Every AI product page assumes you already speak the language. The pricing table bills per token, the changelog celebrates a bigger context window, the sales deck promises agents grounded by RAG — and none of it tells you what those words mean for the only question that matters: will this tool do the job.
So here is AI terms explained for buyers, not researchers. This is deliberately not an A-to-Z dictionary. It is a working vocabulary of about two dozen terms, organized around the four questions people actually ask while evaluating a tool: what am I talking to, why is it wrong sometimes, what do the spec numbers mean, and what does deployment involve. Each entry is a few sentences of plain English plus the reason the term matters at purchase time.
Read it straight through, or jump to the section you need. Where a term is easier to grasp through a real product, the example links to its listing. ToolPotion tracks 13,000-plus AI tools, updated daily, so every example has a current page behind it.
What you're talking to: core AI terms explained
The first confusion is naming. "Model," "LLM," "chatbot," "assistant," and "agent" get used interchangeably in marketing copy, and they are not the same thing.
Model
The trained system itself: a very large statistical function that turns input into output, produced by pushing enormous amounts of data through a training process. Products are wrappers around models: a chat app, an API, and a browser extension can all sit on the same one. This matters when comparing tools, because the model sets the ceiling on quality while the product decides the workflow. You can browse current AI models as their own category in the directory.
Large language model (LLM)
A model trained mostly on text to predict what comes next in a sequence. That turns out to be enough to draft, summarize, translate, and write code. Claude Opus and Gemini 3.1 Pro are current frontier examples. The directory also lists smaller open LLMs you can run yourself. When a vendor says "powered by an LLM," the useful follow-up is which one, and whether you can switch it later.
Chatbot
An interface pattern, not a technology: a message box wired to a model, usually scoped to one job, like answering customer questions. Modern chatbots are built on LLMs, but the word predates them, which is why it can describe anything from a rigid FAQ script to a genuinely capable support system.
Assistant (or copilot)
A general-purpose chat product aimed at one person's daily work: drafting, summarizing, answering, usually living inside the apps where that work happens. ChatGPT is the canonical standalone assistant. "Copilot" typically signals an assistant embedded in an existing product. The buying distinction from a chatbot is audience: assistants serve your staff, chatbots usually serve your customers.
AI agent
Software that takes a goal, plans steps, calls tools, and checks its own results: it acts rather than just answers. Claude Code reads a codebase, edits files, and runs commands. Intercom ships a support agent, Fin, that works conversations rather than just replying to them. The trade-off is blast radius: a system that can act can fail by acting, so permissions and review steps matter more here than anywhere else. The AI agent category is the fastest-moving section of the directory.
| Term | What it does | Buy it when |
|---|---|---|
| Chatbot | Answers questions in one scoped domain | You need customer-facing Q&A at scale |
| Assistant | Helps one person across many tasks | You want staff productivity, low risk |
| Agent | Plans and executes multi-step work | A workflow is worth automating end to end |
Multimodal
A model that handles more than text: images, audio, or video, as input, output, or both. Gemini 3.1 Pro reads and reasons over mixed media. On the generation side, Kling 3.0 turns text and reference images into video with audio. If your work involves screenshots, PDFs with charts, or recordings, multimodal support is a requirement, not a nice-to-have.
Why it's wrong sometimes: hallucination, context, and cutoffs
These four terms explain almost every disappointing AI output you will ever see, and they are the vocabulary of any serious vendor conversation about accuracy.
Hallucination
A confident, fluent, wrong answer. The model is generating plausible text, not consulting a database of facts, so it can invent citations, statistics, and product details without any signal that it is doing so. For factual work, prefer tools that cite retrievable sources, and treat any uncited specific as unverified until you check it.
Context window
How much text, measured in tokens, the model can consider at once: its working memory, distinct from its knowledge. Windows have grown fast. Claude Opus advertises a one-million-token window, enough for an entire codebase or several books in a single request.
A bigger context window raises the ceiling on what you can paste in. It does not guarantee the model weighs all of it.
In practice, models can lose track of details buried in the middle of very long inputs, so "fits in the window" and "gets used well" are different claims.
Training cutoff
The date the model's training data ends. Anything published after it is invisible to the model unless the product bolts on web search or retrieval. Why you care: if the tool must know current prices, laws, or news, ask the vendor exactly how fresh information gets in. "The model is smart" is not an answer.
RAG (retrieval-augmented generation)
Search first, answer second: the system finds relevant passages in your documents, then has the model write its answer using them. This grounds responses in your actual content, shrinks (but does not eliminate) hallucination, and stays current the moment a document changes, with no retraining. When a vendor says "trained on your data," ask whether they mean RAG. Usually they do, and for internal knowledge, RAG is usually what you want.
Prompt and system prompt
A prompt is whatever you send the model. A system prompt is the standing instruction sheet the vendor places in front of every conversation: tone, rules, guardrails. A surprising number of niche AI products are essentially a system prompt over someone else's model. That can still be worth paying for, but it should be priced like packaging, not like research.
