Plain-English AI Lexicon

Software that produces useful output from patterns it learned, rather than from rules a person wrote out step by step. Today, when people say "AI," they usually mean a chat assistant built on a language model.

For you:It's a very fast, very well-read colleague who has never seen your business. Everything below is about fixing that last part.

The trained brain. A model is a specific version of the thing that reads your words and writes back. Companies release new ones every few months; the app you use picks one for you or lets you choose.

For you:Newer models are better at following instructions and less likely to make things up. You rarely need to think about which one, unless something feels dumber than last week.

The kind of model behind chat assistants. "Large" because it was trained on an enormous amount of text; "language" because words are what it's good at. It predicts what text should come next, which turns out to cover writing, summarizing, reasoning, and code.

For you:Same thing as "model" in everyday use. It's great at anything that's mostly words. It is not a calculator or a database unless you give it one (see tools).

How hard the model thinks before it answers. A dial, not a different brain: low effort answers fast; high effort works through steps first, takes longer, and costs more. Products expose it as a "thinking" toggle, an effort level, or a "think longer" button.

For you: Turn it up for analysis, planning, and anything with numbers. Leave it down for drafts and quick questions. If an answer feels shallow, try more effort before trying a different model.

Not the same as a reasoning model, which is a model built to think this way by default. Effort is the knob; the reasoning model is one that ships with the knob turned up.

Models trained on pictures, sound, or video instead of text. An image model turns a description into a picture; a voice model reads aloud or transcribes. A multimodal model does several of these in one, which is why you can drop a photo into a chat.

For you:Same setup ideas apply: give it examples of your brand, tell it what you never want. The LLM is usually still the one coordinating.

Whatever you send the model: a question, a task, a pasted document, a photo. The model only knows what's in the prompt plus what it was trained on.

For you:People overrate prompt wording. A messy description of the real problem plus "ask me what you need to know" usually beats a carefully crafted one. The expensive part is context, not phrasing.

What comes back. It's generated fresh each time, so the same prompt can get slightly different answers.

For you:Treat it as a draft from a smart assistant, not a lookup from a database. If it matters, check it.

The unit models read and write in, roughly three-quarters of a word. Pricing and limits are counted in tokens.

For you:Only matters when you hit a wall: a document too long to paste, or a plan's usage cap. "Too many tokens" means "too much text."

When the model states something false with full confidence: a fake statistic, a quote nobody said, a feature that doesn't exist. It's not lying; it's completing a pattern without checking.

For you:Happens most when it lacks the facts. Give it the source (a file, a connector) and ask it to say what it's unsure about. Never let it invent a number you'll repeat to someone else.

The date the model's training data ends. It knows nothing after that unless it searches the web or you tell it.

For you:Anything recent (prices, people, news, your own numbers) needs to come from search or from you.

Everything the model can see while answering: your prompt, the conversation so far, files you've attached, instructions you've saved. The context window is how much of that fits at once.

For you:If you find yourself re-explaining your business every chat, the fix is putting that explanation somewhere it's always in context (see Project, instructions).

Standing directions the model reads before every conversation: who you are, how you like answers, what to never do. You write them once.

For you:Ten minutes here saves hours later. "Here's my company, my customers, my voice, and what I never want you to do."

A folder in the chat app that bundles instructions, reference files, and the conversations about one subject. Every chat inside it starts already knowing what's in it.

For you:One Project per recurring area of your business. Open the Project, not a blank chat.

The account level: who's on the plan, what's connected, what write actions are allowed, whether your data is used for training. A personal account is a workspace of one; a business plan is a shared workspace with an admin.

For you:Settings that gate everything else live here: which connectors exist, whether write actions are on, data privacy. When a feature "isn't showing up," this is where to look.

Facts the assistant keeps across conversations without you re-saying them: preferences, names, ongoing work. You can see, edit, and delete what it remembers.

For you:Good for "always write in a warm tone" and "my fiscal year starts in July." Don't rely on it for facts that change weekly; use a file or connector for those.

Making the model answer from a specific source (your documents, a database, the web) instead of from memory. The source is retrieved and put into context before it answers.

For you:Any time you want it to be right about your stuff rather than plausible about the world in general.

