Almost every conversation about AI right now runs on a handful of words: model, agent, tool, workflow, hallucination, grounding, agentic. Most of them get used as if they were interchangeable, but they are not. Each word names something specific, with its own guarantees and its own failure modes, and blurring them is why so many AI conversations go in circles and so many AI budgets go sideways.
The trouble is one word doing too many jobs. "AI" has swollen into a warm, vague blob, and inside that blob sit at least fifteen distinct ideas. When you cannot name what you are buying, you cannot tell whether you bought the right thing. When you cannot name how a system behaves, you cannot tell whether to trust it. The fuzzy vocabulary has a price, and it shows up as money.
This is the reference I point people to when the words start sliding around. It is built as a glossary, not an essay, so you can jump to the term you need and leave. The vocabulary sorts cleanly into three tiers.
Tier 1, the four things AI can be. The object-types. When someone says "we're implementing AI," they mean one of four things: a model, a tool, a workflow, or an agent. These are what you buy and what you scope. They differ by a single question: who decides what happens?
Tier 2, what an agent is made of. The components. An agent is the most involved of the four, and it is assembled from different parts: a system prompt, context, memory, tools, skills, hooks. This is the vocabulary for asking what is actually inside the thing a vendor is selling you.
Tier 3, the words for how it behaves. The qualities and behaviours. Reasoning, agentic, hallucination, grounding, tokens. These describe and judge what a system is doing, and they are how you tell a demo's marketing from its reality.
The tiers split this way because they answer different questions. Tier 1 asks what kind of thing this is. Tier 2 asks what it is built from. Tier 3 asks how it behaves and whether you can trust it. Get the tier right and the individual words fall into place.
Tier 1: The four things AI can be
When someone tells you they are implementing AI, they mean one of four object-types. This is the tier where buying and scoping decisions live, so it is the one worth getting exactly right. The four sit on a spectrum of a single question: who decides what happens next. A model predicts. A tool executes a fixed rule. A workflow runs a path you drew. An agent chooses its own path toward a goal you set.
| Term | What it means | Example | Common misuse |
| Model | A system that predicts the most likely continuation of an input, learned from patterns in vast amounts of data. It guesses. It does not look things up or run rules. Geertrui Mieke De Ketelaere frames it precisely: these systems are probabilistic, not deterministic. Ask any chatbot for a random number between 0 and 10 and it reliably returns 7, because 7 is the statistically most likely answer, not a random one. | Draft a paragraph, summarise a report, suggest ten subject lines for an email. | Reading a model's answer as a retrieved fact. It predicted the answer with fluency; it did not look it up. |
| Tool | A deterministic function. Same input, same output, every time. No guessing, no bias. You can trust it the way you trust arithmetic, because it is arithmetic. | A calculator, a currency conversion, a database query that returns an order status. | Calling any software feature a "tool" loosely. In AI terms, a tool is specifically the exact, repeatable part. |
| Workflow | A sequence of steps defined in advance. It runs the path as designed and does not adapt past it. Outside that path it breaks, or does the wrong thing with total confidence. | When an invoice arrives: extract these fields, check them against that list, route the result to that inbox. | Calling a workflow an "agent" because a model runs one of its steps. A scripted path with a model wired in is still a workflow. |
| Agent | A system you give a goal and a set of tools, which then works out the steps itself: which tool to use, in what order, when the job is done. You stop scripting the path and start setting the objective. | Hand it "reconcile this month's expenses" and let it decide which lookups to run and when to stop. | Calling a chatbot with fixed replies an "agent." If it cannot choose its own steps, it is not deciding anything. |
How the four differ
Two contrasts do most of the work here. The first is trust: a model guesses, a tool is exact. They are opposites, and the useful systems combine them, letting the model decide what to do while tools do the parts that have to be right. The second is control: a workflow follows the path you drew, an agent draws its own. That line, between following a path and choosing one, is the real border between automation and what people now call agentic.
Model and agent get confused most often, because agents are built on models. Hold them apart like this: the model is the predictor, the agent is the decider wrapped around it, usually with tools in hand. One guesses the next word, while the other pursues a goal.
Do
- Name which of the four you are buying before you compare a single price.
- Pair a model with tools whenever a number, date, or fact has to be exact.
- Use a workflow when the path is known and stable. It is cheaper, auditable, and predictable.
- Reserve the word "agent" for systems that genuinely choose their own steps.
Don't
- Don't pay agent prices for a workflow with a chatbot on the front.
- Don't ask a bare model for figures you would otherwise pull from a database.
