Kingbase KES AI Database is integrating the NL2SQL semantic layer—turning natural language into SQL queries—with vector retrieval and access controls directly into the database kernel. Its central judgment is that the main bottleneck in AI-powered data querying is no longer SQL generation alone. The real challenge is whether enterprises can help models understand business definitions and access data safely.
What This Is
NL2SQL converts natural-language questions such as “What was revenue last month?” or “How are active customers distributed by region?” into executable database queries. The traditional approach often feeds table definitions directly into prompts and lets the model infer the fields for itself. The result may be syntactically valid SQL that still produces the wrong business answer.
KES AI takes a different approach by adding a semantic layer between the physical database and access views. It explicitly defines metrics, dimensions, table relationships, and calendars. Data teams determine what aggregation “revenue” requires, which order types are excluded by default, and how the metric can be broken down. Permissions are then enforced through row-level security to restrict what the model can see.
Industry View
Our judgment is that this approach reflects enterprise data-application requirements more closely than simply upgrading a model. The ability to generate complex SQL does not mean a model knows how “active customers” should be defined or can detect duplicate charges caused by many-to-many relationships. Putting business semantics and data permissions into the same database can reduce intermediate layers and prevent teams from maintaining multiple competing definitions.
This path still carries risks. First, the semantic layer does not emerge automatically. Metric definitions, dimension values, and relationship rules still require continuous maintenance by business owners, so the cost does not disappear. Second, integrating these capabilities into the database kernel does not mean query accuracy has already been validated. Prompt design, model reasoning, sample evaluation, and exception fallback will still determine the final outcome. More importantly, if complex rules for semi-additive metrics or fiscal-year definitions are configured incorrectly, the generated SQL may still execute successfully and present wrong numbers with an appearance of credibility.
Impact on Regular People
For enterprise IT: The competitive focus of data projects may shift from “Which model should we integrate?” to “Who can maintain a unified semantic and permissions system?” Organizations with mature database governance capabilities are likely to benefit more. Companies that have not accumulated consistent business definitions may initially shift the burden onto their data teams.
For individual careers: Natural-language data querying will become easier for business professionals, but the need to confirm metric definitions will remain. Unauthorized queries, duplicate counting, and misuse of metrics will still require managers to establish review and audit mechanisms.
For the consumer market: Enterprise AI assistants will gradually enter customer service, operations, and sales workflows. Results seen by users will increasingly reflect real-time data, but enterprises will also impose tighter controls over what information models can access. Permission design will directly shape the product experience.