AI is making it easier to query data, generate queries, and build visualizations, even for non-specialists.
For those working in Business Intelligence, this is certainly a significant shift, especially when it comes to topics related to the presentation layer. Some types of analysis may become more accessible to business users who currently depend on reports built in advance by domain specialists.
However, before we can ask a system, for example, what the average margin per customer was, someone first needs to define what "margin" means and what "customer" means, which data should be used, which exceptions should be considered, and how the different data sources should be related to one another.
This process is called modeling, and this is precisely where expertise continues to play a crucial role.
Even today, a generative model can produce SQL very quickly, create a measure, and suggest a visualization. It can translate a question expressed in natural language into a query.
However, it can also produce a query that is technically correct but wrong from a business perspective.
The concept of a customer, for example, may refer to the billing entity, the recipient of the goods, an individual company, the group to which that company belongs, or a consolidated entity built by combining information from different systems.
The data alone rarely determines which interpretation is the correct one. What is needed is extensive knowledge of the context, a deep understanding of business processes, and the ability to translate how the company actually operates into a coherent model.
In Business Intelligence, the quality of data modeling has always been important, but I believe that the rise of AI is increasing the need for rigor in the way data and business concepts are described.
In a traditional dashboard, many decisions are already embedded in the report: metrics, relationships, filters, and analytical paths have been defined during the design phase.
With a conversational interface, users have much more freedom. They can use synonyms, abbreviations, company-specific terminology, and unexpected phrasing. To interpret these requests correctly, the system must be able to rely on clear definitions, consistent naming conventions, well-documented metrics, and unambiguous relationships.
In this sense, modeling also takes on a linguistic dimension: the data model must be understandable to people and interpretable by machines.
In fact, if an AI system is to query data reliably, it needs to know what that data represents, how metrics are calculated, which relationships are valid, and which terms can be considered equivalent within a specific business context.
The semantic layer therefore becomes even more central in connecting an organization's data assets with AI tools, and, of course, the quality of this layer depends on the quality of the underlying modeling.
AI can accelerate many activities involved in building a Data Warehouse, but the need to interpret the business remains.
Integrating ERP, CRM, and vertical applications often means reconciling duplicate data, different business rules, historical exceptions, and definitions that have never been formally documented. It also means deciding which source is authoritative, how a customer should be represented over time, and how operational events should be translated into analytical concepts. All of these activities require experience and domain knowledge.
I expect a broader shift in the way users consume data to happen very soon, and for this reason I believe the underlying foundations will need to become even stronger.
Reliable answers require a model that accurately represents the business.
Top comments (0)