“Potential” is where the promotion gap hides.

Most talent reviews have two axes. Performance is what someone did, usually rated with evidence. Potential is what the room thinks they could do, usually rated without any. The nine-box grid treats both as numbers. One of them is a prediction made by a manager in a meeting.

The study

Benson, Li and Shue followed 29,809 management-track employees at a large retail chain. Women received higher performance ratings than men on average and lower potential ratings, 8.3% lower on average. They were 14% less likely to be promoted, and the difference in potential ratings accounted for about half of that gap. The part that matters most: the lower potential ratings were wrong. Women who were rated lower on potential went on to outperform men with the same rating. The system was systematically underestimating one group, and the record held the evidence all along.

This builds on the same authors' earlier finding, from 131 firms' sales staff, that promotion follows current performance even when it predicts worse management: doubling a salesperson's sales raised their chance of promotion by about a third and lowered the sales of each person they went on to manage by 6.1%. Firms promote on the wrong column, then rate potential in a way that compounds it.

The check any HR system can run

The two columns exist in every performance module: a potential rating from cycle N and a performance rating from cycle N+1. The test is whether the first predicts the second, and whether it does so equally for men and women, for sites, for age bands. It is a few lines of code on data most organisations already hold, and almost nobody runs it, because the potential rating is treated as an input rather than as a claim.

1. For each person with ratings in two consecutive cycles, pair last cycle's potential with this cycle's performance.
2. Within each potential band, compare the average subsequent performance by group. If one group outperforms within the same band, that group was under-rated.
3. Report by band and group, above a minimum group size. Never by name.
4. Put the table in front of the people who sign the promotion list, before they sign it.

What a Coworker does with it, and what it must not

In our design this is a scheduled brief, run once a cycle under the permissions of the head of talent, with every figure computed in code and traced to the ratings it came from. It reports a pattern. It does not re-rank anyone, propose a corrected potential score, or touch an individual's record. The EU AI Act treats systems used for promotion decisions as high-risk, and a system that profiles individuals in that context is always high-risk; a group-level check that informs the people who decide is the opposite of that.

The quiet implication is larger than the gap. If potential ratings do not predict performance, the nine-box's second axis is measuring the rater, not the rated. That is worth knowing before the next calibration meeting, whichever group it falls on.