AI platform
Purpose
The AI works on this system's data, not on general world knowledge — and it sees, without exception, only what the person asking would be allowed to see anyway.
Three decisions shape the outcome more than the choice of language model:
- A single context path. What the AI gets to see is assembled in exactly one place. There is no
hidden instruction somewhere in the program.
- Permissions apply unchanged. The AI is given no separate, wider view.
- Answers carry their sources. An answer without a traceable source is an assertion, and
recognisable as one.
Scope
- Several providers in parallel, plus your own and locally hosted models.
- Models are not hard-coded. They come from a catalogue queried at the provider and refreshed
regularly — a new model appears without a new program version.
- Routing per task: which model handles which task is configurable. That is the lever for cost
and quality.
- A curated knowledge base of company facts, glossary, tone of voice and guidelines that flows
reliably into every context rather than being found by chance.
- Rules per use case and brand governing what a use case may see at all.
- A test console that shows the actual context — not the expected one.
- Access for AI tools over MCP with a distinct identity, so that actions appear in the log as
AI.
Interplay
Provider access lives in the AI settings, the knowledge base in the CompanyBrain, and the working surface is the AI assistant. Actions run over the API and appear in traceability.
Limits
- No data access of its own. What a person may not see, the AI does not see.
- The portal assistant sees considerably less — for customers, a narrow, fail-closed slice
applies.
- Without a configured provider and the flags set, the AI is unavailable. Switched off means
switched off, not "quietly in the background".
- A language model can be wrong. Hence the citations: they make an answer checkable rather than
making it credible.