Modern enterprises integrate heterogeneous data from operational databases, SaaS applications, event streams, and semi-structured sources to power analytics and machine learning workloads. Despite advances in distributed processing engines and lakehouse architectures, data integration remains labor-intensive and error-prone due to schema drift, semantic ambiguity, brittle transformation logic, and slow incident diagnosis. Recent large language models (LLMs) offer new capabilities for semantic reasoning over schemas, SQL code, pipeline logs, and documentation, enabling automated assistance in mapping, rule generation, and failure analysis. This paper proposes a governed architecture for LLM-assisted data integration that combines unified ingestion and lakehouse-style storage zones with an LLM co-pilot for schema mapping and transformation synthesis, and an observability-driven repair loop for fault localization and targeted recomputation. We introduce a confidence scoring framework that integrates semantic similarity with empirical sample evidence and constraint validation to assess mapping reliability. A human-in-the-loop and continuous integration workflow ensures regression safety and reproducibility before promotion to production pipelines. An illustrative evaluation demonstrates reductions in onboarding time and mean time to repair compared with conventional rules-only integration approaches, while maintaining data quality guarantees. The results indicate that LLMs are most effective when deployed as a planning, validation, and debugging layer grounded in empirical evidence and governance controls, augmenting deterministic data processing systems rather than replacing them.
Sankiti et al. (Sat,) studied this question.