AI terms, in plain language.

100 of 100 terms

Access controlpermissions

Rules that determine who or what can view information, change records, or run an action in a system.

Accuracy

How often an output is correct for a defined task. Accuracy should be tested with relevant examples, not assumed.

Action plan

A short, ordered list of next steps for moving a goal forward. It gives you a practical place to start.

AgentAI agent · agentic AI

An AI system that can use instructions and approved tools to complete parts of a task, often in several steps. It still needs clear boundaries and review.

AIartificial intelligence

Artificial intelligence: software that recognizes patterns and helps with tasks such as drafting, summarizing, organizing, or exploring ideas. It can be wrong or incomplete.

Alignment

The effort to make an AI system follow intended goals, rules, and human values instead of producing harmful or unwanted behavior.

Annotationlabeling

Adding useful labels or notes to data, such as marking whether a support message is urgent or which object appears in an image.

APIapplication programming interface

A set of rules that lets one piece of software request data or actions from another. APIs let tools connect without people copying information by hand.

Audit trailactivity log

A dated record of actions, changes, approvals, and events. It helps teams investigate errors and demonstrate accountability.

Authenticationlogin verification

Confirming that a person or system is who it claims to be, often with a password, passkey, code, or security key.

Authorizationaccess permission

Deciding what an authenticated person or system is allowed to see or do.

Automation

A repeatable process that lets software carry out a routine step after you set the rules. Review the setup and results, especially when people, money, or important decisions are involved.

Batch processingbatch job

Handling many items together on a schedule or in one run instead of processing each item immediately.

Benchmark

A standard test used to compare systems on a defined task. A high benchmark score does not guarantee useful results in every real situation.

Bias

A systematic pattern that can make an AI output unfair, incomplete, or less accurate for certain people or situations. Bias can come from data, design choices, or use.

Brief

A clear summary of what you need, who it is for, and what a useful outcome looks like. A good brief gives AI or a collaborator helpful context.

Classification

Assigning an item to a category, such as labeling support messages by topic or marking a review as positive or negative.

Confidence score

A number that estimates how certain a model is about an output. It is not proof that the output is correct.

Consent

A person’s informed, voluntary agreement to a specific use of their information or content. Store and honor consent choices.

Content moderation

Reviewing or filtering content against rules for safety, legality, quality, or community standards. It can be automated, human-led, or both.

Context window

The amount of text, data, or conversation an AI model can consider at one time. Information outside that limit may not affect its response.

Data minimization

Collecting and sharing only the information needed for a task. It reduces privacy and security risk.

Data retention

The policy for how long information is kept and when it is securely deleted. Keeping data longer than needed increases risk.

Data source

The place information comes from, such as a document, database, form, website, or approved application.

Database

An organized collection of information that can be searched, updated, and reused. It might hold contacts, tasks, notes, inventory, or project details.

Dataset

A structured collection of examples or records used to train, test, or operate a system.

Deployment

Making a software system available for real use after it has been built and tested.

Deterministic

Describes a process that gives the same output for the same input and settings. Many generative AI tasks are not fully deterministic.

Embedding

A numeric representation of text, images, or other content that helps software find items with similar meaning.

Engagement

The agreed way people work together on a project or service. It usually covers the goal, communication, timing, responsibilities, and boundaries.

Error handling

The planned way a system responds when something goes wrong, such as showing a clear message, retrying safely, or asking for human review.

Evaluationeval

A planned way to check whether an AI system produces useful, accurate, safe, and consistent results for the task it is meant to do.

Examplefew-shot example

A sample input and desired output included in a prompt to show an AI the pattern or format you want.

Feedback loop

A process where results and user feedback are reviewed to improve a system, prompt, or workflow over time.

Few-shot promptingfew shot

Giving a model a small number of examples in the prompt so it can follow the desired pattern more reliably.

Foundations

The basic concepts and habits needed before adding advanced tools or workflows. In AI, this includes clear goals, useful context, checking outputs, and protecting private information.

Function callingtool calling

A feature that lets an AI request a defined software action, such as looking up an order or creating a calendar event. The application controls what is actually allowed.

Generative AIGenAI

AI that creates new content, such as text, images, audio, code, or summaries, based on patterns learned from data.

GovernanceAI governance

The policies, roles, records, and review processes used to guide responsible decisions about AI systems.

Grounding

Anchoring an AI response in reliable, relevant source material instead of asking it to answer from general patterns alone.

Guardrail

A rule, check, or limit designed to prevent unsafe, irrelevant, or unauthorized AI behavior.

Hallucination

An AI response that sounds confident but includes made-up, incorrect, or unsupported information. Verify facts, quotes, sources, and calculations.

Human in the loopHITL · human review

A workflow where a person reviews, approves, corrects, or can stop an AI-supported action before it has an important effect.

Human oversighthuman supervision

Meaningful human ability to understand, monitor, intervene in, and take responsibility for an AI-supported process.

Identity verificationID verification

Checking that a person is connected to the identity they claim, often before granting access or completing a high-risk action.

In-context learning

A model’s ability to adapt to instructions and examples included in the current prompt without changing its underlying training.

Inference

Using a trained model to produce an output, such as an answer, summary, classification, or prediction from new input.

Input

The information you give a system to work with, such as a prompt, document, image, form response, or API request.

Instruction hierarchy

The order of authority among instructions given to an AI system. Higher-priority rules should override conflicting lower-priority requests.

Integration

A connection that lets separate tools share data or trigger actions as part of one workflow.

