Wildberries & Russ described what level of data maturity is needed for accurate AI agents

Wildberries & Russ proposes viewing data description quality as pragmatically as the degree of steak doneness: from Rare to Well-Done. The logic is simple: the better a company describes its tables, fields, business terms and metrics, the fewer AI agents hallucinate and the closer they come to real business insights rather than plausible but useless answers. The company believes the main problem with big data today is not a lack of models, but poor data governance.

In large organizations, new tables appear faster than teams can describe them manually. As a result, analysts spend time searching for the necessary sources, metrics start to diverge between reports, and access to unmarked datasets becomes a security risk. This is especially acute where personal data is involved and strict access control requirements exist.

The manual approach simply stops scaling in such an environment. The first maturity level, Rare, is considered minimally sufficient for safe data operations in this model. At this level, each dataset should have an owner, a physical model, and confidentiality markup.

The physical model can be automatically extracted from system tables and data catalogs, and if field descriptions are empty, AI can attempt to recover them using naming conventions and corporate knowledge bases. Automation works worse with owners: the model can suggest a candidate, but responsibility is still assigned manually. However, marking sensitive data looks like one of the first practical tasks for an LLM: the model can analyze table names, columns, and business terms and assign security tags even before deep content scanning.

At the Medium level, the focus shifts from technical structure to business meaning. A glossary and logical layer appear here, which translate the language of tables and columns into business entities and attributes that make sense to the business. This layer hides service prefixes, complex joins, and storage details, and data stewards can use AI as a copilot for linking fields to existing terms and finding gaps in descriptions.

If an agent is connected to a metadata catalog through MCP, it can deliver the necessary schemas on request, match them with the glossary, and accelerate work that previously took hours. Additionally, Wildberries & Russ proposes extracting relationships not only from storage structure, but also from SQL query logs: they reveal which tables analysts most often join, which filters they use, and how data is actually consumed. The highest level, Well-Done, is needed not just for navigating data, but for full-fledged text-to-SQL and agent systems.

Here, on top of physical and logical descriptions, a semantic layer is built: facts, metrics, dimensions, relationships, filters, and verified natural language queries with ready-made SQL answers. This is the layer that explains to the model what the business means by "active customer," "gross revenue," or other metrics, rather than forcing it to guess meaning from field names. The article provides an illustrative example: if you ask an AI how many active customers there were in March, a model without semantics might simply count rows with active status, although according to the company's rules, an active customer is one who made at least one order above a specified threshold.

According to the author, this is where open standards like OSI become critically important, because they allow describing data meaning in a portable format compatible with modern semantic layer tools. The practical effect is also noted: in Snowflake materials for Cortex Analyst, there is mention of approximately 20 percent improvement in accuracy when working through correctly described semantics, and the target benchmark for real scenarios is over 90 percent SQL accuracy. What does this mean in practice: the market is gradually moving away from the idea that it's enough to simply connect an LLM to a database and expect magic.

The Wildberries & Russ approach shows a more sober trajectory: first bring order to data owners, structure, and classification, then assemble a logical dictionary, and only after that build a semantic layer for agents. For companies wanting to implement AI assistants in analytics, this sounds like an unpleasant but useful truth: the quality of a model's answer now depends directly not only on the model itself, but on the maturity of data description within the business.