Thirty-plus AI terms explained in plain English, each one tied to a real decision you might actually be making at work.
This is a working AI glossary built for business professionals, not developers: 30+ terms in plain English, organized into AI Fundamentals, Popular AI Tools (ChatGPT, Claude, Copilot, Gemini), Using AI Day-to-Day, and AI Governance & Risk. Every entry ties the definition to a real work decision. Two terms with their own dedicated Future Factors guides, RAG and Constitutional AI, get a short definition and a link here rather than a rewrite.
A benefits coordinator sits in on a vendor demo for a new HR chatbot. The salesperson says the tool “uses RAG to ground its responses and reduce hallucination, all inside a generous context window.” She nods, writes “ask about this” in the margin, and never does. Three months later the tool is live, mostly unused, and when her director asks why adoption is low, she says people are busy. That’s true. It’s also not the real reason. She never fully understood what she’d approved, and wasn’t about to say that out loud in a room full of people nodding along with her.
That scene isn’t really about one nervous meeting. In Future Factors workshops, this is usually the sticking point long before anyone gets to a workflow or a tool decision: capable, experienced professionals who stop asking questions the moment the vocabulary shifts from “what does this do for my team” to “how does the model actually work.” Jargon doesn’t just sound unfriendly. It quietly decides who keeps steering the conversation and who defers.
Not understanding a term isn’t a personal failing. It’s a vocabulary gap, and vocabulary gaps close the same way any other gap does: with a plain-English answer you can actually find when you need it.
You don’t need to memorize a glossary or pass a technical exam to use AI well at work. You need enough of a working vocabulary to ask the second question in a meeting instead of nodding through it, and somewhere to check the rest when an unfamiliar term shows up in a vendor pitch, a policy document, or a colleague’s Slack message. That’s what this article is built to be: a reference you bookmark and actually return to, organized the way these terms come up in practice, starting with the handful everything else depends on.
These seven terms are the ones the rest of this glossary assumes you already have some sense of. Get comfortable with these first; nearly everything else in this article is a variation or an application of one of them.
The broad, umbrella term for any system that performs tasks normally requiring human intelligence: recognizing images, understanding language, making a recommendation, spotting a pattern. It’s worth knowing it’s an umbrella, because “AI” covers everything from the spam filter that’s quietly worked in your inbox for fifteen years to a chatbot that drafts a client email. Those are different tools with different risk profiles, and treating them as one category is where a lot of AI policy conversations go wrong before they even start.
A way of building AI where the system improves at a task by learning from examples, rather than being explicitly programmed with rules for every situation. A fraud-detection model that gets better at flagging unusual transactions the more data it sees is machine learning. The takeaway: an ML system is only as good as what it learned from, so if your historical data has old habits baked into it, the model repeats them.
The subset of AI that creates new content (text, images, audio, code) rather than only classifying or predicting something about content that already exists. ChatGPT, Claude, Copilot, and Gemini are all generative AI, the newer, headline-grabbing half; the older half has quietly run your CRM’s lead scoring and your bank’s fraud alerts for years.
The specific type of generative AI trained on enormous amounts of text to predict what word or phrase comes next, which turns out to be enough to hold a conversation, summarize a document, or draft an email. Every major chat assistant you’ve heard of is built on an LLM. “Which LLM powers this tool” is a fair, specific vendor question, and one that can’t answer it plainly is worth a second look.
The underlying structure, loosely inspired by how neurons connect in a brain, that most modern AI models are built from: layers of simple mathematical units that each adjust how much weight to give a piece of information. You don’t need the math. What’s worth knowing is that a bigger network trained on more data generally gets more capable, but it’s still pattern-matching at enormous scale, not reasoning the way a person does, which is part of why it can sound confident while being wrong.
A set of step-by-step rules a computer follows to complete a task. Every AI system runs on algorithms, but not every algorithm is AI: the rule that sorts your inbox by date is an algorithm with nothing intelligent about it. Worth knowing so a headline like “the algorithm decided” doesn’t automatically mean an AI model was involved.
When a model states something false in the same confident tone it uses for something true. It isn’t lying in the human sense: it’s predicting plausible-sounding text, and sometimes the most plausible answer is simply wrong. This is the term with the most direct dollar cost. It’s why you check the actual invoice total or contract figure before a client email goes out with a number the model drafted, not after.
