BPA expert explains how to prepare corporate data for AI assistant implementation

Modern realities place high demands on government agencies and large corporations regarding data processing speed, quality of service for citizens, and internal process efficiency—and AI assistants are increasingly being considered as an optimization tool. Experts conducted an analysis for the GlobalMedia portal on the practical use of AI assistants in the public sector and corporate management to determine where the technology truly solves business problems and where it remains merely a formal element of a digital agenda.

Data documentation is the foundation of reliable AI: why a metrics registry is essential

Often, the implementation of AI assistants turns into a display of technological prowess without any real impact because automation is launched before tasks are clearly defined. Meanwhile, the technology provides tangible benefits precisely when handling repetitive operations, such as processing ad hoc requests in management analytics, where every extra hour spent searching for data leads to delayed decision-making.

Arkady Froymchuk, founder and commercial director of the domestic development company BPA Tech:

Today, AI assistants are often implemented before it is determined what work they should actually perform. This leads to projects that look high-tech but do not provide a noticeable effect. Practical value emerges where an assistant takes on specific repetitive operations: searching for information in corporate systems, preparing reports, processing standard requests, and reducing the time between a question being asked and an answer being received.

This is particularly evident in management analytics. In large organizations, most regular reporting is already automated using BI systems. However, during meetings, negotiations, and presentation preparation, executives constantly have additional questions: what was a specific metric yesterday, what influenced its change, and where have risks or growth factors appeared?

Usually, such requests are passed on to analysts. A specialist must find the necessary data mart, verify the calculation methodology for the metric, and prepare a brief conclusion. As a result, the analytical team spends a significant portion of their time not on deep analysis, but on urgent and often repetitive ad hoc requests.

To solve such tasks, Text-to-SQL agents are used—systems that translate a question formulated in natural language into a query for a corporate data warehouse. The BPA team implemented one such tool for a federal government client. An executive asks a question, and the system identifies the necessary metrics, accesses the warehouse, and generates a concise answer based on the organization's actual data. This reduces the time required to prepare information for operational and strategic decisions and lowers the workload on analysts.

The main complexity of the project, however, was not in choosing a language model. For an agent to provide reliable answers, it is necessary to describe corporate data in detail: define the meaning of each metric, its calculation rules, the source, and the associated data mart. The model must also have strict limitations and use only the information actually present in the company's information systems.

For this purpose, solutions like OpenMetadata or DataHub are used. They allow for the consolidation of descriptions of warehouses, metrics, dashboards, and BI systems into a single catalog. As a result, analysts gain a holistic view of the data structure, and the AI agent understands where to look for the required metric and how to interpret it.

After documenting the main data marts, creating a metrics registry, and setting up constraints, the agent's response accuracy reached 99.8%. If a user requested a metric not present in the warehouse or asked a question outside the system's scope, the agent did not attempt to fabricate an answer; instead, it reported that the necessary data was unavailable and recommended contacting a subject matter expert. For management analytics, this is fundamentally important: a controlled refusal is safer than a convincing but inaccurate answer that could lead to an erroneous decision.

Operating costs were also addressed separately. Using a single large model to search for data, process the request, and prepare the answer leads to excessive expenses. Therefore, the system was built using a multi-agent architecture: more compact models performed the search and data collection, while the main model interpreted the result. As a result, processing a single request required 36% fewer computing resources than using one large model for all operations. For the client, this means lower operating costs and the ability to increase the number of users and requests without a proportional increase in expenses.

The practical conclusion is that implementing an AI assistant is not just about connecting a language model. It is important to first conduct an audit of available data, form a unified metrics registry, describe the warehouse structure, and define access and operational rules for the agent. These tasks should be part of the project implementation scope from the developer's side, alongside the design of the solution's architecture.