The Plain-English AI Glossary for Business Owners

Model call (a single AI answer)#

Asking the AI one question and getting one answer back. Information goes in, an answer comes out, done. No back-and-forth, no tools, no deciding what to do next.

ExampleSummarize this email. Translate this paragraph. Reword this post. These are one-and-done jobs.

Watch outMost everyday tasks are simple AI answers. Turning them into an “agent” just adds delay and more ways to go wrong.

Agent#

An AI that does more than answer once. It can use tools, take actions, look at the results, and keep going on its own until the job is finished. It works through a task instead of giving a single reply.

ExampleA support assistant that looks up your order, checks the shipping status, and opens a ticket if it cannot solve the problem, all without being told each step.

Watch outOnly reach for an agent when a single answer genuinely will not do. Agents are powerful, but slower and harder to predict.

Tool#

Something outside the AI that it is allowed to use, like a web search, a calculator, your calendar, or your files. “Tool calling” simply means the AI choosing to use one of these and acting on what it finds.

ExampleGiving an assistant a search tool so it can look up today’s prices instead of guessing from memory.

Watch outThe AI only knows what a tool does from how it is described. A vague description leads to the wrong tool being used.

Context window#

How much the AI can keep in mind at one time, including your instructions, the conversation so far, and any documents. Think of it as short-term memory with a firm size limit.

ExamplePaste a 300-page document into a small context window and the AI simply cannot see all of it at once.

Watch outA surprising number of AI mistakes trace back to this. When it fills up, the AI starts forgetting earlier details.

Token#

The unit the AI reads and charges by, roughly three quarters of a word. Both cost and the context window limit are measured in tokens, not words or pages.

ExampleAbout 1,000 tokens is roughly 750 words, or a page and a half of text.

System prompt#

The standing instructions that tell the AI who it is and how to behave, separate from your specific request. It is the job description; your message is the task.

Example“You are a careful assistant for a law firm. Never guess at facts. Always note where information came from.”

Watch outWhen these instructions grow long and full of “but not when...” exceptions, it usually means the AI is being asked to do too many jobs at once.

Single Agent#

One assistant with one job and a small set of tools. The simplest and most common setup, and usually the right place to start.

ExampleA single assistant that answers product questions using two or three tools.

Watch outWhen you keep piling on tools until the instructions sprawl, that is the sign you have outgrown it, not a reason to write longer instructions.

Multi-Agent (in sequence)#

Several specialized assistants working in a fixed order, where each one’s output becomes the next one’s starting point. Like an assembly line.

ExampleA contract review line: one assistant pulls out the key obligations, the next flags the risks, a third writes the plain-English summary.

Watch outBreaks down when the order actually needs to change depending on what turns up partway through.

Multi-Agent (in parallel)#

Several assistants working at the same time on separate pieces, with the results combined at the end. Used when speed matters.

ExampleThree assistants research your competitors’ pricing, their customer reviews, and their ads all at once, instead of one after another.

Watch outOnly works if the pieces truly do not depend on each other. Forcing connected work to run in parallel creates conflicts and gaps.

Coordinator (or Router)#

A central assistant that reads each request and sends it to the right specialist. A receptionist for your AI.

ExampleSupport messages routed to billing, technical, or returns specialists, each focused on its own area.

Watch outIf requests do not fall cleanly into categories, the router itself becomes the bottleneck and starts sending things to the wrong place.

Breaking a big task into smaller ones#

A lead assistant splits a large goal into smaller parts, hands each to a specialist, then pulls the answers back together. (Often called hierarchical task decomposition.)

Example“Which three countries should we expand into?” split into market size, regulations, shipping, and competition, then combined into one recommendation.

Watch outAssumes the parts are independent. If they really need to inform each other as the work happens, splitting them up front loses that.

Swarm (a panel of assistants)#

Several specialist assistants debate a question over a few rounds, challenge each other, and a facilitator settles on a recommendation.

ExampleMarketing, finance, and operations assistants argue the pros and cons of launching a subscription plan.

Watch outThe trendiest and least proven of these. Slow, costly, and unpredictable. The wrong choice when you need a fast, repeatable answer.

Orchestrator and helpers#

The setup that has actually proven itself in real use: one main assistant keeps the full picture and spins up small temporary helpers, each doing a narrow job and reporting back a short summary.

ExampleA research assistant that fires off several quick lookups at once, then writes the report itself from their summaries.

Watch out“More assistants” is not automatically better. Most people get further with one well-built assistant than with a crowd of them.

ReAct (reason, then act)#

The everyday rhythm of most assistants: think about what to do, take an action, look at the result, then decide the next step. The path is not mapped out in advance; it takes shape as the assistant learns more.

Example“Sales dipped last week.” The assistant checks website traffic, then ad spend, then recent price changes, following wherever the numbers lead.

Watch outWithout a cap on how many times it loops, it can keep digging forever. Always set a limit.