These are the four names you’ll hear most in any AI conversation at work, and they’re not interchangeable. Each is built by a different company and genuinely strongest in a different place. Knowing the difference is less about picking a favorite and more about knowing which one to reach for.
OpenAI’s general-purpose chat assistant, and the tool most people mean when they say “AI” without naming one.[1] It’s built for flexibility across a wide range of one-off tasks: drafting, brainstorming, coding, research, explaining something back to you a different way until it clicks. Reach for it when the job doesn’t live inside one particular piece of software. For a closer look at what it can currently do, see our plain-English breakdown of ChatGPT 5.2.
Anthropic’s assistant, built by the company that also developed Constitutional AI, covered further down this page.[2] Claude is strongest when the work is context-heavy, text-heavy, and needs coherence over a long document or an extended back-and-forth: reviewing a long contract, holding a multi-step research task together, editing a report without losing the thread. It’s a weaker fit for a quick lookup or a task that lives entirely inside a spreadsheet. Claude Cowork extends this into a more autonomous working style; we cover what that setup looks like here.
Microsoft’s assistant, built into Word, Excel, Outlook, and Teams.[3] Its real advantage isn’t the underlying model, it’s that it’s grounded in your organization’s own email, documents, and meetings, and it acts inside the apps you already have open. The decision rule: use it to read, compare, find, challenge, and verify the work you already have, not primarily to generate new work from nothing. If the job lives in an Office file, Copilot has an edge no general chat assistant does.
Google’s assistant, built into Gmail, Docs, Sheets, Slides, and the rest of Workspace, the same way Copilot is built into Microsoft 365.[4] If your organization runs on Google rather than Microsoft, Gemini earns its place the same way: drafting and summarizing right where your email and documents already sit, instead of a separate window you copy and paste into.
| Tool | Built by | Lives inside | Reach for it when |
|---|---|---|---|
| ChatGPT | OpenAI | Its own app, browser, or API | The task is general and doesn’t live inside one specific tool |
| Claude | Anthropic | Its own app, or Claude Cowork for file-based work | The job is long, text-heavy, and needs coherence over time |
| Copilot | Microsoft | Word, Excel, Outlook, Teams | The work already lives in an Office file or inbox |
| Gemini | Gmail, Docs, Sheets, Workspace | Your organization runs on Google Workspace |
A starting map, not a ranking. Most professionals end up using more than one, depending on which software the task already lives in.
Match the assistant to where the work already lives, then judge the model itself second.
Once a tool is open in front of you, a different set of words starts showing up: in settings menus, in what a colleague says about “the context window filling up,” in a policy about “fine-tuning.” These ten terms are worth knowing once you’ve moved from “what is AI” to “I’m using this thing right now.”
The practice of writing instructions that reliably get you useful output. It sounds more technical than it is: mostly, it’s giving the same context you’d give a new hire on day one, the goal, the audience, the constraints, and an example of good, instead of a one-line request and a hope it reads your mind.
Weak: “Write a summary of this report.”
Specific: “Summarize this report for a CFO who has three minutes before a board meeting. Lead with the two numbers most likely to get questioned. Flag anything you’re not fully certain of instead of smoothing over it.”
Same document, same tool. The difference is audience, priority, and an explicit instruction to flag uncertainty instead of hiding it.
How much text or conversation a model can actually “see” and hold in mind at once, measured in tokens (more on those next). It’s why a chat running for hours starts forgetting how it began, and why pasting a two-hundred-page report into a tool with a small context window gets you a shallow summary instead of one that engages with page 140.
The unit a model actually reads text in, roughly three-quarters of a word in English. You’ll mostly meet this word on pricing pages and context-window limits; it’s why a tool’s input limit is quoted in tokens rather than pages or words.
A setting that controls how predictable or varied a model’s output is. Lower temperature gives safer, more consistent answers, useful for a compliance summary. Higher temperature gives more varied, surprising answers, useful for brainstorming ten tagline directions. Most everyday chat tools hide this setting and pick a sensible middle ground; it mostly matters if you’re building on top of a model through an API.
