Glossary
AI governance, in plain language
24 terms that keep coming up in AI governance conversations — defined without the jargon.
- Agentic AI
- The pattern of building applications from AI agents that plan, use tools, and act with a degree of autonomy. Agentic AI shifts the governance question from 'is the model accurate?' to 'what is this system permitted to do, and who approved each connection it holds?'
- AI agent
- An AI system that pursues goals by taking actions — calling tools, querying data sources, invoking other models or services — rather than only generating a response. Agents concentrate governance risk because their reach (what they can touch) matters as much as their model quality.
- AI governance
- The policies, processes, and controls that ensure AI systems are developed, deployed, and operated responsibly — covering accountability, risk management, compliance, transparency, and oversight across the AI lifecycle. Effective AI governance starts with knowing what AI exists in the organization and ends with evidence that controls actually operated.
- AI governance platform
- Software that operationalizes AI governance: a registry of AI systems and their dependencies, approval workflows, risk scoring, compliance assessments against regulatory frameworks, and audit trails — replacing the spreadsheets and disconnected documents most programs start with.
- AI registry (AI inventory)
- The authoritative record of an organization's AI estate: agents, models, tools, data sources, and the platforms and services they depend on, each with an owner, lifecycle stage, and governance status. Nearly every framework — EU AI Act, NIST AI RMF, ISO/IEC 42001, model risk guidance — presumes one exists.
- Architecture-as-code
- Describing system architecture in versioned, machine-readable form (for example Mermaid or C4 diagrams as text) so it can be diffed, reviewed, and checked against reality. In governance, it enables detecting drift between the architecture that was approved and what is actually running.
- Audit trail (tamper-evident)
- A chronological, append-only record of governance actions — who approved what, when, and on what evidence — protected so entries cannot be edited or silently deleted. Regulators and auditors treat the audit trail as the difference between claiming governance happened and proving it.
- CI/CD governance gate
- An automated check in the deployment pipeline that fails the build when an AI artifact has not cleared its required governance phase — enforcing governance where software actually ships rather than in documents nobody consults at deploy time.
- Compliance framework
- A structured set of requirements an organization assesses itself against — a law (EU AI Act, GDPR), a standard (ISO/IEC 42001), supervisory guidance (OCC model risk), or a community baseline (OWASP LLM Top 10). Because frameworks overlap heavily, mature programs map one body of evidence to many frameworks instead of running one program per framework.
- Dependency graph
- A map of what each system relies on — an agent's models, tools, data sources, and external services, and theirs in turn. In AI governance, the dependency graph is what turns isolated assessments into systemic ones: it shows what breaks downstream and where risk actually concentrates.
- Full-lifecycle AI governance
- Governance that spans an AI system's whole life — idea, registration, approval, deployment, monitoring, change, and retirement — rather than a one-time pre-launch review. The lifecycle framing matters because most AI risk emerges after approval: drift, new dependencies, changed usage.
- Governance checkgate
- A defined approval checkpoint an AI system must clear before advancing — for example security review, compliance check, or operational readiness — with named reviewers and recorded outcomes. Checkgates turn 'governance' from a policy document into an enforceable path to production.
- High-risk AI system (EU AI Act)
- An AI system in one of the EU AI Act's designated high-risk categories — such as credit scoring, employment decisions, essential services, or safety components — carrying the Act's heaviest obligations: risk management, data governance, documentation, human oversight, and post-market monitoring. High-risk obligations apply from December 2, 2027.
- Human oversight
- The requirement — explicit in the EU AI Act for high-risk systems — that humans can effectively monitor, interpret, and intervene in an AI system's operation. Meaningful oversight requires knowing what the system does and depends on, which is why it begins with inventory and documentation rather than a review meeting.
- LLM observability
- Monitoring the runtime behavior of LLM-powered systems — latency, cost, token usage, output quality, and failure modes — through tools purpose-built for AI workloads. Connected to governance, observability signals can trigger re-review when a system drifts from the behavior it was approved with.
- MCP server
- A service that exposes tools, data, or actions to AI applications over the Model Context Protocol. Every MCP server an agent can reach extends what that agent can do, so mature governance treats MCP servers as registered, approved entities with their own risk assessment — not as invisible plumbing.
- Model card
- Structured documentation of an AI model: its purpose, capabilities, limitations, training context, and appropriate use. The EU AI Act's Article 13 transparency requirements make model-card-style documentation effectively mandatory for high-risk systems — and keeping it current as models change is the hard part.
- Model Context Protocol (MCP)
- An open protocol that lets AI applications connect to external tools and data through a standard interface. MCP has become the common way agents acquire capabilities — which makes the servers providing those capabilities part of the governance surface.
- Model risk management (MRM)
- The discipline — long mandated in banking supervision — of inventorying models, validating them independently, monitoring them in production, and governing them with clear ownership. AI and machine learning models fall inside its scope, and agentic systems stretch it to cover tools and data dependencies too.
- Policy-as-code
- Expressing governance rules as executable, version-controlled code — commonly in Rego, evaluated by Open Policy Agent (OPA) — so every change is validated against policy automatically and consistently instead of through manual review. The same discipline that transformed infrastructure compliance, applied to AI governance.
- Retrieval-augmented generation (RAG)
- An architecture where an AI model's responses are grounded in documents retrieved at query time rather than only in training data. RAG systems add governance surface: the knowledge sources being retrieved from are data dependencies with their own sensitivity, access, and quality concerns.
- Risk propagation
- Scoring risk through dependency chains rather than per system in isolation: a high-risk data source or external service raises the risk of the agents and use cases built on it. Without propagation, the riskiest part of an AI system is routinely the part nobody scored.
- SCIM
- System for Cross-domain Identity Management — the open standard for automated user provisioning and deprovisioning between an identity provider and applications. In governance platforms, SCIM matters because joiner-mover-leaver discipline is itself a control auditors test.
- Shadow AI
- AI models, agents, and services adopted inside an organization without governance sign-off — the AI-era successor to shadow IT. Because cloud AI services are a credit card away, shadow AI accumulates quickly and invisibly; discovering it requires scanning cloud environments for AI usage that never reached the official inventory.
Ready to move from vocabulary to practice?
See how these concepts work as running governance — or measure where you stand with the free maturity assessment.