How Tongcheng Travel’s DataAgent Turns Analysis and Marketing into AI‑Driven Workflows

In an era of explosive data growth and increasingly complex business scenarios, traditional "manual data extraction, manual analysis, manual marketing" approaches have hit efficiency ceilings. Tongcheng Travel’s data team shares their practice and thinking around DataAgent, focusing on the intelligent evolution of its BI system "Lingdong" and marketing system "Xianzhi".

Background and challenges : Over 80% of business users cannot write SQL, leading to high training costs and reliance on the data warehouse team, which spends 70% of its effort on low‑value repetitive data retrieval. Data entry points are fragmented across BI reports, ad‑hoc queries, and data maps, causing inconsistent metric definitions and severe data conflicts. On the marketing side, creating scenarios is complex, tag selection is difficult among more than 4,000 tags, and strategy analysis heavily depends on expert experience, resulting in low efficiency and homogeneous tactics.

Product architecture evolution : Starting in 2018, the BI system "Lingdong" was launched, followed by attribution analysis in 2020, Copilot mode in 2022, intelligent interpretation in 2023, and ChatBI in 2024, with the full‑process DataAgent loop planned for 2025. AI capabilities were woven into each layer, adding Prompt and Skill management, vector‑based dataset retrieval, intelligent routing, NL2SQL query support, and finally AI‑driven interpretation and reporting.

Analysis Agent: NL2SQL to NL2DSL : The traditional BI architecture follows a generic pipeline of system management → dataset → acceleration layer → visualization → dashboard. After introducing AI, each layer received targeted enhancements, but NL2SQL faced five major issues in 2024: accuracy drops below 70% for complex multi‑table joins, AI hallucinations, inconsistent metric definitions, difficulty persisting query results as dashboards, and lack of business logic understanding. The team identified three bottlenecks—accuracy, security (risk of DML injection), and maintainability—and decided to insert a DSL middle‑layer. Instead of letting large models generate raw SQL, the system first produces a structured DSL constrained by syntax and business semantics, converting DSL to SQL (NL2DSL2SQL). Short‑term scenarios can still use NL2SQL, but enterprise‑grade, long‑term applications require NL2DSL for robustness.

Marketing Agent: Multi‑Agent collaboration : The marketing side adopts a five‑layer architecture (traffic/app/web → service layer → AI capability layer (Agent + large‑model base) → data service layer → infrastructure). A Multi‑Agent orchestration framework deploys four core roles: coordination agent (task decomposition), planning agent (step sequencing), execution agent (tool invocation), and analysis‑memory agent (insight generation and learning). Tag recommendation, the most challenging technical problem, uses RAG + Milvus vector search to bridge the semantic gap between user queries and thousands of tags, followed by a small model re‑ranking based on historical conversion data. This reduces token consumption by 70% and API cost by 40%, while supporting over 100 dialogue rounds.

Evaluation and results : The evaluation pipeline covers prompt management, standard test sets, Langfuse‑based tracing, and real‑time production monitoring. Quantitative outcomes include: • Analysis Agent reduces data‑to‑insight time from days to minutes, improving decision accuracy by 15‑25%. • Marketing Agent shortens scenario creation from hours to minutes, raising autonomous operation to 60% and boosting conversion rates to 14.6% (vs. 4.8% manually). • Overall conversion uplift reaches 35%, with labor cost savings of 30% and API cost reductions of 40%.

Future outlook : The team plans four directions: autonomous learning via reinforcement and human feedback, full‑process automated closed‑loop, cross‑domain deep applications (supply chain, risk control, HR, R&D management), and an AI‑driven collaborative network of specialized agents to tackle complex, multi‑domain business problems.