The instructions set behind the scenes before you type a word, defining the assistant’s role, tone, and rules. When your IT team says “we’ve configured your organization’s Copilot,” this is usually what they mean: a standing set of instructions applying to every conversation, not something you repeat yourself.
Software that can take multi-step action toward a goal with some independence, rather than just answering one question and stopping. An agent can check a status, decide what to do next, and act, inside limits you set. The gap between a chatbot and an agent is the gap between “tell me what’s overdue” and “flag what’s overdue and draft the follow-up, but don’t send it without me.” We cover agents versus subagents in more depth here.
A model that can work with more than text: images, audio, sometimes video. This is why you can hand a tool a screenshot of a broken chart and ask what’s wrong, or a photo of a whiteboard and ask it to type up the notes, instead of describing it first.
A technique that grounds a model’s answer in your actual documents at the moment you ask, instead of relying only on what it learned during training. It’s one of the main tools used to reduce hallucination. Read our dedicated guide to RAG for the full plain-English explanation.
Training an existing model further on your own examples so it specializes in a particular task or tone. It’s genuinely powerful and genuinely rare for most business users to need directly. Most of what feels like “customizing” an AI tool at work, a style guide, a set of standing instructions, a connected knowledge base, is prompting or configuration, not fine-tuning. A vendor claiming to have “fine-tuned a model just for you” is worth a follow-up question.
A version of ChatGPT you configure yourself with specific instructions, knowledge files, and a narrower job, no fine-tuning and no code required. It’s the accessible middle ground between a generic chat and a fully custom-built tool. Our step-by-step guide walks through building one.
This is the vocabulary that shows up once a tool moves from “something I use” to “something leadership is asking questions about.” These terms matter because they’re usually what stands between a good pilot and a policy nobody can actually follow.
When a model’s output systematically favors or disadvantages a particular group, usually because the patterns it learned from reflect that same imbalance. A resume-screening tool trained mostly on a company’s past hires will tend to favor candidates who look like those past hires, whether or not that was the intent. For an HR team, this turns “should we use this tool” into “who reviews what it filters out, and how often.”
The broader set of principles guiding whether and how AI should be used at all, not just whether one output is biased. It’s the conversation that belongs before a tool is purchased, not only the audit after something goes wrong.
A design choice where a person reviews or approves an AI-touched output before it goes live, instead of letting the system act unsupervised. It’s the practical difference between “AI drafted this” and “AI sent this,” and usually the cheapest safeguard available for any new AI workflow.
The explicit limits placed around what an AI system is allowed to do or say. For a marketing team, that might mean a brand-voice guardrail that keeps generated copy on-tone, plus a hard rule that it can never touch the CRM or publish without a person clicking send. Guardrails turn “we gave it access” into “we gave it access to exactly this, and nothing else.”
The organization-level policies that decide which tools get approved, who owns oversight, and how risk gets reviewed over time. It’s the phrase that shows up in a board deck, and it usually means “who is accountable if this goes wrong,” not a specific technical control.
What actually happens to the information you type into an AI tool: whether it trains future models, how long it’s retained, who else can see it. Worth checking before you paste a client’s financial details into a personal AI account rather than the enterprise version your organization has reviewed.
Employees using AI tools without IT or leadership’s approval or visibility. It’s rarely malicious. Usually it’s someone in sales or operations solving a real, immediate problem with whatever free tool is available, because the approved option is slower or doesn’t exist yet. It’s also the main reason AI governance policies get written in the first place.
A training approach, developed by Anthropic, that gives a model a written set of principles to check its own responses against, rather than relying only on human feedback after the fact. Short version: it’s one way a model is taught to say no to a harmful request. We cover how it actually works in a dedicated article.
| What AI does | What you still own | How it gets checked |
|---|---|---|
| Drafts a first-pass summary, email, or report from your documents | Deciding what’s worth saying, and whether a flagged number is right | A human reads it fully before it goes out under their name |
| Flags what looks unusual or biased in a dataset | Deciding whether a flag is a real problem or a known exception | Every flagged item gets a human look, not a spot check |
| Takes a defined next step as an agent, within limits you set | Setting those limits, and what it’s never allowed to do unsupervised | Human-in-the-loop approval before anything irreversible happens |
Treat an AI-drafted fact the way you’d treat one from a new hire’s first week: useful, not yet verified.
