The Credit Desk

The science

Anchoring was documented by Tversky and Kahneman in 1974: the first number you see disproportionately shapes every estimate that follows, even when you know it might be wrong, and even when you are warned in advance. A 2025 study of 775 experienced managers found that expertise reduces the effect but does not eliminate it.

In AI-assisted decisions this has a direct design consequence: an interface that shows the AI's recommendation before the evidence turns the human review into a formality. The operator does not evaluate the case; they evaluate their willingness to disagree with the anchor. Showing the raw file first and the AI score second costs more time, and buys back independent judgment.

Two of the six AI scores in this simulation were deliberately miscalibrated (one scored a shaky file too high, one scored a strong file too low). Check the table: did you catch them, or did the anchor catch you?