How is it different from generative AI?

It produces a number rather than content: a likelihood, a score or a forecast. That makes it comparable between cases, thresholdable, and checkable against what actually happened afterwards, which is a set of properties generative output simply does not have.

What makes these projects fail?

Usually not accuracy. It is that nobody was going to act differently, so a good score arrives somewhere with no authority attached to it, and the project is judged on a number while producing no change in what anybody does.

Where should we set the threshold?

Wherever the cheaper mistake is. Moving the line trades false alarms against missed cases, and which of those costs your organisation more is a business judgement rather than something a technical team should be optimising in isolation from the people who bear the cost.

Glossary

Predictive AI

Systems that forecast or score something, such as which customers are likely to leave, rather than producing new content.

In plain terms

Software that estimates what will happen or scores how likely something is. It has been doing useful work for decades, and it went quiet in conversation not because it stopped working but because something newer arrived.

Why it matters

Because a great many problems presented as needing something new are forecasting problems with a settled answer. Reaching for a general model where a score would do costs more, varies more between runs and is harder to check.

How it works

It produces a number rather than content: a likelihood, a score, a forecast. That is the whole distinction from generative systems, and it is why the output is easy to compare, to threshold and to audit after the fact.

A prediction is worth nothing unless somebody acts differently because of it. That is the question these projects most often skip, and a score nobody is allowed to act on is an expensive way of describing a situation people already sensed.

The threshold matters more than the accuracy. Where you draw the line decides how many false alarms and missed cases you get, and that is a business decision about which mistake costs more, not a technical one to be optimised in isolation.

It can be checked against what happened, which generative work usually cannot. Forecasts made last quarter can be compared with the outcome, so this is one of the few places where a system's claims are directly and cheaply testable.

It learns from what your organisation did before, so it inherits your history. A model predicting which customers to pursue, trained on who was pursued successfully, will recommend more of the same, which is a limitation to plan around rather than a defect to fix.

Its predictions can change the thing they predict, which is a trap with no equivalent in generative work. Scoring accounts as likely to lapse and then contacting those accounts means the score influenced the outcome, so measuring it afterwards measures your own intervention as much as the model.

It is frequently the right tool where something more elaborate is being proposed. Where the question is how likely, how many or which of these, a score is cheaper, steadier and easier to defend than asking a general model to have an opinion.

Where the value is, and where the attention goes

Effort concentrates on the first two steps and value is created entirely at the third. A team can spend months improving a score from good to slightly better and change nothing whatever about the outcome, because nobody at the third step has been given the time, the authority or the instruction to do anything differently. The reverse is also true and less obvious: a mediocre score that reaches somebody who acts on it is worth more than an excellent one that reaches a dashboard. That suggests an order of work most projects run backwards. Establish first what somebody would do differently and whether they are able to do it; only then ask how good the score has to be to justify that action, which frequently turns out to be much less good than the team assumed. The fourth step closes the loop and is what makes this category unusual: you can check whether last quarter's predictions were right, and almost nobody does.

Seen in the wild

Scoring which accounts are likely to lapse, so somebody can call them.
Attio
Forecasting volume from historical data rather than asking a model to guess.
Power BI Copilot
A general model asked for a likelihood where a purpose-built score would be steadier.
ChatGPT

Common misconceptions

People assume

It has been superseded by generative AI.

In fact

They answer different questions. One estimates how likely something is, the other produces content, and reaching for the second where the first was the question costs more and produces a less checkable answer.

People assume

The accuracy is what decides the value.

In fact

What decides the value is whether anybody acts differently because of the prediction. An accurate score that reaches nobody able to act on it describes a situation expensively and changes nothing about it.

Questions

How is it different from generative AI?: It produces a number rather than content: a likelihood, a score or a forecast. That makes it comparable between cases, thresholdable, and checkable against what actually happened afterwards, which is a set of properties generative output simply does not have.
What makes these projects fail?: Usually not accuracy. It is that nobody was going to act differently, so a good score arrives somewhere with no authority attached to it, and the project is judged on a number while producing no change in what anybody does.
Where should we set the threshold?: Wherever the cheaper mistake is. Moving the line trades false alarms against missed cases, and which of those costs your organisation more is a business judgement rather than something a technical team should be optimising in isolation from the people who bear the cost.