None of this is a case against using AI: a policy, a guardrail, and a trained facilitator generally get a team to safe, confident use faster than trial and error alone does. Before approving a new AI tool, three questions from this section are worth asking out loud, before anyone signs anything:
You’re not going to remember all thirty of these terms after one read-through, and that was never the point. Bookmark this page. The next time someone in a meeting uses a term you don’t recognize, the move isn’t to nod and hope it doesn’t come up again. Write the word down, and check it here as soon as the meeting ends.
A couple of limits are worth naming too. This glossary won’t make you able to build or audit a model yourself, and it deliberately skips deeper engineering vocabulary, things like training pipelines or vector databases, that a data science team needs and a marketing manager evaluating a vendor doesn’t. What it will do is get you through the moment a meeting turns technical.
If you’re evaluating a new AI tool for your team, the fastest way to use this article isn’t top to bottom. Pull two short lists: from AI Fundamentals, enough to understand what the vendor is actually claiming the tool is; from AI Governance & Risk, the questions to get answered about bias, data privacy, and human-in-the-loop review before anyone signs anything. Those two sections do more work on a vendor call than the other two combined.
Pick one term from this list you’ve been nodding along to for months without actually being sure. Look it up properly today. That’s the whole method, one term at a time, until the room stops feeling like it’s speaking a different language.
An LLM, large language model, is the AI trained on huge amounts of text to predict and generate language; it’s what ChatGPT, Claude, Copilot, and Gemini all run on. A hallucination is when that same model states something false in the same confident tone it uses for something true, because it’s predicting plausible text, not checking facts against reality. Together, those two terms explain most of what people mean by ‘AI got it wrong’: the model is working exactly as designed, it’s just not a fact-checker.
Think of it as nested categories, not three separate things. Artificial intelligence is the broadest term: any system doing a task that normally needs human intelligence. Machine learning is a way of building AI, where the system improves by learning from examples instead of following pre-written rules. Deep learning is a specific technique within machine learning that uses layered neural networks, and it’s the technique behind almost every generative AI tool in use today. So an LLM is deep learning, which is machine learning, which is AI, but the reverse isn’t true: a rules-based chatbot from a decade ago was never machine learning at all.
It’s the practice of writing instructions that reliably get useful output from an AI tool. The name makes it sound like a technical skill reserved for developers, and that’s the main misconception worth clearing up: it isn’t. If you’ve ever rewritten a request to a colleague because the first version was too vague, you’ve already done the core of it. The other habit worth building is treating your first prompt as a draft, not a final answer: asking a follow-up like ‘make this shorter’ or ‘now write it for someone who’s never seen this before’ is still prompt engineering, and it’s usually where the actually useful output shows up.
RAG is a technique that grounds a model’s answer in your actual documents at the moment you ask, rather than relying only on what it learned during training months or years earlier. It’s one of the most effective tools for reducing hallucination, because the model pulls from a real source it can point back to instead of generating an answer purely from memory. See our dedicated guide to RAG for the complete picture.
No, and treating this glossary as a technical requirement misses the point. You can use ChatGPT, Claude, Copilot, or Gemini productively without knowing what a token is. A working vocabulary buys you the ability to ask a sharper question, catch a vague claim in a vendor pitch, or notice when a policy document uses a term incorrectly. If you only have time to learn one section well, make it AI Governance & Risk rather than AI Fundamentals: those terms are the ones most likely to come up in a decision that actually affects your team.
This glossary defines each term independently rather than leaning on any single outside source. Company and product facts, what ChatGPT, Claude, Copilot, and Gemini are and who builds them, are checked against each provider’s own current product pages, listed below, current as of September 2026. This article does not reproduce any third-party glossary’s term list, category names, or wording; the term selection, groupings, and business-use framing are Future Factors’ own. Definitions are intentionally kept business-focused rather than exhaustive: deeper engineering terms (training pipelines, vector databases, model weights) are left out on purpose, since they matter more to a data science team than to the business professional this glossary is written for.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.