AI Foundations for Small Business · Lesson 1 of 11
What AI Actually Is (and Isn't)
Time: ~25 min reading + ~15 min worksheet Goal: By the end, you can explain AI, ML, LLMs, RAG, agents, and RPA in plain English — and you know which term means what when someone throws it around in a meeting.
---
"AI" is a marketing umbrella. It covers everything from the spam filter in your email (1990s technology) to ChatGPT to self-driving cars. When a vendor says "our product is powered by AI," they could mean literally anything.
So the first skill you need is terminology literacy. If you can match the vocabulary to what's actually happening under the hood, you'll never get snowed by a sales deck again.
Here's the map.
Think of it as nested circles, biggest to smallest:
- Artificial Intelligence (AI) — the whole field. Any system that does something we'd call "intelligent" if a human did it.
- Machine Learning (ML) — a *subset* of AI. Systems that learn patterns from data instead of being explicitly programmed. This is where 99% of modern "AI" lives.
- Deep Learning — a subset of ML that uses neural networks with many layers. Responsible for the huge leaps since ~2012.
- Generative AI (GenAI) — a subset of deep learning that *creates* new content (text, images, audio, video, code) rather than just classifying or predicting.
- Large Language Models (LLMs) — the specific kind of GenAI that generates text. ChatGPT, Claude, Gemini, Copilot are all powered by LLMs.
When someone says "we need AI," they almost always mean "we need GenAI, probably an LLM." That's the wave everyone's riding right now.
LLM (Large Language Model)
A neural network trained on enormous amounts of text that learned to predict the next word. That's it. That's the whole trick. It sounds dumb — and it is, mechanically — but at massive scale, predicting the next word well enough produces something that looks a lot like reasoning, writing, and conversation.
Famous LLMs: GPT-4/5 (OpenAI), Claude (Anthropic), Gemini (Google), Llama (Meta), Mistral.
What it's good at: Writing, summarizing, explaining, translating, drafting code, answering questions, pulling structure from messy text.
What it's bad at: Math without a calculator, knowing current events after its training cutoff, remembering things between conversations, being 100% right about anything.
Hallucination
When an LLM confidently makes something up. It's not lying — it literally doesn't know it doesn't know. This is the #1 thing to tell an SMB owner: AI will sometimes be wrong, and it will sound confident when it is. Your job is to design workflows where that's caught.
Context Window
How much text an LLM can "see" at once. Measured in tokens (roughly 0.75 words). Modern models have 100K–1M+ token windows, which means you can paste in entire documents. Older ones could only handle a few pages.
RAG (Retrieval-Augmented Generation)
The pattern that makes LLMs useful for your business. Instead of relying on what the model "knows," you:
- Store your company's documents in a searchable format
- When a question comes in, search for relevant chunks
- Stuff those chunks into the prompt along with the question
- Let the LLM answer using *your* data
This is how you build a chatbot that actually knows your policies, your product catalog, your SOPs. Almost every "AI-powered knowledge base" product you'll see is RAG under the hood.
Agent
An LLM that can *do* things, not just talk. It has access to tools (send email, read a database, call an API, browse the web) and can decide when to use them. Still early and flaky in 2026, but getting better fast.
Fine-tuning
Taking a pre-trained model and training it a bit more on your specific data so it behaves a certain way. Mostly overkill for SMBs in 2026 — RAG and prompting solve most of what people used to fine-tune for.
Machine Learning (classical)
The pre-ChatGPT flavor: predicting customer churn, classifying emails as spam, scoring loan applications, forecasting inventory. Still incredibly valuable, still widely used, still called "AI" by vendors. Doesn't generate anything — just predicts or classifies.
RPA (Robotic Process Automation)
Not AI at all, strictly speaking. Software robots that click through screens the way a human would. Think UiPath, Power Automate Desktop. Often sold alongside AI because the two pair up nicely — an LLM decides *what* to do, RPA does the clicking.
Copilot (lowercase or capital)
Confusingly overloaded. Microsoft uses "Copilot" as a brand for dozens of products. Generically, a "copilot" is an AI assistant embedded in another tool (GitHub Copilot for code, Microsoft 365 Copilot for Office, etc.). The common thread: it's AI that sits *inside* your existing workflow rather than as a separate app.
Prompt
The input you give to an LLM. A prompt can be a sentence or a 2,000-word document of instructions. Writing good prompts is a skill. Module 3.
Token
The chunk of text a model processes. Roughly 4 characters or 0.75 words in English. Pricing and context windows are both measured in tokens.
Model
A specific version of an AI. "GPT-4.5" is a model. "Claude Opus 4.7" is a model. Vendors release new models regularly, and they vary in speed, cost, and smarts.
