International Workshop on Reliable NL2SQL for Enterprise AI (Online)
https://cross-ai.io/conference/2027/preconf/workshops/nl2sql-2026/
(Held in the Cross-AI Symposium 2026)
Introduction
------------
Natural Language to SQL (NL2SQL) has advanced rapidly with Large Language Models (LLMs), achieving strong performance on benchmarks such as Spider and BIRD. However, these gains have not translated into reliable deployment in real-world enterprise and high-stakes application domains, including finance, healthcare, logistics, and enterprise operations.
A key challenge is a persistent gap between benchmark assumptions and production reality. Existing benchmarks typically assume small, clean, and static schemas; fully specified user intent; limited join complexity; few-database settings; and accuracy-centric evaluation. These simplifications support rapid progress on benchmark metrics but fail to reflect real-world conditions.
In practice, enterprise and domain-specific systems are significantly more complex. They span thousands of evolving, poorly documented tables across distributed cloud platforms, with inconsistent business definitions, strict access controls, and heterogeneous data types. This makes both structure and semantics difficult to navigate, requiring user queries to be translated from high-level intent into grounded schemas, disambiguated metrics, and cost-aware execution plans across multiple systems. Despite strong benchmark performance, real-world deployment often fails due to incorrect joins, hallucinated tables or columns, weak grounding in domain metadata, limited scalability, and a lack of cost and system awareness that can lead to unsafe or overly expensive queries.
Meanwhile, NL2SQL is evolving from single-step query generation into agentic analytics systems that must retrieve and reason over schemas, plan multi-step workflows, interact through clarification, execute and validate SQL, and repair failures while operating under cost, latency, and governance constraints.
This workshop frames NL2SQL as a real-world AI systems challenge at the intersection of foundation models, databases, cloud infrastructure, and agentic reasoning. It positions NL2SQL as a core component of enterprise and high-stakes AI infrastructure and aims to shape a research agenda focused on trustworthy, scalable, cross-aware, and production-ready NL2SQL systems, bridging the gap between benchmark success and real-world impact.
Topics
------
Enterprise NL2SQL and Foundation Models
• Large-schema reasoning and retrieval
• Metadata-aware query generation
• Semantic-layer & business metric reasoning
Agentic and Conversational Data Systems
• Ambiguous, interactive, & querying
• Agentic AI for data intelligence & query generation
• Human-in-the-loop structured data reasoning
Cross-Systems and Cloud-Scale Analytics
• Cross-table and cross-database reasoning
• Execution-aware & cost-efficient query generation
• Scalable architectures for NL-to-data interfaces
Safety, Reliability, and Governance
• Reliability and failure analysis in NL2SQL systems
• Cost-aware SQL generation
• Security, privacy, & governance for NL data access
• Deployment-driven evaluation metrics
• Runtime optimization and monitoring
Advanced Reasoning and Applications
• Geospatial & complex structured data reasoning
• High-stakes application domains (e.g., finance, healthcare, enterprise BI)
• Robustness under ambiguity, distribution shift, and noisy schemas
Important Dates
---------------
Please visit the workshop website for more details:
https://cross-ai.io/conference/2027/preconf/workshops/nl2sql-2026/
Submission
----------
All submissions are made through the AirBalloon conference management system and evaluated through a unified process. Accepted archival papers will be published in the indexed Cross-AI conference proceedings.