Prediction vs Moderation: A Better Test for Hiring AI

Most hiring AI is judged on prediction: can it pick who will succeed, scored with a metric like AUC. A study of 10,765 hires suggests that is the wrong test. The behavioral score in that study barely predicted who would produce, with a deployed AUC of 0.57. Yet it strongly predicted who would benefit from a faster ramp, capturing about 2.8 times more production per day for high-scored hires. The score answered a more useful question: who benefits most from the conditions you can change. That is moderation, and it is a different test than prediction.

Source: "Decision Traces," Saad Bin Shafiq, NODES, 2026. Deployment at a Fortune 500 insurance carrier, N=10,765 agents. Read it on arXiv.

The standard test, and why it misses

The default way to evaluate hiring AI is to ask whether it classifies who will succeed, measured by AUC for a binary outcome. By that test, the behavioral score looked modest: a deployed AUC of 0.57 on 714 agents. A team grading the score on prediction alone would conclude it was barely working.

What the score actually did

The score did not change who ramped fast. Using a split at score 70, 42.2% of low-scored producers and 43.1% of high-scored producers reached the production milestone within 60 days, which is nearly identical. What the score predicted was who got paid for ramping fast. High-scored agents captured about $114 per day of speed acceleration, against $41 for low-scored agents, a 2.8 times difference. The score was identifying who converts a fast start into production most efficiently, not who would produce at all. See the speed findings.

Why the distinction matters

If you grade a hiring model only on prediction, you throw away a model whose real value is telling you who to invest in. The more useful question for most hiring AI is whether it identifies who responds most to the conditions you can change: faster ramp, better onboarding, the right lead assignment. A model that does that earns its keep even when its raw classification accuracy is modest.

The idea travels beyond hiring

The same logic applies wherever an individual interacts with conditions you control. A loan model might not predict who defaults, but might predict who benefits most from specific terms. A clinical support tool might not predict who recovers, but might predict who responds most to a specific treatment. A decision trace, which connects inputs to outcomes across systems, supports both kinds of evaluation, prediction and moderation. See decision traces.

What this does not say

This pattern rests on a small high-score band. The high band had 71 agents, and the key comparison cells were smaller still, so the dose-response pattern is suggestive and needs validation on larger samples. The point is the reframing, not a settled magnitude: judging hiring AI on classification alone can miss its most useful contribution. Single carrier, single deployment.

Frequently asked questions

What is the difference between a predictor and a moderator in hiring? A predictor estimates who will produce. A moderator identifies who benefits most from conditions that affect production, such as a faster ramp. The study's behavioral score worked mainly as a moderator.

Is a low AUC a problem for a hiring model? Not necessarily. AUC measures classification of a binary outcome. A model can have a modest AUC and still be valuable if it identifies who responds most to the conditions you can change.

How should hiring AI be evaluated? Ask both questions: does it predict outcomes, and does it identify who benefits most from favorable conditions. Evaluating only the first can undervalue a useful system.

Does this apply outside hiring? The reasoning extends to any decision system where individuals interact with conditions you control, including lending and clinical decisions.

Prediction vs Moderation: Why Hiring AI Should Be Judged on Who Benefits, Not Just Who Succeeds