AI Glossary

An AI system that can pursue a goal by reading instructions, using tools, checking results, and deciding what to do next.

In practice: A support agent that reads a ticket, searches the help center, drafts a reply, and asks for approval before sending.

The underlying AI system that generates text, code, images, audio, or structured outputs from an input.

In practice: When a product says it is powered by AI, the model is the engine underneath the interface.

A large language model is a model trained to understand and generate language. It predicts useful next words based on patterns learned from large amounts of text.

In practice: LLMs power chatbots, writing assistants, coding tools, document search, and many internal automation tools.

The instruction or context you give an AI system so it knows what you want.

In practice: "Summarize this meeting transcript for the sales team" is a prompt.

Background instructions that shape how an AI assistant should behave before a user asks anything.

In practice: A system prompt might tell an assistant to be concise, follow company policy, and avoid giving legal advice.

One unit of conversation sent to or from an AI system.

In practice: A user request, an assistant response, and a tool result can all be messages inside the same workflow.

A small unit of text that models process. Tokens are often shorter than words and are used for pricing, limits, and performance.

In practice: Long documents use more tokens, which usually means more cost and slower responses.

The amount of information a model can consider at one time.

In practice: A larger context window lets a tool read more files, transcripts, or documents before answering.

A model that can work with more than one type of input or output, such as text, images, audio, or video.

In practice: A multimodal model can review a screenshot and explain what a user is seeing.

A model optimized for solving problems that require multiple steps, tradeoffs, or careful analysis.

In practice: Reasoning models are useful for planning projects, debugging tricky errors, or comparing strategic options.

Retrieval-augmented generation is a pattern where an AI system searches trusted content before answering.

In practice: Instead of guessing from memory, an internal assistant can search your policies and respond based on the matching sections.

The step where an AI system finds relevant information from documents, databases, or tools.

In practice: Retrieval is what lets an assistant pull the right contract clause before drafting an answer.

A numerical representation of text, images, or other data that captures meaning well enough for similarity search.

In practice: Embeddings help a search system understand that "refund policy" and "return rules" may be related.

A database designed to store embeddings and find similar items quickly.

In practice: A vector database can help a chatbot find the most relevant pages in a large knowledge base.

A curated collection of information that an AI system can use to answer questions or complete work.

In practice: Company policies, product documentation, sales notes, and support articles can form a knowledge base.

The process of splitting large documents into smaller sections before indexing or retrieval.

In practice: A 50-page manual might be split by section so an assistant can retrieve only the parts that matter.

Connecting an AI answer to reliable information, tool results, or documents.

In practice: A grounded answer explains what it found in the CRM instead of inventing a customer status.

A reference that shows where an AI answer came from.

In practice: Internal search tools often cite the source document so employees can verify the answer.

A capability an AI system can call to do something beyond generating text.

In practice: Tools can search files, run calculations, send requests to software, or update a record after approval.

A way for a model to request a specific action using structured data instead of free-form text.

In practice: If a user asks for account status, the model can call a customer lookup function with the account ID.

A defined way for one software system to communicate with another.

In practice: An AI workflow might use an API to create a ticket, fetch a customer record, or send a Slack message.

An integration that lets an AI system read from or act inside another tool.

In practice: A connector can let an assistant search Google Drive, read Salesforce records, or draft Jira issues.

Model Context Protocol is a standard approach for connecting AI assistants to external tools and data sources.

In practice: A team can expose internal systems through an MCP server so compatible AI tools can use them consistently.

A command line interface lets software be controlled through terminal commands.

In practice: Coding agents often prefer CLIs because commands are faster and more reliable than clicking through web pages.

An isolated environment where an AI system can safely run code or test changes.

In practice: A coding agent can install dependencies and run tests in a sandbox without touching production systems.

Software that moves a repeatable business process from one step to the next with limited manual effort.

In practice: A workflow can watch for new invoices, extract key fields, route exceptions, and notify finance.

The cycle where an agent plans an action, takes it, observes the result, and decides the next step.

In practice: A coding agent edits a file, runs tests, reads the failure, and adjusts the fix.

Coordinating multiple tools, steps, or agents so they work together toward a larger outcome.

In practice: A research workflow might gather data, summarize findings, draft a memo, and route it for review.

A specialized agent used by another agent or workflow to handle part of a task.

In practice: One subagent may research competitors while another checks pricing pages.

A design pattern where a person reviews, approves, or corrects an AI action before it becomes final.

In practice: An assistant can draft a client email but wait for a manager to approve it before sending.

