Problem
Organizations want AI that understands their context — people, roles, files, systems, decisions and history — and can act on it safely. Generic chat assistants answer prompts but know nothing about who may see what, or which actions need a person.
Requirements
- Understand authorized organizational context across documents, communications and systems of record.
- Act only within permissions set by the owner; escalate anything outward-facing, financial or production-changing.
- Deploy in the cloud where appropriate, privately for sensitive environments, and on dedicated hardware only where justified.
- Log and explain every action.
- Stay model-agnostic.
Architecture
| Layer | Responsibility |
|---|---|
| Connectors | Pull from documents, email, chat, CRM/ERP and databases with the organization's own credentials |
| Context graph | People, roles, projects, documents, decisions and their relations — with permissions attached to every node |
| Retrieval | Permission-aware search: a query only sees what the asking person or agent may see |
| Router | Classifies sensitivity and decides where each task may run; masks data before any permitted external call |
| Model runtime | The best available models behind one interface — swappable as models improve |
| Agents | Role agents that research, plan, propose and take actions within limits |
| Permission kernel | Checks every tool call against role, data scope and action class |
| Approval gates | Decision cards for actions that need a person |
| Activity log | An immutable record of inputs, actions and reasons |
Key decisions
- Permissions are data, not prompts. The kernel enforces scope in code; the model never decides what it may access.
- Action classes, not trust levels. Read and analyse are allowed; internal writes are logged; outward-facing, financial or production changes need approval; legal is never delegated.
- The model is replaceable. Differentiation sits in context, orchestration and governance — not in owning a frontier model.
Status
Blu Box is in development. The current focus is one real end-to-end workflow with permissions, logging and human review. No customer deployment is claimed.