Show the manager their own hit rate, not the model's explanation.

The hiring manager has the assistant's shortlist on one screen and her own on the other. The assistant's is better, on the evidence. She takes hers. Nothing in the assistant's explanation, however good, changes that. One thing does: a note that in her last twenty hires, the ones she overrode the shortlist for stayed less long.

The experiment

Dargnies, Hakimov and Kübler ran a hiring task in which managers could decide themselves or delegate to an algorithm that was demonstrably better: it picked the higher-performing worker 66.9% of the time, against 55.9% for the managers. In the baseline, 34.1% delegated. Explaining how the algorithm worked moved that to 32.2%, that is, not at all. Giving managers feedback on their own past accuracy moved it to 49.8%. On the other side of the table, workers chose to be assessed by the algorithm 46.7% of the time, rising to 59.1% when told it ignored gender.

Transparency about the machine did nothing. Transparency about the human did a great deal.

When people accept machine advice, and when they don't

A 2025 meta-analysis in Psychological Bulletin pooled 163 studies and 442 effect sizes across 82,078 participants. People appreciated AI, with a moderate positive effect, when its capability was high and the task did not call for personal treatment. They avoided it, with a larger negative effect, when the task felt personal. Hiring, promotion and pay are about as personal as work gets. A 2026 set of experiments found something more uncomfortable: people rated algorithmic management as less just than human management, yet judged discrimination by an algorithm less harshly than the same discrimination by a person.

What this means for an HR assistant

1. Show the track record, not the reasoning. A brief that reports how the team's past overrides turned out will move decisions more than a paragraph of model explanation.
2. Keep the personal part personal. The assistant prepares the shortlist, the comparison and the sources; the manager meets the people and decides. That division is not a concession; it is where acceptance comes from.
3. Make the exclusions explicit. Workers accepted the algorithm more when told it ignored gender. A brief that states what it did not use is more trusted than one that lists what it did.
4. Treat the reduced-fairness perception as a design input. Algorithmic management lowers justice evaluations even when it is fair, so every brief names the person accountable for the decision.

This is the reason our briefs label the Coworker's own reasoning as reasoning, carry the source of every figure, and end with a named owner. Those are not compliance features. On the evidence, they are what makes a manager take the better shortlist.