How do we know what AI knows?

Understanding the true boundaries of artificial intelligence starts with asking a fundamental question: what does your model actually know? While Generative AI gets most of the spotlight, non-generative AI continues to do the heavy lifting for decision-makers through three distinct analytical lenses. By breaking down Descriptive, Predictive, and Prescriptive analytics, we can simplify how data is processed and set realistic expectations for what to ask your AI.

Descriptive Analytics

Descriptive Analytics is about feeding an AI model with a data corpus and asking it to describe what’s in it. This is sometimes called “BI” — business intelligence. It is commonly used for organizing large data pools, finding patterns that are complex or hidden from the human eye, or for retrospective investigation.

Imagine ultra-efficient statistical tools — regression, classification — that digest the corpus, segment it, and reconstruct it into trends. The trends can be predefined, meaning we could ask the AI model to describe specific patterns that we’re interested in, or we can ask it to surface anything that is statistically significant.

This can be used to deepen your understanding of your own history of performance, and to shine light on biases and tendencies. Descriptive Analytics only gives you visibility of information that already exists in the data.

Predictive Analytics

Predictive Analytics uses historical data to make predictions on future trends. This type of analytics uses your data for training and learning existing patterns and trends, but for the purpose of assigning labels to, or placing, future unseen data.

Predictive Analytics can be used in surveillance systems, credit risk assessment, anomaly detection, and others. It can also be used to train an AI model to follow your investment philosophy — as inferred from training data — and forecast investment decisions.

Like Descriptive Analytics, this uses statistical and ML methods, but usually more advanced ones with higher dimensions and more dependencies, such as Time Series and Probabilistic Modeling. Predictive Analytics will give you information about unseen events that aren’t in the data, but that go along trends from the data.

Prescriptive Analytics

Prescriptive Analytics goes beyond making linear predictions based on past trends and given data. Instead, it evaluates multiple possible worlds based on possible configurations of learned factors and constraints, then outputs predictions for future outcomes for desired scenarios.

In Prescriptive Analytics, you get to “correct” future events by biasing predicted trends. This could be used if you want to make portfolio adjustments or add context to your analytics and predictions, for example — “invest only in companies that benefit the economy in region X,” or any other dimension that you want to incorporate into your predictions.

So, before deploying your next AI model, step back and ask what you truly need: reflection, forecasting, or optimization. Understanding the distinct limits of descriptive, predictive, and prescriptive analytics keeps your expectations grounded and your outputs actionable. When you align your questions with what your AI actually knows, you transform data from a passive archive into your most effective decision-making partner.