Loop#

Repeating a set of steps until a clear finish line is reached.

ExampleAn assistant tidying a messy customer list: it proposes fixes, checks them against your rules, and tries again until it passes or hits a set limit.

Watch outRisky without a reliable stopping point. No clear exit means it can run up cost with no end.

Iterative refinement#

A loop where the work is scored against a quality bar and rewritten until it clears the bar. Not just pass or fail, but steady improvement.

ExampleA copywriting assistant that rates its own draft against your brand guidelines and keeps rewriting until it scores well enough.

Watch outIf the scoring is vague or easy to game, the assistant just inflates its own grade and runs up cost without real improvement.

Review and critique#

A separate “reviewer” assistant checks another’s work, points out weak spots, and gives specific feedback before a person ever sees it.

ExampleA draft report gets checked by a reviewer assistant that flags unsupported claims and missing evidence.

Watch outIf the reviewer has the same blind spots as the writer, it misses the same mistakes. A second opinion only helps if it is genuinely different.

Reflexion (learning from mistakes)#

The assistant looks at its own failure, works out what went wrong, remembers the lesson, and applies it next time, getting better with each attempt.

ExampleAn assistant whose ad draft gets rejected, reads the exact reason it was flagged, fixes that, and resubmits successfully.

Watch outOnly helps when failures follow a real pattern. If the errors are random, there is nothing to learn from.

Plan first, then do#

The assistant writes out a full plan first, which you can review or approve, and only then carries it out. The opposite of ReAct, where the plan takes shape step by step. (Often called plan-and-execute.)

Example“Launch the email campaign: clean the list, write three emails, schedule the sends, then update the CRM.” The whole plan is visible and approvable before anything goes out.

Watch outIf things change faster than the plan runs, an out-of-date plan followed blindly can be worse than no plan.

Human in the loop#

The assistant investigates and recommends, but a person makes the final call on anything risky or unclear. The AI proposes; you approve.

ExampleSmall, clear refunds are handled automatically; large amounts or anything suspicious pauses for a person to review.

Watch outTreat this as a real part of the setup, not an afterthought. Decide in advance which actions always need your sign-off.

Fixed rules (where the AI does not decide)#

Using plain, fixed rules for the parts that must never be wrong, while letting the AI handle the judgment and the wording. A deliberate split between what is locked down and what is flexible.

ExampleRefund eligibility and fraud checks follow fixed rules; the AI only drafts the friendly reply to the customer.

Watch outEligibility, permissions, and anything involving money should never be left to the AI’s judgment alone.

Guardrails#

The hard limits and safety checks placed around an assistant: what it is allowed to touch, say, or do, and what is off-limits.

ExampleA rule that stops the assistant from deleting files or spending money without explicit approval.

Watch outMissing guardrails are how an assistant ends up doing something it cannot undo that nobody intended.

Event-driven assistant#

An assistant that does not wait to be asked. It watches for something to happen and acts the moment it does.

ExampleAn assistant that flags a suspicious charge the instant it appears, rather than when someone complains days later.

Watch outIf the triggers are poorly defined, it either cries wolf constantly or misses the one signal that mattered.

Memory#

How an assistant holds on to information across steps, or across whole sessions, instead of forgetting everything between turns.

ExampleAn assistant that remembers your preferences from last week without you repeating them.

Watch outWeak memory is behind a huge share of “the AI forgot what I told it” moments.

RAG (answering from real sources)#

Letting the AI pull in real documents or data before it answers, so it responds from actual sources instead of general memory. (Short for retrieval-augmented generation.)

ExampleA support assistant that searches your help articles and answers from them, with the source attached.

Watch outThis is the main defense against confident wrong answers. If it pulls the wrong source, the answer is wrong too.

MCP (how AI connects to your tools)#

The shared standard for connecting an AI to outside tools and data. It is the plumbing that lets an assistant reach your apps and files. (Short for Model Context Protocol.)

ExampleThe connectors that let an assistant read your calendar or documents run on MCP.

Watch outWhen someone says an assistant “connects to” your software, this is usually how.

Evaluation (testing it properly)#

Testing an assistant against real cases, especially the hard and unusual ones, not just the situations where everything goes right.

ExampleRunning an assistant through 200 tricky past customer messages to see where it fails before trusting it with live ones.

Watch outGetting the easy cases right is a demo. Getting the hard cases right is a product.

Hallucination#

When the AI states something false with complete confidence, having effectively made it up.

ExampleCiting a study, a quote, or a statistic that does not actually exist.

Watch outConfidence is not the same as accuracy. This is exactly why real sources, fixed rules, and human review matter.

Latency (speed)#

How long the assistant takes to respond. Every extra tool, loop, and helper adds to it.

ExampleA single answer is fast; a panel of assistants debating over several rounds is deliberately slow.

Watch outMore autonomy almost always costs more time and money. Spend it only where it clearly pays off.