Intent

The underlying goal behind a person’s request. Identifying intent helps a system choose an appropriate response or route a task.

Iteration

A repeat cycle of trying, reviewing, and improving. With AI, you may refine the prompt, context, or output format after seeing a result.

JSONJavaScript Object Notation

A common text format for structured data. It uses named fields and values so software can reliably exchange information.

Knowledge base

A maintained collection of approved information, such as policies, guides, and answers, that people or AI tools can search.

Large language modelLLM

An AI model trained on large amounts of text to recognize language patterns and generate or transform text. It does not automatically know your private or current information.

Machine learningML

A branch of AI in which software learns patterns from examples to make predictions, classifications, or generated outputs.

Metadata

Information that describes other information, such as a file’s author, date, topic, source, or permissions.

ModelAI model

The trained software system that processes input and produces an output. Different models have different strengths, limits, costs, and safety controls.

Monitoring

Ongoing observation of a system’s performance, errors, costs, safety signals, and outcomes after it is released.

Multimodal

Able to work with more than one kind of content, such as text, images, audio, video, or files.

Named entity recognitionNER

Finding and labeling names in text, such as people, companies, places, dates, or product names.

Natural language processingNLP

Technology that helps computers work with human language, including reading, writing, translating, classifying, and extracting information.

Opt-in

A clear choice to receive a service, message, or data use after being told what it involves. It should not be assumed from silence.

Orchestration

Coordinating several tools, data sources, or steps so they work together in a larger automated process.

Output

The result a system returns, such as text, an image, a classification, a file, or a completed action.

Personally identifiable informationPII · personal data

Information that can identify or reasonably be linked to a person, such as a full name, email address, government ID, or account number.

Privacy by design

Building privacy protections into a product or workflow from the start instead of adding them only after a problem appears.

Prompt

The instruction or question you give an AI tool. A useful prompt explains the task, relevant context, and the kind of response you need; it does not guarantee a correct answer.

Prompt injection

A malicious or unwanted instruction hidden in content that tries to make an AI ignore its rules or reveal information. Treat untrusted content as data, not instructions.

Prompt template

A reusable prompt with placeholders for details such as audience, goal, source material, or desired format.

Quality assuranceQA

A structured review process that checks whether work meets agreed requirements before it is shared or used.

Regression test

A repeatable test used after a change to make sure a feature that previously worked still works as expected.

Responsible AIAI ethics

The practice of designing and using AI with attention to safety, fairness, privacy, transparency, accountability, and real-world impact.

Retrieval-augmented generationRAG · retrieval augmented generation

A method that finds relevant documents first, then gives them to a language model to help it answer with current, organization-specific context.

Retry

Trying a failed action again, usually with limits and delays so a temporary problem does not create duplicate work or overload a service.

Role-based access controlRBAC

A permissions method that grants access based on a person’s job role, such as administrator, editor, or viewer.

Sandbox

An isolated environment for safely testing software, automations, or data connections without affecting live customers or records.

Scope

The work included in a project, along with what is not included. Defining scope helps set a realistic shared expectation before work begins.

Semantic search

A search method that looks for similar meaning, not just exact matching words. It often uses embeddings.

Sensitive data

Information that could cause harm if exposed or misused, such as health details, financial records, passwords, or personal identifiers.

Single sign-onSSO

A login method that lets a person use one trusted account to access several connected services.

Source material

The notes, documents, examples, or data used as input for a task. Use material you are allowed to share, and check that it is current and accurate.

Structured output

An AI response returned in a predictable format, such as a table, JSON object, or fixed set of fields, so people or software can use it reliably.

Summarization

Reducing a longer piece of content to its most important points. Check a summary against the original when details matter.

Synthetic data

Artificially created data that resembles real data without directly copying individual records. It still needs privacy and quality review.

System prompt

High-priority instructions set by an application that define an AI assistant’s role, limits, tone, and allowed behavior.

Taxonomy

A shared system for naming and organizing categories. A useful taxonomy makes information easier to classify, search, and report on.

Temperature

A setting that affects how varied or predictable a language model’s responses are. Lower settings are usually more consistent; higher settings may be more varied.

Template

A reusable starting structure for a common task, such as a plan, message, or workflow. It is guidance to adapt to your situation, not a promise of a particular result.

Toolkit

A small set of related resources, such as templates, prompts, checklists, or instructions, designed to support one type of work. It helps you get started; your judgment shapes the outcome.

Traceability

The ability to follow a result back through its inputs, rules, data sources, and changes so it can be checked or explained.

Transparency

Giving people understandable information about how an AI system is used, what it can and cannot do, and when its output may affect them.

Trigger

An event that starts an automated workflow, such as a submitted form, a new row in a table, or a scheduled time.

User acceptance testingUAT

Testing by the people who will use a system to confirm it supports their real tasks before a wider release.

Validation

Checking that data, a prompt, or an output meets required rules before the next step. For example, a form may validate that an email address is present.

Vector databasevector store

A database designed to store embeddings and quickly find content with similar meaning. It is often used in RAG systems.

Version control

A system for tracking changes to files over time so teams can review history, collaborate, and restore an earlier version if needed.

Webhook

An automatic message one service sends to another when a specific event happens. It is commonly used to start workflows in real time.

Workflow

A repeatable sequence of tasks, decisions, and handoffs used to achieve a goal. A workflow can be manual, automated, or both.

Zero-shot promptingzero shot

Asking a model to complete a task using instructions alone, without giving it examples of the desired output.