Workload Management in Hierarchical Knowledge-Based Service Systems

Abstract

We analyze the structural ability of service systems that rely on agent knowledge to respond quickly and effectively to unanticipated spikes in system workload. Specifically, we consider two-level hierarchical systems in which agents process cases and make decisions concerning service. If a lower level agent is uncertain about a case, s/he can refer it to a more knowledgeable upper level agent. Such a referral tends to increase the quality of the decision, but at the expense of increasing workload and delay at the upper level. We show that, when agents’ assessments are rational, optimal referral decisions have a threshold structure that describes the conditions under which cases should be sent to the higher level. Moreover, these thresholds are monotone in system workload, and depend on agent knowledge. However, because the optimal policy is complex to implement, we develop simple heuristic policies that can respond to a workload spike by adjusting the referral decision criteria based on partial real-time queue length information. Using an extensive numerical study, we find that (a)~there is significant opportunity to improve the performance of knowledge-based service systems by implementing policies that efficiently respond to changes in workload, (b) our proposed heuristics are effective in exploiting this opportunity, (c) sharing assessments across levels can be a potent mechanism for mitigating the effect of workload spikes, particularly under certain conditions, and (d) agent training can improve system performance overall, although moderately effective two-sided training has a stronger impact than highly effective one-sided training when the focus is on agents’ consistency (improving assessment variance), but the opposite is true when the focus is on agents’ accuracy (improving assessment mean).