Retrieval-augmented large language models for bank complaint prioritization: accuracy, bias transfer and fairness-aware retrieval

Abstract Banks must decide which customer complaints to escalate, and large language models (LLMs) now offer to automate this decision. Zero-shot LLMs rank complaints that end in monetary relief close to chance, because the outcome depends on institutional history that the complaint text does not contain. This study tests whether retrieval-augmented generation (RAG), which supplies similar past complaints and their outcomes, closes that gap, and whether it transfers historical disparities between consumer groups. Using 309{,}348 complaints from 2021--2024 as retrieval memory and complaints from 2025--2026 as queries, drawn from the Consumer Financial Protection Bureau archive, we evaluate Qwen3-8B and Claude Haiku~4.5 under six context conditions against three non-LLM baselines. On a pilot set, similar cases with their outcomes raise ROC AUC from 0.61 to 0.76--0.77 for both models; similar cases without outcomes, or random cases with outcomes, give no gain. On the 3{,}000-complaint test set, Qwen3-8B rises from 0.56 to 0.81, matching supervised models (0.80--0.82) but not the historical relief rate used directly as a score (0.84). Institutional memory also widens the recall gap between servicemembers and other consumers, from +0.6 percentage points without context to -4.1 with full context, approaching the -6.0 to -8.1 points of the base rate and supervised models, because servicemembers' meritorious complaints retrieve neighbours with lower historical relief rates. For complaint prioritization, what the model is shown matters more than which model is used, and retrieval memory requires a fairness audit before deployment.