A plain-English reference for the AI and AI governance terms you'll run into when evaluating a gateway, a guardrail, or a policy. We'll keep adding to this as the vocabulary around AI governance settles.
A model trained on large amounts of text to understand and generate natural language. It's the core technology behind chatbots, coding assistants, and most modern AI agents.
The input sent to a model, including instructions, questions, or context, that shapes what it responds with.
The call an application, agent, or user sends to a model provider: the prompt, plus parameters like which model to use and authentication. It's the point where a gateway can inspect and control traffic before it ever reaches the provider.
What the model provider sends back after processing a request, usually generated text, but it can also include tool calls, token usage, or an error. Guardrails can check a response just as they check the request that triggered it.
The basic unit a model reads and generates text in, roughly a word or part of a word. Usage-based AI pricing is usually billed per token.
The maximum amount of text, measured in tokens, a model can consider at once: the prompt, conversation history, and its own response, combined.
A setting that controls how random or deterministic a model's output is. Lower values produce more predictable answers; higher values produce more varied ones.
Additional training on a pretrained model using a narrower dataset, to adapt its behavior for a specific task or domain.
A numerical representation of text, or other data, that captures meaning. Embeddings let systems compare or search content by similarity instead of exact keyword matches.
A technique where a model's answer is grounded in specific documents retrieved at request time, rather than relying only on what it learned during training.
When a model produces an answer that sounds plausible but is factually wrong or unsupported by its sources.
A system that uses a model to take multi-step actions, such as calling tools, running code, or making requests, toward a goal, rather than just producing a single text response.
An open standard that lets AI models and agents call external tools and data sources in a consistent way, instead of every integration needing custom code.
A model capability where the model can request that a specific tool, API, or function be run, then use the result to continue its task.
The practices and controls that let an organization know where AI is used, apply policy to that use, and produce evidence of what happened, across every model, agent, and tool involved.
Infrastructure that sits between AI callers (apps, agents, employees) and model providers, so requests can be inspected, controlled, and logged in one place instead of scattered across every integration.
Automated checks applied to AI requests and responses, such as detecting sensitive data or blocking a request, enforced before or during a call rather than reviewed after the fact.
AI tools, agents, or API keys in use across an organization without IT or security's knowledge, often set up directly by individual employees or teams.
An attack where malicious instructions are hidden in content a model processes, such as a document, a web page, or a tool response, to make the model act against its intended instructions.
Directing an AI request to a specific model or provider based on rules like cost, availability, or task type, instead of hardcoding one model per application.
Automatically shifting traffic to a backup model or provider when the primary one is slow, unavailable, or returning errors.
A credential issued by a governance layer, instead of a raw provider API key, that lets a team or application call AI providers without ever holding the underlying key directly.
A spend limit enforced on AI usage, typically scoped to a team or provider, so a single integration or feature can't run away with cost.
A permissions model where what a user or system can do is determined by the role they're assigned, rather than being configured person by person.
The recorded detail, such as identity, policy decision, data involved, model used, and outcome, that lets a compliance or security reviewer verify what actually happened during an AI interaction.
A security approach that verifies every request based on identity and context, rather than assuming anything inside a network perimeter is automatically trusted.
Short for AI Trust, Risk, and Security Management, a category (coined by Gartner) covering the practical work of governing AI: visibility, runtime inspection, policy enforcement, data protection, access control, and audit evidence.
Tell us what's missing, or tell us what you're trying to govern and we'll help map it to the right controls.