaks. The model has to support function calling. That is how it asks CtrlScout to search and read, and CtrlScout checks every request against the vault before carrying it out.
Your M‑Files server needs to reach that service, and your administrator points CtrlScout at it in the vault's settings. Which model and which serving software to use is your organisation's choice. Running them is work for your own IT people: the server, the model, its updates and keeping it available.
Settle these before you choose
Our post on sovereign AI in document management lists the questions to put to any vendor. These are the ones to settle inside your organisation.
-
Step 1: Which vaults hold what the rule protects
The model is chosen for each vault, so the answer can differ from one vault to the next.
-
Step 2: Who gets the panel
Everyone in a vault, or only the user groups you name.
-
Step 3: Whether people may add their own files
A file someone adds to a question goes to the model along with it.
-
Step 4: Whether it may change objects
It can stay with reading and answering, or also change and create objects, each change approved first.
-
Step 5: Whether to search by meaning
That keeps an index of the passages you choose, which can sit in a store you run.
-
Step 6: Who looks after the model
Someone owns the server, the model's updates and any later decision to change model.
What it needs to run
- Any M‑Files edition, in M‑Files Cloud or on-premise, with M‑Files 25.1 or later on the server and in the clients.
- CtrlScout itself: one vault application, which brings its panel with it.
- An AI service for CtrlScout to work with: Scaleway Generative APIs, OpenAI, or any other OpenAI-compatible service, a model on your own servers among them. The model must support function calling.
- For suggested details on new documents, the optional CtrlScout Intelligence Service, which needs the vault's IML Core licence.
Questions about AI on your own servers
Can CtrlScout use a model that runs on our own servers?
Which models can we run?
What does CtrlScout send to the model?
Where are conversations and indexes kept?
Can CtrlScout read more than the person asking?
Can different vaults use different models?
Is there a trial?
More about CtrlScout
Everything CtrlScout does
Answers, charts, summaries, email drafts and suggested details for new documents.
See CtrlScoutSovereign AI, the reasoning
Why the place a model runs matters, and the questions to ask any vendor.
Read the postAn AI assistant for M‑Files
What people ask CtrlScout, and what comes back.
Read the post