The AI Glossary for Business Leaders (No PhD Required)

Every AI vendor pitch comes with the same problem. Halfway through the deck, the slides start using words that sound technical, look authoritative, and quietly hide what's actually being sold. "Our orchestration layer leverages a vector store to enable retrieval-augmented inference across heterogeneous data sources." That sentence is not impressive. It's expensive.

This glossary is for the meeting after the meeting — when you're trying to figure out what was actually proposed and whether you should pay for it. Every entry has the plain-English definition, then a one-line "what this means for you" that translates the term into a decision you might actually make. Salesforce and Workday and Columbia Business School each publish solid glossaries; this one adds the buying-decision layer that the others leave out.

Listed alphabetically. Skim, search, or read end to end.

Agent (AI Agent). A piece of software that uses an AI model to pick from a set of possible actions and run them. The choices are the work. Without choices, it's a script. What this means for you: Most things called "agents" don't actually need to be agents. If your task is "read this, write that, save it" with no decisions, you're paying for an agent label. Push back.

Agentic AI. AI systems that operate autonomously toward a goal, often through multi-step planning and tool use. What this means for you: "Agentic" is the 2026 marketing term for "we built an agent." Productive translation: ignore the adjective and ask which workflow specifically the agent owns end-to-end.

Algorithm. A set of instructions or rules a computer follows to solve a problem. In AI, often refers to the underlying mathematical method (e.g., gradient boosting, transformer attention). What this means for you: The algorithm matters less than the data and the workflow. Vendors who lead with algorithms are usually selling sophistication, not outcomes.

Benchmark. A standardized test that compares AI models on the same task. Useful for ranking models in research papers. Almost never matches how the model performs on your specific data. What this means for you: Don't pick a model based on benchmarks. Pick based on a 50-example test set built from your real workflow.

Bias. Systematic errors in AI results, often caused by prejudiced or skewed training data. What this means for you: Bias isn't fixed by "better models." It's mitigated by careful data selection, diverse evaluation sets, and ongoing audits. Vendors who claim their model is bias-free are misleading you.

Big Data. Datasets so large or complex that traditional data processing tools struggle to handle them. What this means for you: Most business problems don't actually have big data. They have moderate data with messy structure. Vendors selling big-data solutions to small-data problems are a common overspend pattern.

Computer Vision. AI that interprets visual data — images, videos — to recognize objects, faces, defects, etc. What this means for you: Mature, well-understood technology with API access from every major cloud. Don't train your own unless you have a unique image type that off-the-shelf models can't handle.

Context window. How much text a model can read in one go. Measured in tokens. A 200,000-token window holds about 600 pages. What this means for you: Modern context windows hold a lot. Many "you need RAG" pitches are obsolete because the context is now big enough to skip retrieval entirely on small to mid-sized document sets.

Deep Learning. A subset of ML using neural networks with many layers, especially effective on unstructured data. What this means for you: The right tool for images, audio, video, and language. The wrong tool for tabular prediction. Don't pay for deep learning on a problem XGBoost would solve.

Embedding. A way of representing text or images as a list of numbers so a computer can compare them mathematically. What this means for you: The math behind semantic search. You don't need to understand it. You just need to know it's how a system finds "similar" things even when the words don't match.

Evaluation set (eval set). A small collection of real input-output examples used to grade an AI system. What this means for you: If your AI vendor cannot show you the evals their system passed before deploying, walk away. Evals are how grown-ups ship AI.

Fine-tuning. Continuing to train a pretrained model on your own data to specialize it for a task. What this means for you: You probably don't need this. RAG handles 90% of what people think they need fine-tuning for, at a fraction of the cost.

Foundation Model. A large model trained on huge amounts of data that can be adapted to many tasks. Claude, GPT-5, Gemini, Llama are foundation models. What this means for you: These are the off-the-shelf AI products you call through an API. Almost every AI project sits on top of three or four of these.

Generative AI (GenAI). AI that creates new content — text, images, audio, code — instead of just predicting or classifying. What this means for you: Powerful for content tasks. The wrong tool for most decision tasks. Don't put generative AI on problems that need a number, a class, or a probability.

Hallucination. When an AI model produces output that sounds confident and turns out to be made up. What this means for you: Hallucinations don't get fixed by "better models." They get managed by retrieval, validation, and human review on the cases that matter. Plan for them.

Inference. Running a trained AI model on new inputs to get an output. What this means for you: When you're paying per API call, you're paying for inference. Your cost lever is inference volume.

Latency. Time between sending an input and getting an output. What this means for you: Generative AI is slower than traditional AI by orders of magnitude. If your application needs sub-second responses, design accordingly.

