All metrics and P0 questions in a unified context layer: Qonto has set this goal for the end of 2026.
Against a backdrop of implementing an agent-based data analytics architecture. The neobank describes this journey in four phases, the first of which was text to SQL. A bot trained on specific instructions translated natural language, but did not itself execute the queries on Snowflake. The task fell to users, who had to understand SQL.
In a second phase, Qonto granted everyone direct access to the data warehouse via Claude. It judged whether queries and data were correct. But it did so from an inconsistent foundation: naming inconsistencies across domains, duplication of business logic, metric definitions that varied by team, etc.
Qonto provides an example question posed by the customer success: “How many active customers in Germany in the last quarter?” Claude identified an is_active column within a table and used it. Except that it did not account for a customer success parameter: for a given quarter, only customers who had been active every month count as active. Similarly, Claude filtered “Germany” based on the customer’s country of domicile, whereas the customer success team segments based on the portfolio’s country affiliation. One can add the aspect of “last quarter,” which Claude interpreted as the last three months rather than the calendar quarter.
Une couche Omni pour introduire du déterminisme
Qonto then switched to Claude-based agents that queried a context file in Markdown. It contained approved metric definitions, validated join paths, and documented entities. SQL was compiled from these definitions. Questions that did not map to any defined metric returned a response of “not yet modeled.”
This approach did not scale. The agent lacked the necessary context to answer complex questions. Qonto therefore embarked on a refactor of its data model architecture. It moved away from BI-focused models toward dimensional modeling: a row represents a payment, a billing month, etc., all organized around events and their descriptions.
The neobank chose Omni for its context layer. This enabled it to move toward deterministic results, breaking each question down into a combination of predefined metrics and dimensions. It stands in contrast to the initial iteration (text to SQL): the LLM selected tables, inferred relationships, chose aggregations, and wrote the SQL. Four probabilistic steps with as many chances to err.
Data literacy before AI literacy
The approach comes with a clearly visible cost period, admits Qonto. For several months, business teams saw engineering effort focused on the existing models, with no delivery of new features. More broadly, other projects had to be deprioritized.
Moreover, before cultivating AI literacy, there was a need to cultivate data literacy. Indeed, the agentic architecture no longer intermediates an analyst; the business user inherits the responsibility of arbitration.
Qonto asserts that coverage is now its principal constraint. It aims for a rate above 90% via its Omni context layer. Between governed semantic queries and exploratory free use, it continues to adjust the dial. This boundary, it anticipates, will shift as model capabilities progress.
