Data preparation—the tedious process of denoising and organizing raw information—is undergoing a fundamental paradigm shift as large language models (LLM) replace rigid, rule-based pipelines. Researchers from Shanghai Jiao Tong University have released a comprehensive survey detailing how context-aware AI agents are now handling complex tasks like entity matching and data imputation with minimal human intervention. This move toward Agentic AI—systems capable of autonomous action to achieve specific goals—allows for flexible workflows that understand the semantic nuances of data rather than just following hardcoded logic. The survey introduces a task-centric taxonomy that splits the field into three pillars: data cleaning, integration, and enrichment. In the cleaning phase, models handle standardization and error processing, while integration focuses on identifying how different datasets relate to one another. Enrichment involves generating Synthetic Data or annotations to enhance existing records for downstream analytics. These methods offer superior generalization across diverse domains compared to traditional software, though they require careful prompt engineering to ensure accuracy. However, the transition involves significant hurdles. The authors warn that the high cost of running massive models at scale remains prohibitive for many organizations, and the risk of Hallucination—where the AI confidently generates false or fabricated information—persists even in the most advanced setups. To move forward, the roadmap suggests developing more scalable systems and robust evaluation protocols to ensure that the data processed by AI is actually more reliable than the raw mess it started with.

Data preparation—the tedious process of denoising and organizing raw information—is undergoing a fundamental paradigm shift as large language models (LLM) replace rigid, rule-based pipelines. Researchers from Shanghai Jiao Tong University have released a comprehensive survey detailing how context-aware AI agents are now handling complex tasks like entity matching and data imputation with minimal human intervention. This move toward Agentic AI—systems capable of autonomous action to achieve specific goals—allows for flexible workflows that understand the semantic nuances of data rather than just following hardcoded logic. The survey introduces a task-centric taxonomy that splits the field into three pillars: data cleaning, integration, and enrichment. In the cleaning phase, models handle standardization and error processing, while integration focuses on identifying how different datasets relate to one another. Enrichment involves generating Synthetic Data or annotations to enhance existing records for downstream analytics. These methods offer superior generalization across diverse domains compared to traditional software, though they require careful prompt engineering to ensure accuracy. However, the transition involves significant hurdles. The authors warn that the high cost of running massive models at scale remains prohibitive for many organizations, and the risk of Hallucination—where the AI confidently generates false or fabricated information—persists even in the most advanced setups. To move forward, the roadmap suggests developing more scalable systems and robust evaluation protocols to ensure that the data processed by AI is actually more reliable than the raw mess it started with.