How much an AI system can do without human approval.

In practice: A low-autonomy assistant drafts recommendations. A higher-autonomy assistant can update systems within approved limits.

A scheduled AI workflow that runs at a regular time or after a recurring trigger.

In practice: A weekly routine can summarize open support themes every Monday morning.

A reusable set of steps for handling a common AI-assisted task.

In practice: A sales research playbook might gather account news, review CRM notes, and create a briefing.

The transfer of work from an AI system to a person, another system, or another agent.

In practice: A triage assistant can hand complex legal questions to a human reviewer.

Information saved outside the current conversation so it can be reused later.

In practice: A company assistant might remember preferred report formats or team-specific terminology.

A provider-side optimization that can reuse repeated prompt content to reduce cost or latency.

In practice: Long standing instructions may be cheaper to reuse across repeated calls when caching is available.

Additional model training on task-specific examples to influence style, format, or behavior.

In practice: Fine tuning can help a model produce outputs that match a company's support tone or tagging taxonomy.

Model output constrained to a predictable format such as JSON.

In practice: Structured output makes it easier to send an AI result into a database, form, or downstream workflow.

An AI response that sounds confident but is not supported by facts.

In practice: A model may invent a policy if it is asked a question without access to the real policy document.

A setting that influences how varied or predictable a model's responses are.

In practice: Lower temperature is useful for extraction and classification. Higher temperature can be useful for brainstorming.

The time it takes for an AI system to respond.

In practice: A customer-facing chat experience needs lower latency than an overnight research workflow.

The total cost to complete one AI-assisted job, including model usage, tools, retries, and human review.

In practice: Cost per task matters more than token price when deciding whether an automation is worth shipping.

A repeatable test that measures whether an AI system performs a task well enough.

In practice: An eval can check whether a support assistant gives correct answers across 100 real customer questions.

A standardized comparison used to measure models or systems on the same set of tasks.

In practice: Benchmarks can help narrow options, but internal evals matter more for company-specific work.

A trusted set of examples with expected answers or outcomes.

In practice: A team can use a golden dataset of past tickets to test whether a new assistant is improving.

A check that makes sure a change did not break behavior that previously worked.

In practice: After updating a prompt, regression tests can confirm the assistant still handles common requests correctly.

A record of the steps an AI system took during a task.

In practice: A trace can show the prompt, retrieved documents, tool calls, and final output for debugging.

The ability to monitor and inspect how an AI system behaves in production.

In practice: Observability helps teams see failure rates, tool errors, latency, cost, and user feedback.

A process for turning user corrections and outcomes into improvements.

In practice: If reviewers keep fixing the same type of draft, that feedback should update prompts, data, or workflow rules.

Rules or checks that limit what an AI system can say or do.

In practice: Guardrails can block unsafe actions, require approval, or prevent the assistant from discussing restricted topics.

Controls that determine what data and actions an AI system can access.

In practice: A finance assistant should not read HR files unless the user and workflow are authorized.

Sensitive values such as API keys, passwords, and tokens that should not be exposed to AI outputs or logs.

In practice: A coding agent should use secure environment variables instead of printing credentials.

Personally identifiable information is data that can identify a person.

In practice: Names, emails, phone numbers, addresses, and account IDs may require special handling.

How long inputs, outputs, logs, and tool results are stored.

In practice: Teams should know whether vendor logs are retained and whether sensitive prompts can be deleted.

Testing an AI system by intentionally trying to make it fail, leak data, or break rules.

In practice: Red teaming can reveal unsafe prompts, weak permissions, or confusing approval flows before launch.

A specific business problem where AI can create measurable value.

In practice: Reducing time spent on proposal drafts is a stronger use case than "use AI in sales."

A limited launch designed to test value, risk, and adoption before scaling.

In practice: A team might pilot an internal research assistant with 20 users before rolling it out company-wide.

Return on investment compares the value created by an AI system with the cost to build and operate it.

In practice: ROI can include saved hours, faster cycle times, higher conversion, fewer errors, or better customer experience.

The work required to help people adopt a new process or tool.

In practice: Training, clear ownership, workflow redesign, and manager support often matter as much as the model choice.

How prepared a team is to use AI effectively based on data quality, process clarity, tooling, permissions, and culture.

In practice: A team with clean documentation and clear workflows can usually move faster than one with scattered knowledge.

The structure for owning, improving, and governing AI work across a company.

In practice: An operating model defines who approves use cases, who maintains systems, and how teams report outcomes.