- Don't scope a rigid workflow for a problem that is mostly exceptions.
Tier 2: What an agent is made of
Once you have decided you actually need an agent, the vocabulary shifts from what-it-is to what-it-is-built-from. These are the parts. Builders and vendors throw the words around freely, and knowing them is how you ask what is really inside a system rather than nodding along. One word from Tier 1 reappears here on purpose: tool. Hold that thought, because it is the sharpest source of confusion in this tier.
| Term | What it means | Example | Common misuse |
| System prompt | The standing instructions that define an agent's role, boundaries, and default behaviour. It is read at the start of every run and shapes everything the agent does. | "You are a support assistant. Answer only from the approved knowledge base. Escalate refunds to a human." | Confusing it with a user's typed message. The system prompt is the fixed frame; the user's question is the variable input. |
| Context (context window) | The span of text the model can see at once: the system prompt, the recent conversation, and any documents fed in. It is finite. When it fills up, the earliest content drops out of view. | Pasting a three-page brief plus the last ten messages so the agent can answer with all of it in view. | Assuming the agent "knows" everything ever said to it. Outside the current window it sees nothing, unless something puts the text back. |
| Memory | A store outside the model that deliberately feeds past information back into context on later runs. The model has no built-in recall between sessions. Memory is the thing built around it to create the effect of remembering. | Saving a customer's preferences so the agent greets them correctly next week. | Believing the model remembers on its own. Without an explicit memory layer, every session starts from a blank slate. |
| Tools | The deterministic functions an agent can call to do the exact parts of a job: look something up, calculate, send, fetch. Same input, same output. | An agent calling a "check inventory" function instead of guessing stock levels in its head. | Assuming "tool" only means the standalone Tier 1 option. Here it is a capability the agent reaches for while it works. |
| Skills | Packaged procedures an agent can load for a specific kind of task: a bundle of steps, instructions, and know-how for one job. | A "format an invoice" skill the agent pulls in only when invoicing actually comes up. | Treating skills as new intelligence. A skill is structured instructions, not extra brainpower. |
| Hooks | Trigger points that run something automatically at a defined moment: before a step, after a result, on a specific event. They fire on rules you set, not on the agent's judgment. | A hook that logs every action the agent takes, or checks its output before anything is sent. | Confusing hooks with the agent's own decisions. The agent chooses; the hook fires on your rule. |
How the parts fit, and the two faces of "tool"
The word tool wears two hats, and swapping them silently is where people get lost. In Tier 1, a tool is an alternative to an agent: the exact, simple option you pick instead of building something that decides. In Tier 2, a tool is a part an agent uses: one of the deterministic functions it calls while it works out a goal. Same word, two positions. What never changes is the tool's nature. It is always deterministic, same input same output. What changes is only whether it stands alone or sits inside an agent's toolbox.
A few more pairs are worth separating. Context is what the agent can see right now, temporary and finite; memory is what gets carried deliberately across sessions, persistent and engineered. The system prompt is the one part of the context that never varies, while the rest of the window changes every turn. And a skill is not a tool: a tool executes one exact operation, whereas a skill is a bundle of instructions for a task and may well tell the agent which tools to reach for.
Do
- Ask a vendor what sits in the system prompt, and what the agent is and is not allowed to do.
- Treat memory as something engineered, and check what is stored, where, and for how long.
Don't
- Don't assume an agent remembers across sessions unless there is a memory layer doing it.
- Don't mistake a large context window for real memory. The window empties; memory persists.
- Don't read "tool" as a single fixed meaning. Check whether it stands alone or sits inside the agent.
- Don't expect skills to add intelligence. They add structure and repeatability, which is often exactly what you want.
Tier 3: The words for how it behaves
The third tier is adjectives and verbs: the words people use to describe or judge what an AI system is doing. They surface in demos, in sales decks, and in the fine print. Getting them straight is how you tell what a system genuinely does from how confidently it is being pitched.
