{
  "name": "AI Governance Glossary",
  "description": "A reference glossary of AI governance terms — from policy enforcement to deterministic governance and audit trails.",
  "url": "https://gianlucabusato.com/datasets/ai-governance-glossary/",
  "creator": {
    "name": "Gianluca Busato",
    "id": "https://gianlucabusato.com/#person"
  },
  "dateModified": "2026-07-15",
  "inLanguage": "en",
  "columns": [
    "term",
    "definition"
  ],
  "records": [
    {
      "term": "AI Agent",
      "definition": "A software entity that perceives context, plans, and takes actions toward a goal with limited or no human intervention."
    },
    {
      "term": "Agentic AI",
      "definition": "AI systems composed of one or more autonomous agents that can chain reasoning, call tools and act on the world."
    },
    {
      "term": "AI Governance",
      "definition": "The set of policies, controls and processes governing how AI systems are built, deployed, monitored and held accountable."
    },
    {
      "term": "Policy Enforcement",
      "definition": "The runtime mechanism that ensures an AI agent can only perform actions explicitly permitted by governance rules."
    },
    {
      "term": "Deterministic Governance",
      "definition": "Governance whose rules produce the same, repeatable outcome for the same inputs, enabling audit and reproducibility."
    },
    {
      "term": "Identity Control",
      "definition": "Assigning a verifiable identity and roles to each agent so its actions can be attributed and scoped."
    },
    {
      "term": "Operational Transparency",
      "definition": "The property of a system that produces a complete, reconstructable record of decisions and actions."
    },
    {
      "term": "Audit Trail",
      "definition": "An immutable, time-ordered log of every decision and action taken by an AI system."
    },
    {
      "term": "Human-in-the-Loop (HITL)",
      "definition": "A control pattern requiring human approval at defined checkpoints before high-impact actions execute."
    },
    {
      "term": "Guardrail",
      "definition": "A constraint that bounds an AI system's behavior to keep it within safe and permitted limits."
    },
    {
      "term": "Multi-Layer Governance",
      "definition": "An architecture enforcing controls at independent layers (identity, policy, execution, audit) for defense in depth."
    },
    {
      "term": "Agent Orchestration",
      "definition": "Coordinating multiple agents with defined roles and permissions to collaborate on complex workflows."
    },
    {
      "term": "Scoped Permissions",
      "definition": "Narrowly defined privileges granted to an agent for a specific task, following least-privilege principles."
    },
    {
      "term": "Model Risk",
      "definition": "The risk that an AI model produces harmful, biased or non-compliant outputs or actions."
    },
    {
      "term": "Prompt Injection",
      "definition": "An attack where crafted input manipulates an AI agent into ignoring its instructions or policies."
    },
    {
      "term": "Tool Use",
      "definition": "An agent's ability to invoke external functions, APIs or services to accomplish a task."
    },
    {
      "term": "Accountability",
      "definition": "The ability to attribute an AI action to a responsible identity and explain why it occurred."
    },
    {
      "term": "Compliance-by-Design",
      "definition": "Embedding regulatory and policy requirements directly into a system's architecture from the outset."
    },
    {
      "term": "Observability",
      "definition": "Instrumentation that makes an AI system's internal state and behavior measurable in real time."
    },
    {
      "term": "Least Privilege",
      "definition": "A security principle granting the minimum access necessary to perform a function."
    },
    {
      "term": "Reversibility",
      "definition": "The capacity to undo or roll back an action taken by an autonomous system."
    },
    {
      "term": "Sandboxing",
      "definition": "Running an agent in an isolated environment to limit the blast radius of its actions."
    },
    {
      "term": "Drift",
      "definition": "Gradual change in model behavior or data distribution that can degrade governance assumptions."
    },
    {
      "term": "Attestation",
      "definition": "A verifiable claim about the state or identity of a system or agent at a point in time."
    },
    {
      "term": "Policy-as-Code",
      "definition": "Expressing governance rules in machine-executable form so they are enforced automatically."
    },
    {
      "term": "LLM (Large Language Model)",
      "definition": "A neural network trained on large text corpora to generate and reason over natural language."
    },
    {
      "term": "AIO (AI Optimization)",
      "definition": "Optimizing content to be surfaced and cited by AI answer engines and search overviews."
    },
    {
      "term": "LLM_O (LLM Optimization)",
      "definition": "Optimizing an entity's presence so large language models describe and cite it accurately."
    },
    {
      "term": "Entity",
      "definition": "A distinct person, organization or concept that search engines and LLMs model as a node of knowledge."
    },
    {
      "term": "Knowledge Graph",
      "definition": "A structured network of entities and relationships used by search engines to understand the world."
    }
  ]
}