Temperature
A setting that controls how "creative" or random an LLM's output is. Low temperature = consistent, predictable. High temperature = varied, creative, sometimes weird. Default 0.7ish. You'll rarely touch it, but now you know what it means.
- Sentient. It's math. Very big, very impressive math, but math.
- Reliable without supervision. It's a first draft generator, not a final answer machine. For anything that matters — legal, medical, financial, safety — a human needs to be in the loop.
- One thing. "AI" is not a product. Treat anyone who sells "an AI solution" the same way you'd treat someone selling "a software solution." What *kind* of software?
When you walk into an SMB, you're not looking for places to "add AI." You're looking for:
- Repetitive text-generation tasks → LLM
- Scattered knowledge that people can't find → RAG
- Repetitive clicking between systems → RPA + maybe LLM
- Prediction or classification of things (leads, churn, fraud) → classical ML
- Complex multi-step tasks with decisions → agent (cautiously)
If you know the vocabulary, you know which shelf to reach for.
---
- AI is the umbrella; most "AI" conversations in 2026 mean LLMs and GenAI
- LLMs predict the next word — very well, at massive scale. That's the whole magic trick.
- Hallucinations are real. Always design for human review on important outputs.
- RAG = LLM + your data. This is how most useful business AI gets built.
- Agents = LLMs that take actions. Promising, still rough.
- RPA, classical ML, and GenAI are different tools. Matching the tool to the problem is the core skill you're building.
---
Next up: Module 2 — The AI Tool Landscape. We'll turn these concepts into specific products you'll actually use and recommend.
---
Time: ~15 min Goal: Lock in the vocabulary so it sticks.
---
For each scenario, write which AI term best describes it. Choices: LLM, RAG, Agent, RPA, Classical ML, Fine-tuning, Hallucination, Context Window.
- A chatbot on a law firm's website that can answer questions using only that firm's case history and policies. → ______________
- ChatGPT confidently cites a court case that doesn't exist. → ______________
- A system that scores incoming sales leads 1–100 based on likelihood to close, trained on five years of CRM data. → ______________
- You paste a 40-page contract into Claude and ask it to summarize — it handles the whole thing in one shot. The thing that makes that possible is the ______________.
- A bot that logs into three different websites, downloads reports, and emails them to an accountant every Monday morning. → ______________
- An AI system that, given a customer complaint, decides whether to refund, escalate to a manager, or draft a reply — and then *does* it. → ______________
- A company trains a smaller model on 10,000 examples of their customer service tone so it writes replies that sound exactly like their brand. → ______________
- The thing that's actually predicting the next word in a sentence is the ______________.
A vendor sends your (imaginary) client this email:
> "Our proprietary AI-powered solution leverages cutting-edge neural technology to deliver intelligent insights across your enterprise data fabric, enabling next-generation decision intelligence."
Rewrite this in plain English, using the vocabulary from this module. What are they *probably* actually selling? (Hint: it's likely one of two or three things. Take a guess.)
Your translation: _____________________________________________ _____________________________________________ _____________________________________________
You're walking through a 30-person insurance brokerage. The owner shares three headaches. For each, name the AI category you'd reach for.
Headache 1: "My CSRs spend hours writing summary emails to clients after every policy change. It's the same structure, just different details each time." → Category: ______________ → Why: ______________
Headache 2: "I have 20 years of underwriting notes in PDFs, emails, and a shared drive. When we get a new prospect, my people waste an hour hunting for similar past cases." → Category: ______________ → Why: ______________
Headache 3: "Half my quotes never close. I wish I knew which prospects were worth chasing before my team spends two weeks on them." → Category: ______________ → Why: ______________
---
Exercise 1:
- RAG
- Hallucination
- Classical ML (specifically classification/scoring)
- Context Window
- RPA
- Agent
- Fine-tuning
- LLM
Exercise 2 (sample answer): "We have an LLM chatbot that runs on your documents." Or possibly: "We bolted a RAG pipeline onto your existing data." Or: "We have a classical ML tool that scores your business data." The vague language usually means they don't want to commit to a specific category because the specifics are less impressive than the marketing.
Exercise 3:
- LLM — repetitive text generation with templates. Perfect LLM use case.
- RAG — retrieval over their own unstructured documents. Textbook.
- Classical ML — prediction/classification on historical data. Lead scoring is a classic ML problem. (You could also pitch an LLM analysis of prospect profiles, but for a real statistical prediction, classical ML wins.)
---
How'd you do? If you got 7+ right on Exercise 1 and nailed the shelf-picking on Exercise 3, you're ready for Module 2. If not, re-read the lesson's "Terms You'll Hear Most" section — the vocabulary has to be automatic before we move on.