Decision Science

Generative models sample the next unit by probability rather than look up a record. That explains two things at once: output is fluent without being guaranteed true, and the same property makes them strong precisely where no unique answer exists.

The practical split follows: anything verifiable — numbers, regulation text, client data, prices — must be reconciled against a primary source; anything divergent — options, analogies, first-draft translation, structuring — can safely start with a model. Blurring the two categories is where most "the AI is unreliable" complaints actually come from.

A quick calculation: the chance of randomly calling eight of ten outcomes correctly is only about 5.5%. But ask fifty people to guess ten times each, and the chance that at least one of them gets eight right is roughly 94%. Identical result, entirely different meaning depending on whether you are looking at one person or the best of a crowd.

This does not mean results are luck. It means that before treating an outcome as evidence of skill, establish the denominator. The same question recurs in client development, channel evaluation and hiring; the remedy is to move the criterion from the result to a verifiable process.

The classic case comes from graduate admissions: an overall admission rate looked clearly unfavourable to one group, yet within nearly every department that group was admitted at a favourable rate. The difference came from applying disproportionately to departments with low rates to begin with.

The practical use is immediate. When overall margin falls, split it by customer tier, product line and region before concluding anything — the whole may be declining while every part improves, or the reverse. Compressing a high-dimensional fact into one reported number is one of the most common systematic misjudgements in organisations.

From very high up, a complex technology looks like a simplified box: clean logic, clear path. Come closer and the key trade-offs appear. Use it hands-on and the sheer number of details exceeds every expectation.

Each altitude has a correct use — direction from the top, sequencing from the middle, problem-solving from the ground. Mismatching them is the most common organisational pathology: leaders specifying details from high altitude, or frontline staff unable to move the whole. So for any important call, ask explicitly which altitude you are standing at and which input is missing.

Clustering groups similar objects and is the most underrated of the four, especially for market segmentation, customer profiling and cost-versus-value tiers. Prediction seeks certainty inside uncertainty, and the closer to fundamentals and the shorter the horizon, the more certain it gets. Association mining finds co-occurrence, useful for cross-selling and risk flags. Anomaly detection isolates what deviates.

One cost must be understood: association rules describe correlation, not causation, and false positives carry ethical weight. The most-cited historical case is a retailer inferring a customer's pregnancy from purchase changes and mailing marketing material to her home before her family knew — a demonstration of the method's commercial value and of the harm it can do in the same motion.

Red Capital (Walter & Howie) dissects a financial system from the inside: banks acting as policy tools channel credit to favoured sectors; bad loans are shuffled between "good bank/bad bank" structures; beyond official debt, off-balance-sheet obligations accumulated through financing platforms run far higher — the risk never vanishes, it changes carriers and final bearers. Headline indicators can stay presentable for years; the real fragility hides in the answer to one question: if it breaks, who pays last?

The method generalises: every system has on-book and off-book halves — assess health by tracing where risk finally settles, not by its intermediate carriers; stability bought with efficiency has a price — controls breed arbitrage, arbitrage breeds tighter controls, so judge any reform by which link of the chain it actually cuts; and for enterprises, when assessing counterparties, partners or the macro environment, first establish who stands last in line on every chain.