Generative AI vs predictive AI

When most people say “AI” today, they mean the tools that write, draw and chat — the ones that produce something new every time you ask. But a huge amount of the AI that quietly runs businesses doesn’t create anything at all. It looks at data and tells you what is likely to happen next: which customer might leave, how much stock you’ll need in December, whether a transaction looks fraudulent.

Those are two different families of AI — generative and predictive — and mixing them up is one of the most common reasons I see projects head in the wrong direction. They are built differently, judged differently and priced differently. In this post I want to give you a simple way to tell them apart, and a feel for when each one is the right tool.

The one-sentence difference

Here’s the version I use on whiteboards:

Predictive AI is the older, more established of the two. It’s the classic machine learning we covered earlier in this series: you show a model lots of labelled examples, it learns the pattern, and then it makes predictions about new cases. Google’s Machine Learning Crash Course frames most of these problems as either regression (predicting a number, like next month’s sales) or classification (predicting a category, like “spam” or “not spam”)^[4].

Generative AI learns the patterns in its training data well enough to produce new examples that look like they belong. Two research breakthroughs helped make this practical for images. In 2014, Generative Adversarial Networks (GANs) set two models against each other — one generating fakes, the other trying to spot them — until the generator got good enough to fool its opponent^[1]. In 2020, diffusion models showed a different approach: learn to gradually remove noise from an image, then start from pure noise and “denoise” your way to a brand-new picture^[2]. For text, the large language models we covered earlier in this series do the same job one token at a time.

Side by side: everyday business examples

The easiest way to build intuition is to look at real use cases. Notice how the outputs on the left are decisions or numbers, while the outputs on the right are content.

You’ll notice something in every row: the two are complementary. The predictive model decides who or what needs attention; the generative model helps you act on it. Many of the most useful systems I help build combine both.

They are judged differently

This is the part that trips teams up. Predictive AI has a reasonably clear definition of “right”. Either the customer churned or they didn’t. Either the forecast was close to actual sales or it wasn’t. That means you can measure accuracy on historical data before you ever go live, and keep measuring it afterwards.

Generative AI rarely has a single right answer. There are thousands of acceptable ways to write a customer email. So instead of “right or wrong”, you’re judging “good enough”: Is it accurate? Is it on brand? Is it safe to send? That usually needs a mix of automated checks, sample reviews by people and clear guidelines — and it’s why generative outputs that face customers almost always need a human somewhere in the loop, at least at first. (We looked at why generative models can confidently make things up in the post on hallucinations earlier in this series.)

They cost differently too

A predictive model is usually small, trained on your data, and cheap to run once built — scoring a million customers might cost cents. Most of the effort is up front: getting clean historical data and building the model.

Generative AI is typically used through a large pre-trained model from a provider, so there’s little up-front training, but you pay per use, often per token. As usage grows, so does the bill. I’ve written about why that matters at scale in the hidden risk of Pure AI products, and about blending traditional software with AI in SaaS vs AI vs Hybrid. The short version: use generative AI where its flexibility genuinely earns its cost, and lean on predictive models and ordinary code for the rest.

Why the hype is mostly about generative

Predictive AI has been delivering value for decades in fraud detection, recommendations, credit scoring and forecasting — it just isn’t very exciting to look at. Generative AI is new, visible and touches work that used to be considered uniquely human: writing, design and coding. The economic estimates reflect that excitement. McKinsey, for example, estimated that generative AI could add the equivalent of US$2.6 trillion to US$4.4 trillion annually across the use cases it analysed^[3].

That’s a big number, and I’d treat it as a sign of potential rather than a promise for any single business. In practice, the quickest, most reliable wins I see still often come from predictive models on data a business already has — with generative AI layered on top to make the results easier to act on.

What I want you to take away

1. Predictive AI forecasts or classifies. Its output is a number, a label or a probability: churn risk, demand, fraud likelihood.
2. Generative AI creates. Its output is new content: text, images, audio or code.
3. They’re judged differently. Predictive models can be scored as right or wrong against history; generative outputs need quality checks and often human review.
4. They’re priced differently. Predictive models are usually cheap to run once built; generative AI is usually pay-per-use and grows with volume.
5. The best systems often combine both. Use prediction to decide where to focus, and generation to help people act on it.