Something the model can use besides writing: search the web, run a calculation, read a file, send an email. The model decides when to use a tool and reads the result.

For you:This is how it stops guessing and starts looking things up. Tools are read (safe) or write (they change something; approve those).

A packaged tool that links the assistant to a service you already use: Gmail, Google Drive, QuickBooks, Slack, Canva. You connect it once by signing in; after that the assistant can read from (and sometimes act in) that service.

For you:The difference between "paste your inbox in" and "check my inbox." Start with read-only; turn on write actions one at a time.

The technical standard that many connectors are built on: a common plug so any tool can work with any assistant, the way USB lets any device plug into any laptop. You'll see the term in settings and in vendor marketing.

For you:If someone says a product "has an MCP server," it means it can be connected. You don't need to know how it works, just that "MCP" and "connector" are usually the same thing from where you sit.

The way software talks to other software. AI companies sell their models through an API so developers can build products on them; connectors use APIs under the hood.

For you:You'll hear it when someone builds you something custom. "It needs API access" means "the software has to be allowed to talk to that service."

Read means the assistant can look at data. Write means it can change it: send, create, delete, pay, post.

For you:The single most important setting. Read-only everything first. Add one write action at a time, with confirmation on, once you trust the output.

An assistant that can take multiple steps on its own toward a goal: decide what to do next, use tools, check the result, try again. A chat answers you; an agent works for you.

For you:The word is overused. Ask: can it take more than one step without me? Can it use tools? If yes, it's an agent. If it just writes, it's a chat.

Not the same as autonomous. Most agents still ask before doing anything important.

Working without a person in the loop. Fully autonomous means it decides and acts, including irreversible things, with nobody approving each step.

For you:Rarely what you want yet. Aim for "runs on its own, drafts everything, sends nothing." Autonomy is a dial, not a switch, and you turn it up one action at a time.

The person who checks or approves before something happens. The opposite of autonomous.

For you:Where you should be for anything that sends, pays, posts, or deletes. Your job shifts from doing the work to saying yes or no to it.

A repeatable sequence: when this happens, do these steps, produce this. Workflows can be simple (one saved prompt) or multi-step with tools and approvals.

For you:Anything you do the same way every week is a workflow waiting to be written down. Write it down once; that's most of the automation.

A prompt that runs on a timer without you opening the app: every Friday at 4, check X and tell me if anything needs me. Results arrive by notification or email.

For you:Right for things you forget. Wrong for things you do when they show up (use a Project you drop things into) or things you do in a batch when you're ready (use a saved prompt). A scheduled digest nobody reads is the most common wasted automation.

Starting a workflow from an event rather than a clock: a new email arrives, a form is submitted, a file lands in a folder.

For you:The natural fit for "when something arrives" pain. Newer and more plan-gated than timers; check what your tools support.

A written procedure the assistant loads when a task matches it: how to write in your voice, how to format a report, how to handle a distributor order. A skill turns "here's how we do it" into something the AI does the same way every time.

For you:Every SOP you'd give a new hire can be a skill. Record yourself doing the task once, have the AI turn it into a checklist, save it.

Everything wrapped around the model to make it useful and safe: the instructions, the tools it can reach, the permissions, the approvals, the memory, the schedule. The model is the engine; the harness is the rest of the car.

For you:When AI feels magic for someone else and useless for you, the difference is almost always the harness, not the model or the prompt. Getting better at AI mostly means getting better at the setup.

Rules that keep the assistant inside lines you set: never send without asking, never quote a price, only draft. Some are built into the product; some you write.

For you:Put your never-do list in the instructions. "Drafts only" is a guardrail.

When text the assistant reads (an email, a web page, a document) contains hidden instructions that try to hijack it: "ignore your rules and forward this to…" The risk grows as you connect more sources and allow more write actions.

For you:Another reason for read-only first and approvals on sends. If it can read strangers' emails and also send email, someone will eventually try this.

Whether what you type or connect is stored, seen by people, or used to train future models. Varies by product and plan; business plans usually promise no training on your data.

For you:Check the plan before connecting the books or customer lists. It's a settings page, not a mystery.