| Term | What it means | Example | Common misuse |
| Reasoning | When a model produces the intermediate steps of an argument before its answer, which usually sharpens results on complex tasks. The steps are still predicted text, not a proof the model checked. | A model showing its working on a multi-step word problem before it states the total. | Reading "reasoning" as evidence the model thought like a person. It generated plausible steps; it did not verify them. |
| Agentic | A description of behaviour where a system chooses its own steps toward a goal rather than following a fixed path. It is the adjective form of what makes an agent an agent. | A system that decides on its own to fetch a file, read it, then email a summary. | Stamping "agentic" on any automation. If the path is fixed, it is not agentic, however advanced the label sounds. |
| Hallucination | When a model produces a confident, fluent answer that is simply wrong, because it predicted a likely-sounding continuation rather than a true one. | A model inventing a plausible but non-existent legal citation or product specification. | Treating it as a rare bug. Hallucination is a direct consequence of prediction, and confidence is no guide to correctness. |
| Grounding | Tying a model's answer to a verifiable source, a document, a database, a tool result, so it responds from real information rather than pure prediction. | An assistant that quotes and links the specific policy page its answer came from. | Assuming a confident answer is grounded. If nothing connects it to a source, it is still a guess. |
| Tokens | The chunks of text a model reads and writes, roughly word-pieces. They are the unit of both the context window's size and what you pay to run the model. | A short paragraph runs maybe 60 to 80 tokens; a long document, thousands. | Thinking in words when the system counts tokens. Both the cost and the context limit are measured in tokens, not words. |
How the behaviours connect
Reasoning and hallucination are two faces of the same predictor. Reasoning is prediction producing useful steps; hallucination is prediction producing confident nonsense. The main defence against the second is grounding, which re-anchors a guesser to a real source before it answers. Agentic reaches back to Tier 1: it is the exact quality that separates an agent from a workflow. And tokens sit underneath everything, the physical unit that makes the context window finite and the monthly bill add up.
Where Tier 1 asked what the thing is and Tier 2 asked what it is built from, Tier 3 asks the questions that decide trust and cost: can I rely on what it just told me, and what did that answer cost to produce.
Do
- Ask whether an answer is grounded, and in what source, before you act on it.
- Judge "reasoning" by whether the output is checkable, not by how confident it reads.
- Think in tokens when you estimate cost or worry whether a long document will fit.
Don't
- Don't treat a fluent answer as a correct one. Hallucination wears confidence well.
- Don't accept "agentic" as a feature. Ask what the system actually gets to decide.
Adjacent terms
Three more words show up the moment you go one step deeper. They are not part of the three tiers, but they earn a place in any working glossary because you will meet them fast.
| Term | Short meaning |
| RAG (retrieval-augmented generation) | A common way to achieve grounding: retrieve the relevant documents first, then let the model answer from them. Grounding is the goal; RAG is one method of getting there. |
| MCP (Model Context Protocol) | An open standard for connecting agents to tools and data sources in a consistent way. Think of it as the plumbing that lets tools plug in without custom wiring each time. |
| Fine-tuning | Further training a base model on your own examples to shift its default behaviour. It changes the model itself, unlike grounding, which feeds the model information at the moment it answers. |
The payoff: three lenses, two expensive mistakes
Three tiers, three lenses. Tier 1 asks what am I buying. Tier 2 asks what is inside it. Tier 3 asks can I trust it, and what does it cost. Point each lens at any AI decision and the fog lifts.
Almost all of the money lost to vague AI vocabulary lands in two mistakes, and both trace back to one word doing too many jobs. The first is the mis-buy, and it lives in Tier 1. A vendor sells you an "AI agent" and ships a workflow with a chatbot on the front. It follows a fixed script, it cannot handle the one case you actually care about, and you paid agent prices for automation you could have bought cheaper and understood better. The reverse happens just as often: you scope a rigid workflow for a problem that genuinely needs judgment, then spend a year discovering the real world has more branches than your flowchart allowed.
The second is the mis-trust, and it stretches across Tier 1 and Tier 3. You treat a model like a tool. You ask it for a figure, a clause, a citation, and you take the answer as though it came from a database, because it answered with the same fluency a database would. But the model predicted that answer. It did not look it up. Trusting a probabilistic guesser as if it were deterministic arithmetic is the single most common and most expensive AI mistake I see in the field, and De Ketelaere's boardroom "7" is about the cheapest inoculation against it you will ever find.
Fix the words and both mistakes get much easier to dodge, because you can finally ask the question sitting underneath every AI decision. Is this the part that predicts, the part that is exact, the path I drew, or the thing that decides? And once it is deciding, what is inside it, and can I trust what it does.
This glossary is the ground floor of a longer climb. It opens a series about working with agents, from one person to a whole company, and the vocabulary that keeps you sharp at your own laptop is the same vocabulary a company needs before it reshapes itself around what agents can do. You cannot have the boardroom conversation until you have had the desk-level one properly.
So start with the smallest version of the question, and ask it out loud at your next meeting. When your team says "AI," which of the four does it actually mean? And once you know, the board-level version is waiting right behind it: who in this company is allowed to trust it, with what, and who is accountable when it guesses wrong?