Why Machine Learning Fails at Prioritization Problems

Machine learning has been enormously successful at perception tasks: vision, speech, translation, classification. But over the last few months, I’ve become convinced it’s the wrong abstraction for a large class of problems that show up in real businesses.

These aren’t problems about recognition. They’re problems about prioritization.

What I mean by “prioritization problems”

In many operational settings, the task isn’t to predict an outcome in isolation. It’s to decide what to act on next given limited time and resources.

Examples: • Which customers should a team focus on today? • Which accounts are most at risk right now? • Which leads are worth human attention this week? • Which cases should be escalated first?

In practice, these problems reduce to ranking entities, not labeling them. The top of the list matters far more than global accuracy metrics.

Why accuracy is a misleading objective

Most ML systems optimize for accuracy (or AUC, log loss, etc.) on historical data. But high accuracy doesn’t necessarily produce a useful ranking.

Small changes in data distribution, feature availability, or labeling timing can cause large changes at the top of the list — exactly where decisions are made.

From an operator’s perspective, a system that is slightly less “accurate” but stable and interpretable is often far more valuable than one that optimizes a metric but behaves unpredictably in production.

The retraining tax

Another issue is the operational cost of retraining.

Real business data is: • Sparse and delayed • Full of policy changes • Affected by human behavior reacting to the system itself

As a result, models need to be retrained frequently. This introduces a constant tax: • Engineering overhead • Monitoring complexity • Organizational mistrust when outputs change without explanation

For many teams, this cost outweighs the theoretical benefits.

The explainability gap

Explainability is often treated as a UX feature — a nice-to-have for dashboards. In practice, it’s operational.

Teams need to understand why something is prioritized in order to act on it. Without that understanding, outputs are ignored, overridden, or quietly worked around.

A ranking that can’t be explained is rarely used for long.

A different way to think about these problems

Prioritization problems have different requirements than prediction problems.

What matters most is: • Stability over time • Determinism (the same inputs produce the same outputs) • Explainable structure, not opaque scores • Low operational overhead

This suggests a different mental model — one closer to constraint-based or physics-inspired reasoning than to probabilistic model fitting.

Instead of asking, “What is the predicted probability of X?” the more useful question is often:

“Which combinations of conditions consistently precede this outcome, and how strongly does each entity match them right now?”

Closing thought

Machine learning is a powerful tool, but it’s not a universal one. For problems where the goal is deciding what to do next — not modeling the world — optimizing for accuracy may be the wrong objective entirely.

I suspect many teams don’t actually need better models. They need better ways to prioritize action.

Why machine learning fails at prioritization problems