Foundation · 6 min
What you will learn
- Why the distinction changes what you have to check, not just what you call it
- What each one does, how it fails and how you tell it is working
- How a single ecommerce feature can use both
- Which to choose for a given problem
Why the distinction matters
They are sold under one word. A product page that says AI-powered could mean either, and often means both in different places. The label does not tell you what the feature does, what data it needs or what happens when it is wrong.
They fail in opposite ways. Predictive AI fails quietly and keeps running. Generative AI fails in full sentences that read as correct. A check designed for one catches nothing in the other.
They are judged by different evidence. A predictive feature is evaluated against what actually happened afterwards, which means recording outcomes over time. A generative feature is evaluated against the source it drew from, which means sampling and verifying. Neither method works on the other.
So the useful question about any AI feature is not whether it uses AI. It is whether it is estimating something or producing something, because the answer decides how you supervise it.
The two families side by side
Predictive AI: estimating what is likely
What it does: Produces a number about something it has not seen: a score, a rank, a class or a forecast. It never writes anything. It tells you how probable, how many or which one.
Ecommerce examples: Scoring how likely an order is to be fraudulent, forecasting how many units to stock next month, ranking which products to show first, or flagging which customers look likely to stop buying.
How it fails: Quietly, by drifting. It keeps producing confident numbers after the conditions it learned from have changed. A fraud model trained before a new payment method launches will keep scoring, just worse, and nothing in the interface will say so.
How you tell it is working: Compare its calls against what actually happened. Of the orders it flagged, how many were fraudulent? Of the stock it forecast, how much sold? You need outcomes recorded over time, so measurement has to be set up before launch, not after.
Generative AI: producing something new
What it does: Produces content from instructions and context: text, images, summaries, translations. It does not estimate anything. It writes, rewrites, or condenses.
Ecommerce examples: Drafting a product description from approved attributes, summarizing a long run of reviews, translating a catalog, or drafting a support reply from your published policies.
How it fails: Loudly but invisibly, by stating something wrong in a well-formed sentence. The output is not flagged as uncertain because it does not read as uncertain. A missing specification becomes an invented one.
How you tell it is working: Sample and check against the source. Take a batch of output, verify each factual claim against the product data or policy it came from, and track how often a reviewer has to change something. Fluency tells you nothing.
When one feature uses both
The two often sit in sequence, with predictive choosing and generative explaining. A merchandising feature can predict which products are most relevant to a shopper, then use generative AI to write a short line about how two of them differ. Search can rank results predictively, then summarize what the shopper appears to be looking for.
Keep the two jobs separate when you set it up. If the ranking is poor, better writing will not rescue it, and if the writing invents a difference between two products, the ranking was never the problem. Measuring them together hides which half is failing.
How to choose
Choose predictive AI when the core problem is estimating, classifying or ordering: how many, how likely, which first. Choose generative AI when the core problem is producing or condensing language: write this, summarize that, translate the other. If the correct answer is already known and could be written down, neither is the right tool, and a rule or ordinary automation will be cheaper and more reliable.
Key takeaways
- Predictive AI estimates. Generative AI creates or transforms.
- Predictive AI drifts silently. Generative AI is wrong fluently.
- Each needs its own evaluation method, set up before launch.
- One workflow can use both, with predictive choosing and generative explaining.
- Generative AI cannot supply a fact your store never recorded.