The future is the ultimate unknown. There’s no more coveted business knowledge than, “What’s going to happen?” Yet, since we can’t eliminate uncertainty, we can only do the next best thing: Manage it.
Large-scale operations constitute a “numbers game” that involves millions of decisions. We can play this game more effectively by putting odds on each individual outcome. Which transactions should we block as potentially fraudulent? Which machine parts should we replace before they break? Which customers should we discount before they cancel?
Enter predictive AI. It learns from data to gauge the odds for each case. While there is no magic crystal ball, we do have the next best thing: probabilities, which tell us the chances that each individual will click, buy, lie, die, commit an act of fraud or turn out to be a bad debtor.
GenAI is not suited. It’s newer, sexier and more advanced – and yet it doesn’t replace predictive AI, it only augments its. The two are intrinsically destined to remain distinct endeavors and disciplines, even as they blend into a unified technology ecosystem. Here’s why.
GenAI Doesn’t Replace Predictive AI
There will always be uncertainty. No matter how advanced algorithms become – even including large language models composed of trillions of parameters – we cannot predict future outcomes with very high confidence in general. Rather, we can only place odds on outcomes. As algorithm sophistication increases, prediction improves, but with diminishing returns. Ultimately, we face a ceiling on how well we can predict the behavior of humans, corporations, machines and other kinds of artifacts.
To manage uncertainty with predictive AI, an enterprise must follow a very particular end-to-end paradigm. These projects are inherently “predictive” – they need per-case odds, not genAI. Their function is to drive many decisions, each by way of an estimation of the odds for that particular case. To do so, a predictive AI project must be highly customized, largely by determining three things: 1) what it should predict, 2) how well it must predict and 3) how the predictions are to be used to drive decisions.
GenAI is not well suited for predicting on this granular, per-case level. Razi Raziuddin, CEO of FeatureByte – which uses genAI to improve predictive AI – helps the world understand this limitation. “LLMs and other genAI models deliver value for many business problems,” he told me, “but they weren’t designed to analyze large tabular data, much less run machine learning algorithms across such data.”
GenAI is made with ML, not for ML. Even though it’s built with ML methods that are more sophisticated than those normally used for predictive AI, genAI itself doesn’t constitute the same kind of “prediction machine.” At its heart, an LLM is an ML model that predicts the next word (token) in a sentence. As such, it works well with human languages and it accomplishes a certain amount of “reasoning” (by some definitions). But using an LLM to perform a well-defined data-analysis task – including ML itself – is usually inelegant overkill and, indeed, usually ineffective.
GenAI can’t single-handedly execute on predictive AI projects. To use predictive AI, an organization must work explicitly with its predictive capabilities in three ways: 1) Train the ML model for the prediction goal at hand, 2) evaluate the model as to its value for operational improvement and 3) operationalize the model, using it to predict for individual cases and drive decisions accordingly. Without special modification, LLMs aren’t well suited for any of these three stages.
But GenAI Helps Predictive AI
While GenAI doesn’t perform the core analysis of predictive AI, it supports predictive AI projects in various ways. After all, genAI can code, design and explain. GenAI has been applied to serve as a plain-spoken copilot that elucidates and that explains how ML models decide, and as a predictive AI coding assistant and predictive feature generator.
These developments bring predictive AI and genAI together within an emerging, unified ecosystem. This works by embedding predictive AI capabilities within genAI systems. This way, for example, users can ask a conversational AI which customers are at risk of defection and how best to design a targeted marketing campaign to retain them.
The unification of predictive and generative AI hasn’t yet received its due. Justin Swansburg, former VP at DataRobot, points out that “there’s a lot of opportunity and I don’t think, to date, it has really gotten as much attention as it has deserved… in terms of engineering context, explaining the output, integrating them into workflows [and] incorporating predictive models as tools.”
Predictive AI will always have a role in this world, because uncertainty is an indelible aspect of life and business. GenAI systems can only realize state-of-the-art capabilities for managing uncertainty by explicitly incorporating predictive AI functions. This will allow genAI access to an established, well-structured paradigm for generating – and acting on – granular predictions. Once they’re integrated in this way, genAI serves to support and supercharge predictive AI.
For more about combining these flavors of AI, attend my presentation, “Seven Ways to Hybridize Predictive and Generative AI,” at the free online event IBM Z Day (live on November 12, 2025 and available on-demand thereafter). If your work involves hybrid AI, consider submitting a proposal to speak at Machine Learning Week 2026.
