"Can't we just use AI for this?"
It is the question every leader hears now, usually in a meeting, usually with a budget attached. Sometimes the answer is a resounding yes - machine learning is genuinely the right tool. Just as often the honest answer is "a simple rule would work better, faster, and cheaper," and knowing which is which is worth more than any model. Lumen Skincare, the direct-to-consumer brand we follow all course long, wants to predict which website visitors will buy. That is a real ML problem. But half the things pitched as "AI" around it are not. This session gives you the judgement to tell them apart.
Ordinary software is told the rules. ML learns them. 8 min live
Here is the whole distinction in one picture. In classic software, a person works out the rule and writes it down as code. In machine learning, you hand the computer a pile of past examples - inputs paired with the right answers - and it works out the rule for itself. That flip is the entire idea; everything else in this course is detail.
LiveA definition you can actually use in a meeting3 min▶
Machine learning is a system that learns its rules from examples, so it can make predictions on new cases it has never seen. No one writes "if the visitor came from email and spent money last month, they are likely to buy." Instead you show the system thousands of past visitors and whether each one bought, and it discovers that pattern - and many subtler ones - on its own.
- Examples - the past data, each with its known outcome (a visitor, and whether they converted).
- Training - the once-off process of finding the pattern in those examples.
- The model - the learned rule, saved and reused to score new visitors in real time.
- Prediction - the model's best guess for a fresh case, always with some uncertainty.
Lumen's team could hand-write rules for "likely buyer" - and they did, for years: "returning customer + added to cart = email them." It caught the obvious cases and missed the rest. A model trained on two years of visits found combinations no one had thought to write down (a particular channel, on desktop, late in the month, with a mid-range prior spend) and roughly doubled the hit rate. Same data the humans had; the machine just read more of it.
LiveWhy now - what changed to make this work3 min▶
The maths behind ML is decades old. Three things made it practical and unavoidable in the last few years:
- Data - businesses now record everything, so there are millions of labelled examples to learn from.
- Compute - cheap, fast hardware makes training on that data affordable.
- Tools - libraries like scikit-learn (the one the builder track uses) turned a research project into a few lines of code any analyst can run.
Three kinds of learning - and when to use none of them 7 min live
Almost every ML project you will meet is one of three shapes. Naming the shape tells you what data you need, what the model produces, and how you will judge it. Then the most valuable slide of all: the cases where machine learning is the wrong answer.
LiveSupervised learning: classification vs regression2 min▶
Inside supervised learning - the box that matters - there are two flavours, split by what you are predicting:
- Classification - predict a category. "Will this visitor convert: yes or no?" The answer is a label. Lumen's conversion model is a classifier.
- Regression - predict a number. "How large will this order be, in dollars?" The answer is a quantity. Lumen's order-value model is a regression.
You will meet both all course long, because they need different metrics to judge (Session 4). Keep the pair in your head: category = classification, number = regression.
Lumen runs both on the same visitor. The classifier decides who to show a free-shipping offer to (the likely-converters). The regression decides how much of a discount is worth it (the big-order visitors). One question is a yes/no, the other is a how-much - two models, two shapes.
LiveWhen machine learning is the wrong tool2 min▶
The most senior instinct in the room is knowing when not to reach for ML. Skip it when:
- A simple rule already works. "Flag orders over $500 for review" needs no model. If a human can write the rule in one sentence and it holds, write the rule.
- You have too few examples. ML learns from volume. A few dozen past cases will not teach it a reliable pattern - it will just memorise them.
- The stakes demand an explanation you cannot give. If a wrong call is dangerous or must be justified by law, a black-box prediction may be unusable however accurate.
- The world just changed. A model learns from the past. If the ground shifted (a new product, a pandemic, a new market), the past may no longer predict the future.
- You actually need a cause, not a prediction. "Will sales fall?" is ML. "Will this price cut cause sales to rise?" is an experiment - the next course, not this one.
This week ◐ 25 min total
- Find one "AI" claim in your org or inbox. Decide, in one sentence, whether it is really a prediction problem - or whether a rule, a report, or an experiment would do better.
- Name the shape. Pick a prediction your team would love to have. Is it classification (a category) or regression (a number)? Write it down; you will defend it in Session 2.
- Spot the data. For that same prediction, list what past examples you would need - and whether you actually have them, with honest labels. This is usually the real bottleneck.
- Optional: find a case where your team used a model but a simple rule would have been better, or vice versa. Bring the story.
Three questions before you go 🎯 ◐ 90 seconds
1 · In one line, what makes machine learning different from ordinary software?
The defining flip: classic software is given rules and produces answers; ML is given answers (labelled examples) and produces the rules. More compute helps, but it is not the definition - and ML is not automatically more accurate.
2 · Lumen wants to predict how many dollars an order will be. What shape of problem is that?
Predicting a quantity (dollars) is regression. Predicting a category (convert: yes/no) is classification. There are labels - the past order values - so it is supervised, not unsupervised.
3 · A manager asks "will cutting the price CAUSE sales to rise?" Why is that not really a machine-learning question?
ML finds patterns and predicts; it cannot establish that one thing causes another. A cause-and-effect question needs a controlled experiment (an A/B test) - that is the experimentation course, not this one.
What this session covers
This leader session teaches the framing behind the "what is ML" week of the major intro courses - the ~80% a decision-maker needs. The hands-on Python (setting up scikit-learn, building the dataset) lives in Builder Session 1; the sources below go deeper.