Human-in-the-Loop AI for Exposure Detection in Loan Portfolio Monitoring

Abstract

Automated detection also allows different aspects of what a bank perceives about its loans to change․ Alerts are provided continuously about concentration build-ups‚ early indicators of credit deterioration‚ weakening collateral and covenant positions‚ and anomalies in the data sent by servicers of loan tape data․ Despite this‚ the next binding constraint to be solved is whether an alert is received by a person when it is generated․ Research in Human-Computer Interaction found that reviewers accept automated feedback‚ ignore it when it makes an error‚ and accept it when it confirms their prior beliefs․ Explanatory output does not eliminate biases and may even reinforce them․ As such‚ having a human reviewer at the end of the automated pipeline ends up being oversight in name only․ The remainder of the article argues the human reviewer should be folded into triage‚ not be assigned at the end․ Combining the literature on detection‚ human-AI decision studies‚ financial alert reviews and supervisory expectations leads to: a taxonomy of loan portfolio exposure signals and human roles; a seven-stage human-in-the-loop reference architecture with an analyst reaching an independent conclusion before seeing the model evidence; an alert decision contract document describing all the information the alert needs to convey to the human decision-maker; and a set of human-oversight governance metrics implemented in line with the model risk guidance‚ the NIST AI Risk Management Framework and the human-oversight obligations in the EU AI Act․ An example using a commercial real estate concentration cycle illustrates the principles․ From this‚ it appears that the quality of monitoring depends at least as much on how alerts are reviewed as on how they are detected․