Large language models (LLMs) are increasingly embedded in organizational workflows, yet their mistakes often slip through human review. Prior work attributes these failures to users' limited review capability or engagement. We propose a retrieval‑based oversight model, arguing that error detection improves when oversight‑relevant information is readily accessible at review time. To test this, we conducted two randomized lab‑in‑the‑field experiments with 640 customer‑facing employees. Results show that self‑generated explanations boost error detection and reinforce recall of verification‑relevant reasoning, while cues that reactivate such reasoning help sustain detection performance across repeated LLM use. Theoretically, we define information retrievability as a distinct precondition for effective oversight and identify generative encoding and cue‑supported reactivation as mechanisms that build and maintain it. Practically, lightweight onboarding self‑explanations and daily retrieval cues can make human oversight more resilient as LLM use becomes routine.
Review