Capabilities

The topology

Everything in Sovereign AI hangs off one architecture. Your people reach it from the browser, the CLI, mobile, and Microsoft 365. Between them and the models sits the Governed Control Plane: the layer that screens every request against your content policy, routes it to the right model, meters the cost, and writes the audit record. Nothing reaches a model around it.

Read it top to bottom. Users enter through Chat, Build, Flow, and Data: every surface funnels into the same plane. The plane itself is where the twelve governance controls live: policy, approval gates, content controls, secrets, PII detection, tenant isolation, spending caps, model routing, cost optimization, audit, run receipts, and data residency. Models, local or frontier, connect only to the plane: never to users, never to your data. Below the sole communication path sits the Isolated Agent Orchestrator: your dedicated instance where agents, swarms, and flows execute, grounded by the Knowledge Base (RAG), live Data Connectors (MCP), Org SLMs trained privately on your data, and Microsoft 365 add-ins.

That is the whole trick of the platform: one control plane, one audit trail, one policy, no matter which surface asked or which model answered.

Work with it: chat, agents, swarms, flows

• →Governed chat: policy-screened conversation with the best model for the job, grounded answers with sources, charts and documents as replies, image generation in the seat.
• →Agents and swarms: describe an outcome and a composed crew plans, builds, and proves the work against your Definition of Done, with adversarial review strategies and Oversight seats.
• →Flows: durable automations with schedule, webhook, and event triggers, agents and swarms as steps, human approval gates, and a full run history.
• →Build surface: run commands, Git access under policy, a skills library, API keys and webhooks, and data queries for the technical seats.

Grounded in your knowledge

• →Sources: upload documents or sync connectors into the indexed knowledge base, with per-source status in the open.
• →MCP connections: live bridges to internal systems, queried in place rather than copied.
• →Workspace files: every artifact agents produce, versioned, revertable, and delivered where you work.

Governed by design

• →Content policy: enforced on every request, with PII handling and output redaction.
• →Plan gates and approval cards: nothing expensive happens on momentum; humans decide at the moments that matter.
• →Audit and agent logs: everything on the record, filterable, built for the regulator as much as the operator.
• →Isolation: an instance per user, region pinned at setup, enforcement at the database, transport, and compute layers.

Administered without surprises

• →Users, SSO, and access groups: your identity, your roles, seat management in bulk.
• →Connected apps and provider secrets: OAuth apps with per-tool permission modes, a write-only vault for provider and ERP credentials. Keys never touch endpoints.
• →Costs and spending limits: metered usage in the open, per-role estimates before a swarm runs, caps and budgets so invoices stay boring.
• →Model routing: the right model for every request, self-hosted models on your GPUs at zero marginal cost, BYOK ready.