Before you can decide whether AI belongs in your business, you need a working mental model of what it actually is. The good news: the core ideas are simpler than the headlines suggest. No maths, no jargon — just the handful of concepts every decision-maker needs.
AI, machine learning, deep learning: nested ideas
These three words get used interchangeably, but they are not the same thing. Picture them as circles inside circles.
- Artificial intelligence (AI) is the broad goal: getting computers to do things that normally require human intelligence — understanding language, recognising images, making recommendations.
- Machine learning (ML) is the most successful way of reaching that goal today. Instead of a programmer writing out every rule by hand, the system learns patterns from examples.
- Deep learning is a powerful style of machine learning that uses large "neural networks" — loosely inspired by the brain — and is what powers most of the AI you hear about, including the chatbots.
So every LLM is deep learning, all deep learning is machine learning, and all of it sits under the umbrella of AI.
Rules you write vs. patterns the machine learns
The clearest way to understand machine learning is to contrast it with the old way of building software.
In traditional software, a human writes explicit instructions: if the invoice total is over €1,000, flag it for review. Every case has to be anticipated and coded by hand. That works well for tidy, predictable problems.
But how would you hand-write rules to tell a cat from a dog in a photo, or to spot which support emails are angry? You can't enumerate every possibility. Machine learning flips the approach: you show the system thousands of labelled examples, and it figures out the patterns on its own. You don't tell it the rules — it infers them from the data.
Traditional software follows rules a human wrote. Machine learning discovers the rules itself by studying examples. That single shift is what makes modern AI possible.
This also explains a quirk: an ML system is only as good as the data it learned from. Biased or thin data produces a biased or unreliable model — a theme we return to in Responsible AI & humans in the loop.
Large language models: very good next-word guessers
A large language model (LLM) — the technology behind tools like the ones in our live demo — is a deep-learning system trained on an enormous amount of text. At its heart, it does something surprisingly humble: it predicts the next piece of text (a "token" — roughly a word fragment), over and over, one step at a time.
Type "The capital of France is…" and the model predicts "Paris" because that pattern appeared countless times in its training text. String enough of these predictions together and you get fluent paragraphs, summaries, and answers.
This matters for two practical reasons:
- An LLM generates language; it does not look anything up in a database. That is why it can write a polished email but can also state something false with total confidence — a behaviour known as hallucination.
- Because it works from patterns, the quality and recency of those patterns is everything — which brings us to what a model actually "knows."
Training vs. inference: two very different moments
People often picture an AI as "learning" while you chat with it. It isn't. There are two distinct phases:
- Training is the one-time, expensive process where the model studies its data and adjusts itself until it predicts well. Think of it as the model's education — months of study, then a final exam.
- Inference is what happens every time you use it: you give it an input, and the already-trained model produces an output. The model is not learning from your question; it's applying what it already learned.
A useful everyday analogy: training is medical school; inference is the doctor seeing a patient. The doctor doesn't go back to school for each appointment — they apply fixed knowledge to a new case. Your private conversations don't quietly become part of the model, which is reassuring for governance, but it also means the model can't update itself with your latest facts on its own.
Supervised, unsupervised, and generative learning
Machine learning comes in a few flavours, easy to tell apart by what the examples look like.
- Supervised learning uses labelled examples — each one tagged with the right answer. Show it 10,000 emails marked "spam" or "not spam," and it learns to sort new ones. Predicting house prices from past sales works the same way.
- Unsupervised learning gets data with no labels and finds structure on its own — for example, grouping your customers into natural segments nobody defined in advance.
- Generative AI (the LLM category) learns the patterns of language or images so well that it can produce new content that fits those patterns. We go deeper on this in Generative AI & foundation models.
Why the model doesn't know your business
Here is the single most important takeaway for any SME. A model only knows what was in its training data, up to a fixed cut-off date. It has never seen your contracts, your price list, your internal handbook, or anything published after that cut-off.
An off-the-shelf LLM is brilliant at general language but blind to your private documents — they simply weren't in its training data.
Ask a raw model about your refund policy and it will either admit it doesn't know, or worse, invent a plausible-sounding answer. This isn't a flaw to fix with a "smarter" model — it's a structural fact. The solution is not to retrain a giant model on your files (slow, costly, and a privacy minefield); it's to give the model your documents at the moment it answers, under your control. That technique — retrieval-augmented generation — is exactly what Module 3 covers, and it's the foundation of how a sovereign, governed agent answers on your own knowledge base.
With this mental model in place — patterns over rules, training apart from inference, and a model that's powerful but blind to your data — you're ready to look closer at the generative models themselves and where they come from.