Find the pay gap before the audit does

The first output of the model was ridiculous, which is why everyone remembers it. A 400-person software company, prepping for the EU pay-transparency rules, ran a regression on its own compensation data. Top predictor of current salary, after role and level: the year the employee was hired. Not performance ratings. Not tenure, exactly. Hire year, because people hired in hot markets came in high and stayed high, people hired in downturns came in low and stayed low, and every annual raise cycle since had preserved the entry gap like sediment.

The model was not broken. It was describing the company honestly for the first time: compensation was, to a measurable degree, a fossil record of negotiating conditions. And folded inside that fossil record sat the number the general counsel actually feared. Controlling for role, level, location, and tenure, women in the engineering org earned 4.1 percent less than men in like-for-like positions. Nobody had decided that. The org chart had accumulated it, one anchored offer and one un-negotiated raise at a time.

An unexplained pay gap does not need a villain to be real, and it does not need to be intentional to be a liability.

Companies mostly argue about the raw gap, the difference in average pay between groups, and the argument is always the same: "that's mix, not discrimination; the groups sit in different roles." The defense is half right, which is what makes it dangerous. The raw gap is indeed the wrong number. But the reflex answer, comparing averages within job titles, is also wrong: titles are coarse, levels drift, and a comparison that controls for nothing but a label can hide a real gap or invent a fake one with equal ease.

Regulators and plaintiffs' experts do neither. They run controlled regressions: model pay as a function of the legitimate factors a company says it pays for, level, function, location, tenure, relevant experience, and measure what group membership predicts after those factors are accounted for. The EU directive now forces disclosure and remediation around exactly this kind of analysis, with reporting thresholds at 5 percent. Walking into that regime having never run the regression on yourself means the first rigorous analysis of your pay system will be performed by someone whose incentives you do not share.

The method is the compensation specialist's standard toolkit, applied in-house and on a schedule. A controlled wage regression per job family, with the company's own stated pay factors as covariates. A decomposition, Oaxaca-Blinder is the classic, that splits the raw gap into the portion explained by legitimate factors and the residual that is not. Diagnostics a statistician insists on and a dashboard never mentions: whether the "legitimate" covariates are themselves contaminated, because if level assignments or starting salaries already encode the bias, controlling for them launders it into the explained column. That contamination check, modeling promotion velocity and entry compensation by group, is the part that separates a defensible analysis from a comforting one.

Then the remediation arithmetic, which is an optimization problem, not a spreadsheet sort. Blanket raises to every below-line employee overshoot the budget and under-target the actual inequity; the constrained version finds the minimum-cost set of adjustments that brings every group's residual gap under threshold, flags the outlier individuals driving each pocket, and prices the fix precisely. We built this stack for our own comp reviews, hire-year fossil and all, and the first run reshaped an adjustment budget we had already believed was generous.

(The same regression, run quarterly, becomes a control system: every offer and raise cycle lands as a data point, and drift gets caught at one basis point per quarter instead of compounding silently for five years.)

The 400-person company fixed its 4.1 percent for about 310,000 dollars a year in targeted adjustments, quietly, inside a planned comp cycle. The identical gap, surfaced by a regulator or a class action after the transparency rules bite, prices at back pay, penalties, legal spend, and a public number attached to the company's name in perpetuity. Same regression, wildly different invoice.

TensorLabs builds pay-equity analysis pipelines that run the auditor's math on your data before the auditor exists. If your last equity review compared averages inside job titles, you have not yet seen your residual gap. Send us your headcount and reporting deadline, and we will walk you through what controlled analyses typically surface at companies your size.