AI in DDM

DDM uses AI – more precisely, a language model – only in a few, clearly named places. This page lists them all: what goes to the model in each case, what comes back and who switches it on and off.

In short

• Only when you click. Every request to a language model starts with a deliberate action – a question, a click on Suggest, a command in the AI menu. Nothing runs automatically: not on upload, not on saving, not in the background or on a schedule.
• Only with your organisation's provider. DDM brings no model of its own and uses no provider in the background. Your organisation enters provider, model and key itself – until it has, nothing goes to an AI.
• Off until someone switches it on. The AI assistant and the AI apps start switched off. Only administrators and owners can switch them on.
• Nothing changes without you. What the AI suggests – a text, a translation, a change – DDM only applies once a person confirms it.
• No content stored. DDM stores neither questions nor answers nor the texts that went to the model – only how many requests there were and what they cost.
• Labelled. Answers and suggestions from the AI are recognisable as such.

Where the AI comes from

Your organisation chooses under Settings › AI assistant which model DDM asks:

• The data goes directly from the DDM server to this provider – nowhere else. Which rules apply there (storage location, retention, use for training) is set by your contract with the provider, not by DDM.
• If you do not want to give anything to an external provider, run a model in your own network and enter its address. The data then does not leave your network.
• The key is stored encrypted and never shown again. If the provider or address changes, DDM deletes it.
• DDM sends no details about the person asking with a request – no name, no e-mail address, no identifier.

Setting up and setting limits: see AI assistant.

Where AI is used

The AI assistant

Where: in the CMS through the Assistant button at the bottom right and the search; in the app through Assistant and – with a document open – through Ask.

What goes to the model when you send a question:

• Your question and the last steps of the conversation.
• Up to eight matching documents, with name, location, date and a short excerpt from the text. They are found with the normal full-text search – across the documents you are allowed to see. In the app, only documents you can open there qualify; drafts that are only distributed after approval do not go in with their text.
• Matching categories with the number of their documents.
• If you ask in the app about an open document, its text goes along – up to 60,000 characters, for longer documents the matching passages and the beginning.
• If an administrator asks in the CMS about people or groups, group names and the names of the members go along – no e-mail addresses. The app does not have this.

Owners and administrators see all documents in DDM – their questions can therefore draw on all documents.

What comes back: an answer, below it Files used, Categories and, where applicable, From the user guide. DDM compiles the sources itself from its own data; the model can only refer to them, not invent any. If the assistant suggests to an administrator that people be added to a group, that only happens after a click on Apply to …. Otherwise the assistant cannot change anything.

Answers from the user guide

If you ask how something works in DDM, the assistant can read matching sections of this guide and refer to them (From the user guide). This is not a separate request – the sections go to the model together with your question. Readers in the app only get sections about the app. Your organisation's data is not in the guide.

Switch on and off: Settings › AI assistant › Answer from the user guide.

AI in the note editor

Where: in notes in the app and in note templates in the CMS – in the AI menu above selected text (Spelling & grammar, Longer, Shorter, Simpler, Summarise, Translate) and with Continue writing (Ctrl/⌘ + J).

What goes to the model: only the selected text, with Continue writing the text before the cursor – a few thousand characters at most. DDM first replaces references to documents, images and placeholders with markers; they do not go along. What you have written into the selected text yourself – including names or customer data – does go along.

What comes back: a suggestion in a card of its own with the note Suggested by a language model – please check. It only goes into your note with Accept.

Switch on and off: Settings › AI assistant › AI in the note editor.

Translation suggestions

Where: in the CMS under Translations, with Suggest, Suggest selection and Suggest all missing. Administrators only.

What goes to the model: only names – of categories (with the location as context) or file names of documents – and the target language. No document content.

What comes back: suggestions, with a dashed border next to the row. Only what you confirm with Accept suggestion or Accept all is saved. The audit log records Translations suggested – with the number, without the names.

Translating system e-mails

Where: in the CMS under Settings › System e-mails, with Translate from … with AI. Administrators only.

What goes to the model: subject and text of the template in the other language – with placeholders such as {{ name }}, not with real recipients' data.

What comes back: a translation in the editor. DDM checks that all placeholders are kept and otherwise discards the translation. Nothing is saved until you save.

Testing the connection

Where: Settings › AI assistant › Test connection. Sends a fixed test request without any of your organisation's data, to check that provider, model and key are right.

AI apps

Through the MCP server you connect your own AI app – Claude, ChatGPT, Cursor and others – to DDM. Here DDM does not ask a model; your app asks DDM:

• Each person connects their app themselves and explicitly allows access. The app sees exactly what the person sees in DDM.
• What the app reads – documents including their text, search results, for administrators also people with e-mail address, groups and the audit log – goes to the provider of that app. You decide which apps your data-protection rules allow.
• An app cannot change anything directly: it suggests changes, and the person approves each one individually – in the app or in the CMS under AI actions. Carried-out and rejected suggestions are in the audit log (AI action carried out, AI action rejected).
• The AI apps are off until an administrator switches on Enable MCP server under Settings › AI apps. Connections are disconnected by the person themselves under My account and by administrators under Settings › AI apps.

More: MCP server › Administration & security.

How to recognise AI results

• Under every answer from the assistant it says how many documents it comes from and which model answered. Below the input field it always says: Answers come from a language model and can be wrong. The assistant changes nothing without your confirmation.
• Suggestions in the note editor carry the note Suggested by a language model – please check.
• Translation suggestions appear as a Suggestion next to the field until you accept them.
• Changes through AI apps are in the audit log as AI action carried out, with the app's name.

Language models can be wrong. Check answers against the sources named before you rely on them.

What DDM stores

Who switches what on and off

Only administrators and owners can change any of the four; every change is in the audit log. Under Spending limits administrators cap what the AI may cost per month – for the whole organisation and per person. If the assistant is off, readers in the app see no way into it.

Where no AI is involved

Some things seem "intelligent" but work without AI:

• Search – in Documents, in the app and in the assistant – is a classic full-text search. There is no semantic search and no vector database.
• Text extraction for search reads the text from the files themselves. There is no text recognition in images or scans (OCR).
• A document's language is detected by DDM from common words, not with a model.
• Preview images are made with classic tools.
• Document content is never translated – only names, e-mail templates and selected note text are translated, on request.
• There is no automatic tagging, classification, summarising or filing of documents.
• External sources, approvals, permissions and usage statistics work by fixed rules.