Large Language Model (LLM). A foundation model specialized in text. Reads, writes, reasons over text. What this means for you: This is what most people mean when they say "AI" in 2026. Vendor differences (Anthropic, OpenAI, Google) are real but smaller than the marketing implies.

Machine Learning (ML). Software that learns patterns from data instead of being explicitly programmed with rules. What this means for you: "AI" and "machine learning" are not interchangeable. ML is one approach to AI. Most non-generative AI in your business is classical ML.

Multimodal AI. A model that processes more than one kind of content — text, images, audio, video — in the same call. What this means for you: Buying separate vendors for vision and language is increasingly unnecessary. Most frontier models in 2026 handle both.

Natural Language Processing (NLP). AI that understands and processes human language. What this means for you: In 2026, NLP is largely subsumed by LLMs. Don't buy a separate "NLP solution" if a foundation model does the job.

Neural Network. A computing system loosely inspired by brain neurons, made of layers of mathematical units that transform data. What this means for you: The architecture under deep learning. Mostly the vendor's problem, not yours.

Orchestration. The layer that decides which AI model gets called when and stitches results together. What this means for you: Don't pay for an orchestration platform when one model is doing all the work.

Predictive Analytics. Using historical data and AI to forecast future outcomes — churn, demand, default risk. What this means for you: This is classical ML territory. Foundation models are the wrong tool here.

Prompt. The instructions you give an AI model. What this means for you: Prompt quality has more impact on output quality than model choice for most tasks.

Prompt Engineering. The discipline of designing prompts to get reliable, high-quality outputs. What this means for you: If your team has nobody whose job it is to write and improve prompts, you're flying blind. This is a real role.

Retrieval-Augmented Generation (RAG). Connecting a language model to your own documents so it can look things up before answering. What this means for you: How most enterprise AI knowledge applications work. Cheaper, faster to build, more controllable than fine-tuning. Start here.

Recommendation Engine. AI that suggests products, services, or content based on user behavior. What this means for you: Mature predictive AI category. Don't reinvent — most ecommerce platforms have built-in or readily-integrated solutions.

Robotic Process Automation (RPA). Software that automates rule-based, repetitive tasks (data entry, form filling). What this means for you: Often the right tool when "AI" was the proposal but the work is actually deterministic. Cheaper than AI for the kinds of work it handles.

Sentiment Analysis. AI that determines emotional tone in text (reviews, social media, customer messages). What this means for you: Mature category, available as a single API call from major providers. Don't build custom unless your domain has very specific terminology.

Structured Output. Forcing the model to return data in an exact format (usually JSON) your code can parse. What this means for you: This single feature turns AI from a demo into production software. If your vendor can't return structured outputs reliably, they aren't ready.

Synthetic Data. Artificially generated data used to train AI when real data is scarce, private, or insufficient. What this means for you: Useful in narrow cases (rare-event modeling, privacy-sensitive domains). Not a substitute for real data when you can get it.

Token. How AI models measure text. About three-quarters of a word. A 2,000-word article is around 2,700 tokens. What this means for you: You're billed by tokens. The pricing page tells you cost per million in and per million out. The model cost is usually 10-20% of total project cost.

Tool Use (Function Calling). When an AI model decides to call an external function — a database query, an API call, a calculator — instead of just generating text. What this means for you: Tool use is what turns a chat model into something useful for business workflows.

Vector Database. A database designed to store and search embeddings. What this means for you: You probably don't need a separate vector database for your first AI project. Modern context windows are big enough to skip retrieval on small to mid-sized document sets.

Zero-shot, few-shot, fine-tuned. Three ways of teaching a model. Zero-shot: ask with no examples. Few-shot: include 2-5 examples in the prompt. Fine-tuned: train on hundreds or thousands of examples. What this means for you: Try zero-shot first. Few-shot if mediocre. Fine-tuning only after exhausting prompt design. Most teams should never reach the third step.

What this glossary leaves out

I've left out terms that don't change a business decision. You don't need to know what an attention head is to buy AI. You don't need to know the difference between a transformer and an MLP to write a procurement memo. The vocabulary that actually matters is the vocabulary that affects what you build, what you buy, and how much you pay.

The pattern across these terms: AI in 2026 isn't a magic black box. It's a stack of tools with specific names, specific tradeoffs, and specific costs. The vendors who use the words to confuse you are usually the ones with thinner offerings. The vendors who explain things plainly are usually the ones who've shipped the most. Use this glossary to tell those two apart.

If you walk into your next AI vendor meeting and they use a word you don't see in this list, ask them to define it without using other AI words. Watch what happens. The good ones can. The bad ones can't.