Topic

Deterministic Workflows and Agent Collaboration: AI Implementation Practices in Fliggy Air Ticket Fulfillment

In air ticket fulfillment, AI-generated results still need to go through business rules, system states, and handoffs to human operators before they can be truly transformed into actual processing. From the perspective of fulfillment R&D, this presentation focuses on three types of tasks: material review, rule execution, and solution judgment. Through three practical cases—medical refund review, AI-SOP, and an intelligent fulfillment assistant—it introduces how AI capabilities can be integrated into existing systems and operational workflows in the forms of recognition services, deterministic workflows, and Agents, respectively. The presentation focuses on common challenges such as integrating with existing systems, aligning with business semantics, handling exceptions, and handing off to human operators. It also distinguishes three levels—successful service invocation, result adoption, and business completion—to demonstrate the engineering trade-offs involved in moving AI capabilities from being callable to being truly integrated into operations. Outline What needs to be addressed when AI capabilities enter fulfillment systems? Using three typical types of tasks to illustrate the respective roles of AI services, workflows, Agents, and domain systems, and clarify the prerequisites for integrating them into existing processes. Medical Refund Review: How do AI recognition results enter the fulfillment process? Combining AI material recognition results with order information and business rules to generate review conclusions, along with the strategies and practices for automated handling and exception handoffs to human operators. AI-SOP: How can the delivery of executable workflows shift from R&D roles to business roles? Business teams take the lead in defining and maintaining executable workflows, while R&D teams focus on providing tools and acceptance testing, together with version execution, checkpoints, and exception recovery mechanisms. Intelligent Fulfillment Assistant: How can Agents enter customer service operations in complex business scenarios? The organization of context, knowledge, and tools; the ability for business personnel to independently revise capabilities; and the differences in human handoffs between conclusion-based and solution-based tasks. Engineering Methodology Summary and Business Validation Horizontally comparing the applicable conditions of the three types of solutions, summarizing common issues, and clarifying capability boundaries and areas for improvement. Key Takeaways Gain an AI selection framework for industry-specific tasks: Understand the applicable conditions of recognition services, deterministic workflows, and Agents, and clarify the capability boundaries of each type of solution. Understand a business-led, R&D-supported collaboration model: Learn the engineering approach to transforming business rules and experience into maintainable system capabilities. Establish an effectiveness evaluation methodology for real-world operations: Identify key issues between technical usability and business completion, and determine improvement priorities accordingly.