Beyond GenAI: AI's big picture

Generative AI is having its moment, and honestly, it has earned it. LLMs like Claude, Gemini and ChatGPT are genuinely incredible, advancing at an unprecedented speed, and it’s accessibility to millions of hands has already propelled real, impactful innovation. While that undoubtedly deserves celebrating, the hype tends to drown out that the chatbot in your browser is the newest, smallest branch of a bigger tree (Machine learning or ML). ML is GenAI’s older and quieter parent — and it already reads medical scans, prices markets in real time, routes global logistics, and predicts which sustainable materials best resist humidity long before anyone runs a lab test. GenAI is a spectacular branch. This blog is about the whole tree, in plain words, and the 3 ways it learns.

Reaching for the newest tool before you understand the problem is exactly how most AI pilots quietly die. Understanding your problem deeply to match with the specific toolbox solution is how you avoid it.

First, zoom out: AI is bigger than the chatbot

Ask most people to picture AI and they imagine a chatbot: type a question, get a paragraph back. That box is real and useful, but it is one small, recent corner of a much larger field. The Generative AI and large language models (LLMs) behind those chatbots are the youngest subset of machine learning — itself just one part of artificial intelligence. Before we zoom in, here is the whole family on one map, drawn roughly to the footprint each part occupies in real-world use:

Now zoom in: what machine learning really is

Traditional software is a recipe: you write the rules, the computer follows them. Machine learning flips that. You show it a mountain of examples and it works out the rules itself — often ones far too subtle for a person to find out. You don’t tell it “an invoice over €10k from a new supplier is risky”; you show it hundreds of thousands of past invoices and it learns a pattern richer than any checklist you could author by hand.

That is the part worth taking seriously: at scale, a well-fed model doesn’t just automate human rules, it can find signal people never spotted. This umbrella splits into three big styles, defined by what kind of feedback the machine gets:

1. Supervised learning: learning from labelled examples

The most common style, and often the most valuable. You feed the machine examples that already carry the answer — invoices tagged “paid late / on time,” photos tagged “defect / fine” — and it learns to predict that label on cases it has never seen, with a consistency no team could ever meet. It’s what powers inventory optimisation: forecast demand well and you never miss a sale, without tying up cash or letting shelf life run out. The one hard rule is clean labels — it’s only ever as good as the examples you feed it. Rich enough to earn its own blog later in this series.

2. Unsupervised learning: finding structure no one labelled

Here you hand over data with no answers attached and ask the machine to find the shape of it on its own — natural groups, odd outliers, hidden patterns. The best-known example is clustering — grouping customers, packaging types or material streams that behave alike, at a resolution no spreadsheet would reach.

3. Reinforcement learning: trial, error and reward

The machine tries something, gets a reward or a penalty, and adjusts — the way you’d train a dog, or learn a video game by dying a lot. It shines wherever a system makes a sequence of decisions, each shaping the next: it beat the world champion at Go and sits under dynamic pricing, where every price you set influences demand, which changes the best price to set next. Some of the most powerful and delicate work in the field.

Geeky aside: this “explore something new or exploit what already works, and in what order?” question is old and deep — Brian Christian and Tom Griffiths give it whole chapters in Algorithms to Live By, one of our studio’s favourite books.

We lead with the problem, not the technology

This is the part most miss. Understanding AI's big picture matters, but choosing the technology before you understand the problem is how most AI pilots quietly die. So we never start with the model. We start with a problem worth obsessing over, and let the right technique earn its place inside a disciplined build. One storyline, 5 stages:

Gen AI is a real breakthrough. But “let’s do AI, roll out Copilot to the whole company” can still be the wrong first move. Starting with the tool instead of a use case is exactly how a pricey pilot quietly dies 6 months later — licences renewing, siloed productivity gains but nobody quite sure what it was for. So don’t lose the bigger AI picture, or the agile approach: fall in love with the problem first, then let the right technology — GenAI or one of its quieter cousins — turn it into business value. That is precisely where an AI Value Scan™ comes in: it maps and scores where AI would actually pay off in your operation, on profit and planet, before you spend a cent licensing or building.