This paper suggests the multi-objective optimization framework, which is defined formally, to the dynamical task allocation of human and chatbot in customer-support systems. The model develops a constrained utility maximization model that maximizes the performance, cost and ethical risk in a joint manner. Normalized response accuracy, latency, user satisfaction, cost of operation and ethical risk are all objectives of the objective vector. An adaptively weighted-sum scalarization is performed under restrictions of the highest allowable bias, probability of toxicity, and service-level latency. Simulated annealing with adjusted temperature schedule is used to solve the non-convex optimization problem. This framework is tested using 18,742 annotated customer-support interactions of low-risk telecom and billing domain. The data were coded to resolve correctly, satisfaction (5-point Likert scale), complexity of the task, and outcome of the escalation. Performance in ethics is measured by both the toxicity classification scores and the disparity in bias and fairness is measured through statistical parity difference and equalized performance variance. A standardized A/B experiment will be used to compare the proposed model with rule-based escalation and performance-only optimization control groups. Findings indicate statistically significant increases in aggregate normalized utility ( +8.6, p < 0.01), task completion rate ( +5.2), and less exposure to the ethical risk. Sensitivity- This establishes consistency with the weight configurations. The paper adds a clear mathematical model, reproducing assessment strategy, and politically consistent analysis of scalable human-in-the-loop system structure.
