Improving service efficiency without sacrificing quality: a multi-agent AI system for hospitality email response automation

Purpose This study examines how multi-agent artificial intelligence (AI) architectures can automate hospitality email responses while maintaining service quality standards. The research addresses the growing challenge of email volume management in hotel operations where response time expectations increasingly conflict with operational capacity constraints. Design/methodology/approach A three-month pilot study evaluated a multi-agent system across 200 properties within a large global hospitality enterprise operating thousands of properties across multiple segments. The architecture employed three specialized agents using Claude 3.5 Sonnet via AWS Bedrock: a triage agent for classification and security, a knowledge retrieval agent with hybrid search capabilities and a drafting agent with validation mechanisms. Human evaluation of 2,500 emails (10% sample) assessed accuracy, tone, clarity, policy compliance and response effectiveness against human-only baselines. Findings The multi-agent system reduced average response time by 77% (from 18 hours to 4.2 hours) while improving customer satisfaction scores from 73.5% to 78.9%. First contact resolution increased from 68% to 83.6%, and factuality accuracy achieved 96.4% in human evaluation. The architecture's agent-specific validation layers prevented information hallucination while maintaining 100% personally identifiable information protection compliance. Research limitations/implications Hotel operators can implement multi-agent architectures to address email volume while maintaining quality by wrapping large language models with specialized validation logic rather than deploying multiple AI models. Properties should establish confidence thresholds (0.70 for human review, 0.85 for automation) and prioritize agent-specific policy enforcement over model selection. The approach proves particularly effective for select-service and extended-stay brands facing high inquiry volumes with limited staffing. Originality/value While task decomposition, validation layering and retrieval grounding are established principles in AI systems research, their integrated application to asynchronous, compliance-sensitive hospitality email automation at enterprise scale represents the primary contribution of this study. Drawing on a deployment across 200 properties and 25,000 emails, this research provides empirical evidence that architectural design, specifically agent-specific validation mechanisms and coordinated workflows, drives reliability in production service environments without requiring multiple foundation models, offering a replicable framework for maintaining service quality at scale.