AI Glossary
Plain-language definitions of AI and automation terms — no jargon left unexplained.
Agentic Workflow— A multi-step automated process where an AI agent makes decisions along the way.AI Agent— An AI system that can take multiple steps and actions toward a goal, not just answer a single question.API (Application Programming Interface)— A defined way for two pieces of software to communicate with each other.Blue-Green Deployment— Running two identical production environments and switching traffic between them, so a bad release can be reversed instantly.Chain-of-Thought Prompting— Prompting a model to reason through intermediate steps before giving a final answer.Chunking— Splitting a long document into smaller pieces before storing it for retrieval.CI/CD (Continuous Integration / Continuous Deployment)— Automatically testing and deploying code changes as they're made, instead of manually bundling and releasing them on an irregular schedule.Cold Start— The delay before a serverless function or newly-scaled system is actually ready to handle a request.Confidence Threshold— The minimum certainty level required before an AI system acts on its own output automatically.Context Engineering— Deliberately deciding what information a model sees for a given task — not just how the prompt is worded, but what's included at all.Context Window— The maximum amount of text an LLM can consider at once in a single request.Data Residency— The requirement that data be stored (and sometimes processed) within a specific geographic or legal jurisdiction.Data Silo— Data trapped in one system that other systems and teams can't easily access, forcing manual re-entry or exports to bridge the gap.Embedding— A numerical representation of text (or other data) that captures its meaning.Eval (Evaluation Harness)— A structured test set used to measure how well an AI system performs on real, representative examples.Feature Flag— A toggle that turns a feature on or off in production without deploying new code, letting you ship code dark and release the feature separately.Fine-Tuning— Further training an existing model on a specific dataset to change its behavior or style.Function Calling / Tool Use— A model's ability to call external functions or APIs as part of generating a response.Grounding— Anchoring a model's output in specific, verifiable source material instead of letting it answer purely from what it memorized during training.Guardrails— Explicit checks that constrain what an AI system is allowed to say or do.Hallucination— When an AI model generates confident-sounding but factually incorrect or invented information.Human-in-the-Loop— A system design where a person reviews or approves an AI's output before it takes effect.Idempotency— A property where running the same operation multiple times produces the same result as running it once.Inference— Running a trained model to produce an output, as opposed to training it.Infrastructure as Code (IaC)— Defining servers, networks, and cloud resources in version-controlled configuration files instead of clicking through a console by hand.Knowledge Graph— A network of real, distinct facts connected by explicit relationships, rather than a flat list of unconnected pages.Legacy System— An existing system a business depends on that's become hard to change, poorly documented, or built on an outdated stack.LLM (Large Language Model)— A machine learning model trained on large amounts of text to predict and generate language.MCP (Model Context Protocol)— An open standard for connecting AI models to external tools and data sources.Model Drift— A model's real-world accuracy quietly degrading over time as the data it sees diverges from what it was trained or tuned on.Multi-Agent System— Multiple AI agents, each handling a distinct sub-task, coordinating toward one overall goal instead of one agent doing everything.Multimodal AI— AI models that can process more than one type of input — text, images, audio, or video together.MVP (Minimum Viable Product)— The smallest real version of a product that lets you validate whether it actually solves the problem, before investing in the full build.No-Code Automation— Building automated workflows via a visual interface instead of writing code.Open-Source / Open-Weight Model— An AI model whose underlying parameters (weights) are published and can be run on your own infrastructure.Orchestration Platform— Software that coordinates multi-step workflows across several tools or services.Ownership Transfer— An engagement model where the client ends up owning and able to independently run/extend what was built.PII (Personally Identifiable Information)— Any data that can identify a specific individual — names, emails, ID numbers, and similar.Prompt Engineering— The practice of designing inputs to an AI model to reliably get the output you want.Prompt Injection— An attack where malicious instructions hidden in input data trick an AI system into ignoring its actual instructions.RAG (Retrieval-Augmented Generation)— A technique where an AI model retrieves relevant information from a knowledge source before generating an answer.Rate Limiting— A cap on how many requests a system will accept or send in a given time window.Reranking— A second pass that re-scores retrieved results before they're used, to improve relevance.Scope Creep— A project's requirements quietly expanding beyond what was originally agreed, without a corresponding change in budget or timeline.Semantic Search— Search based on meaning rather than exact keyword matching.Shadow AI— AI tools employees use for work without IT or leadership's knowledge or approval.Single Source of Truth— One system designated as the authoritative record for a given piece of data, so other systems reference it instead of holding their own conflicting copy.SLA (Service-Level Agreement)— A defined commitment on response time, uptime, or support quality between a vendor and client.SOP (Standard Operating Procedure)— A documented, repeatable process for how a specific task is done — the prerequisite for automating it reliably.Structured Output— Forcing a model's response into a fixed, machine-readable format like JSON.Synthetic Data— Artificially generated data used to train or test a system when real data is scarce, sensitive, or doesn't cover enough edge cases.System Prompt— Instructions given to an AI model that set its behavior for the whole conversation, before any user input.Technical Debt— The accumulated cost of shortcuts taken during a build that make future changes slower or riskier.Temperature (LLM parameter)— A setting that controls how random or deterministic a model's output is.Token— The basic unit of text an LLM processes — roughly a word or word-fragment.Vector Database— A database optimized for storing and searching numerical representations of meaning (embeddings).Vendor Lock-In— The real cost and difficulty of switching away from a vendor once your business depends on their specific platform.Webhook— A way for one system to automatically notify another system the moment something happens.