Governance
The policy layer for what an AI system is allowed to read, call, decide, and ship — encoded as configuration the runtime enforces, not as a document on a shared drive.
Production AI is not a prompt. It is a system of context, tools, permissions, traces, evals, and feedback loops.
Autonomy with boundaries
The right level of autonomy depends on the action's risk and reversibility. Reading internal documentation: usually autonomous. Filing an issue: typically autonomous. Sending an external email, modifying production data, completing a financial transaction: gated. Risk levels are declared per tool in the registry and per workflow in the runtime, not left to model judgment.
What is enforced
Data access (which sources an agent or principal can see), tool permissions (which actions are exposed and at what scope), approval gates (which calls require a human), audit retention (how long traces are kept and who can read them), tenant isolation, secret boundaries, and deployment promotion paths. Each rule is declarative; the runtime is what makes it real.
- Access controls inherited from the source graph
- Tool scopes pinned in the MCP registry
- Approval routing for high-risk actions
- Tenant isolation and per-tenant retention
Standards we map to
SOC 2 Common Criteria for security and confidentiality controls, the NIST AI Risk Management Framework for risk identification and mitigation, ISO/IEC 42001 for AI management systems, and sector-specific overlays (HIPAA, PCI DSS, Quebec Law 25 / Canadian PIPEDA / GDPR) when the engagement requires them.
What it works with
Enforces access boundaries declared in the Source Graph, scopes in the MCP Tool Registry, routes in Model Routing, and approval rules in Human Approval. Reads audit events from Observability. Sets retention rules on the conversation tables Conversation Intelligence writes. Without governance the other layers are unaudited; without the other layers governance is a document nobody enforces.
When you need it
Signals: an internal audit asking 'who can see what'; a regulator asking which data left your network and to which model provider; a sector overlay (healthcare, finance, public sector) that requires documented controls; a customer asking for SOC 2 evidence on AI workflows.
Related resources
Approval gates that put a human in the loop where correctness, risk, or accountability actually require human judgment — designed as part of the workflow, not as a panic button bolted on after launch.
A governed catalog of every tool an AI agent can call — your APIs, your databases, your internal systems — with typed schemas, permission scopes, audit trails, and the standard protocol (MCP) that turns 'we exposed it to the LLM' into 'we know exactly who called what when'.
How an AI system decides which model to call for each step — based on privacy, cost, latency, quality, and what happens when a provider goes down.
A capability in the Group e-media information AI stack. This resource connects the subject to data substrate, agent runtime, evals, and operations.
A capability in the Group e-media information AI stack. This resource connects the subject to data substrate, agent runtime, evals, and operations.