Lesson 1: What AI Actually Is

Every department is getting the same pitch right now: "We're rolling out AI, here's a tool for your team." Then the training shows a chatbot demo, someone says it's "just autocomplete," and half the room leaves convinced it's either magic or a gimmick. Neither is right, and the reason people land on both is that the words around AI are doing more confusing than clarifying.

This first lesson clears that up. By the end you'll be able to look at any AI product and answer the two questions that actually matter: is this a model doing the work, or a set of fixed rules? And am I talking to a chatbot, or am I using a tool that does something? The rest of this series builds on those two answers, so it's worth getting them straight now.

What AI actually is

Strip away the hype and "AI" mostly refers to a model: a software program trained on a large set of examples to do one kind of task. Stanford's teaching guide on AI puts it plainly: a model is "an AI software program that has been trained on datasets to perform a specific task." ¹

For the tools most of us use at work, that task is language. A large language model (LLM) is a model trained on enormous amounts of text. IBM describes how it works: the model is "a giant statistical prediction machine" that repeatedly predicts the next word in a sequence, one token at a time, based on patterns it learned in training. ²

That last part is the sentence to hold onto. An LLM does not look up an answer. It guesses, statistically, what text is most likely to come next, given everything you just typed. When you ask it to summarize a report, it is not reading the report the way you do. It is producing the most probable sequence of words that a summary would look like. That is a powerful trick, and it is also the reason the tools get things wrong: a confident guess is still a guess.

Model vs automation: the fork in the road

Here is the first distinction that decides what any AI product can and cannot do. Automation follows rules that a person wrote in advance. A model produces answers for situations nobody pre-coded.

Retool's breakdown of the two is the clearest version I know: "automation is best when the steps are known in advance, and AI is better when the work is less predictable or needs human-like judgment." ³

Concretely:

• Automation: "If a customer marks their order as urgent, send an email to the fulfillment team." A person wrote that rule. It never surprises you, and it never adapts.
• Model: "Read this customer's email and tell me whether they sound angry and whether they might cancel." No one can write that rule in advance, because the email is different every time. The model guesses at the sentiment and the risk.

Most "AI" products in an office are actually a mix. A tool that sorts your inbox into "bills, receipts, newsletters" may combine a model (it reads the email) with automation (it files it into a folder once it decides). The useful habit is noticing which part is doing what. When a tool fails in a way that feels dumb and mechanical, it is usually the rules failing. When it fails in a way that feels confidently wrong, that is the model.

Chatbot vs tool: wrapper and worker

The second distinction is about how you interact. A chatbot is a conversation: you type, it replies. Stanford defines a chatbot as "a computer program that uses an LLM to simulate a conversation with human users, typically through typed text." ¹

A tool is a button. It takes your input and does something with it: turns a messy transcript into clean notes, drafts the first version of a contract clause, translates a deck. ²

The trap is thinking they are the same. A chatbot will happily chat all day, but it does nothing until you copy its answer somewhere. A tool has a job: you give it a document and it produces a result in the format you need. Both are useful. They are useful in different ways. The chatbot is for exploring and drafting; the tool is for getting a specific piece of work done. When someone says "our team has an AI chatbot," they are usually telling you about the wrapper, not the work. Ask what it actually does.

The word pile-up: LLM, generative AI, prompt, token, agent

Half the confusion is that five terms get used as if they mean the same thing. They don't.

• AI is the umbrella. Anything that simulates some human ability, from a spam filter to a self-driving car.
• Machine learning is the subset of AI where the program learns from examples instead of following hand-written rules.
• Generative AI is the subset that creates new content, text or images, rather than just classifying. Georgetown's explainer notes that generative AI and LLMs overlap heavily but are not identical: generative AI is the broader idea of producing content, while an LLM is a specific language-focused system. ⁴
• LLM is a model trained on text to predict and produce language. ²
• Prompt is simply what you type into the model. The instruction, the question, the draft. There is nothing technical about it.
• Token is the unit the model actually reads and produces, usually a chunk of a word. IBM describes the pipeline: the model breaks your prompt into tokens, then predicts the most likely next token over and over. ²
• Agent is the current buzzword, and the word you should be most suspicious of. For this series, keep a working definition: an agent is a model that can take multiple steps and use tools on its own, deciding what to do next. Many products labeled "agents" are really just chatbots in a trench coat. You will learn to tell the difference in later lessons.

You do not need to memorize that list. You need the one insight underneath it: an LLM predicts the next token. Everything else, from "generative AI" to "agent," is a layer of description and features built on top of that.

Worked example: two tools, one job

Say your team needs a short first-draft reply to a customer who is frustrated that a delivery ran late. Watch how two different products handle it.

The rules-based tool. The company built a response generator from a decision tree. It looks for keywords: "late," "delivery," "refund." If it finds "late," it returns a fixed apology template with the standard "we're sorry, here's a 10% credit" block. It is fast, consistent, and identical every time. It cannot respond to a customer who is upset for a reason nobody pre-coded, and it has no idea the apology reads as scripted.

The LLM-based tool. The company points an LLM at the same customer message with the instruction "write a reply in the brand voice, acknowledge the specific delay, and offer the standard credit." The model reads the whole message, not just keywords, produces a reply that references the actual date and reason, and varies its wording. It also might invent a reason the customer never mentioned, because it is guessing what a good reply looks like. ²

Now put the two distinctions to work. The rules tool is automation with no model; the LLM tool is a model doing the writing. The chatbot-versus-tool question is separate: whichever product you use, if it only produces text in a chat window, it is a chatbot that needs you to copy the answer into your CRM. A real tool takes the customer message and writes the reply back into the ticket itself. ²

That is the whole skill this series teaches: seeing which part is rules, which part is a model, and which part is a person pasting results around.

Try it

Open any AI assistant your company already provides (ChatGPT, Claude, Gemini, or an internal tool all work the same way here) and run this prompt on a task you actually do:

Here is a piece of work I do regularly: [paste the task, for example "writing a short status update for my manager each Friday"]. Describe the steps you would take to do it. Then tell me which of those steps need judgment and which could be done by a fixed set of rules, and why.

Then look at a tool you already use and ask the two questions from this lesson out loud: Is a model doing the work, or is it a fixed rule? And is it a chatbot that talks, or a tool that does? Most office tools are a mix, and naming which part is which is the skill. Watch for the model part to be the part that sometimes answers confidently and wrongly; that is not a bug, it is what a statistical guess looks like. ²

What's next

Lesson 2 is about how to talk to the model, which is the single highest-leverage skill in this whole series. You will learn why the same tool gives one person a useful draft and another a pile of fluff, and a prompt structure that works no matter which assistant you use. For now, keep the mental model from today: an LLM is a next-word prediction machine, automation is pre-written rules, and the two get mixed together inside most tools you touch. Once you can spot which part is which, the rest of the series is detail.

Note: model names and product capabilities change quickly. The definitions in this lesson are stable; verify the specific features of any tool you adopt against the vendor's current documentation.