Reading the spec sheet: tokens, parameters, inference
The numbers on an AI pricing page are not marketing decoration. Three of them directly set your bill.
Token
The unit models read and emit: a chunk of text, where a common English word is often one token. A workable rule: 1,000 tokens is roughly 750 English words. API pricing is quoted per million tokens, with input and output priced separately, which is why long documents in and long answers out are what drive cost.
Parameters
The internal numeric weights a model learned during training. Parameter count is the standard shorthand for model size. Muse Glimmer, for instance, is a 30-billion-parameter model built to run agent tasks on consumer hardware. More parameters generally buys capability at the price of speed and cost, and a small model that handles your task well beats a large one you are overpaying to underuse.
Inference
Running the trained model to produce an output. Training happened once, at the vendor's expense; inference happens on every request, at yours. Latency and per-request cost are inference properties, so when a vendor advertises cheaper or faster inference, that claim is about your bill and your users' wait time, not about answer quality.
Fine-tuning
Additional training on your own examples to shift a model's behavior: tone, format, domain habits. It is good for consistent style and structured output, and a poor way to teach facts that change, which is RAG's job. The hidden cost is maintenance: every base-model upgrade can mean redoing the work.
Reasoning model
A model that spends extra compute working through intermediate steps before answering, which measurably helps on math, planning, and hard code. Some are hybrids: Claude Opus can answer quickly or think longer as needed. The trade-off is real: reasoning passes are slower and cost more tokens, so routing a simple extraction task through one is paying for thought you don't need.
Benchmark
A standardized test that produces the leaderboard scores in launch posts. Treat them as rough capability signals, not predictions about your workload: vendors naturally showcase the benchmarks they win. Twenty of your own real tasks, run against two candidate tools, tell you more than any leaderboard.
Deployment terms: APIs, open weights, and where it runs
The last set of terms decides cost structure, privacy posture, and how much engineering you are signing up for.
API
Programmatic access: your software calls the model over the internet and pays per token, instead of a person paying per seat for an app. The same vendor often sells both, and they are different budgets: seats scale with headcount, API costs scale with usage. Know which one you are actually shopping for before comparing prices.
Open weights vs. closed
Open-weight models publish the model files: download them, run them on your own hardware, keep data in-house, pay no per-token fee to the maker. Gemma 4 offers lightweight open models built from the same technology as Gemini, and Stability AI publishes open generative models for images, video, and audio. Closed models are rented through the vendor's servers: less control, far less operational burden, and the strongest closed models still tend to lead on raw capability.
On-device vs. cloud
Where inference physically happens. Cloud means requests leave your network; on-device or self-hosted means they don't, which is why compliance-heavy teams care, and why models like Muse Glimmer that run on consumer hardware exist. Expect a capability gap: the largest models do not fit on a laptop.
Embedding
Text converted into a list of numbers positioned so that similar meanings land near each other. Embeddings are how software searches by meaning instead of exact keywords, the reason a query for "refund rules" can find a document titled "returns policy."
Vector database
Storage built to hold embeddings and find nearest matches fast. It is the retrieval half of most RAG systems. Chroma is an open-source example combining vector, full-text, and metadata search. If a vendor says they "RAG over your docs," one of these is in the stack. Ask where it runs and what it stores.
Framework
Code scaffolding for wiring models, tools, and retrieval into an application: the build-it-yourself path, chosen early by teams doing in-house work. Browse the AI framework category to see the current options. If nobody on the team writes code, visual builders like Microsoft Copilot Studio cover common cases, and the directory keeps a full task page of tools for building AI agents spanning both approaches.
Observability and evals
Observability is tracing what a model or agent actually did: every step, tool call, token cost, and delay. Evals are graded tests of output quality, run repeatedly so regressions surface before customers do. Langfuse is an open-source platform covering tracing, prompt management, and evaluation. If you are buying an agent rather than building one, the equivalent question for the vendor is simple: when the agent gets something wrong, what exactly will you show me?
Frequently asked questions
Is ChatGPT an LLM?
Strictly, no: ChatGPT is an application built on top of LLMs from OpenAI's GPT family. The distinction matters because features like memory, file uploads, and web browsing belong to the app, not the model, and a different product using the same model may lack all of them.
What is the difference between an AI agent and a chatbot?
A chatbot answers; an agent acts. Agents plan multi-step work, call other software, and change things in real systems, which makes them more useful and riskier at once. The practical buying question is permissions: what an agent is allowed to touch, and what gets reviewed before it does.
Do more parameters mean a better AI model?
Not for your task, necessarily. Larger models are generally more capable in the abstract, but they cost more and respond slower, and plenty of everyday work — classification, extraction, routine drafting — runs well on small models. Test the cheapest model that might work before defaulting to the biggest.
Can I run an AI model on my own computer?
Yes, if it has open weights and a modest size: small open models such as the Gemma family are designed for it, and some 30-billion-parameter models now target consumer hardware. Expect weaker results than frontier cloud models, in exchange for zero per-token cost and data that never leaves your machine.
