KOHO's Claude-Powered Analytics Tool Rests on Years of dbt Governance

At a dbt Summit 2026 session, KOHO's Célia Bru, senior director of data, and Gabriel Gambacorta, technical product manager, made the case that the difficult part of self-serve analytics was never the AI itself. KOHO is a Canadian fintech with roughly 250 employees, more than 10 percent of them data staff, and is in the second of three phases toward a Canadian banking license.

Bru focused on data team culture and data foundations, which she said made KoDa possible at all. KoDa lets any KOHO employee ask a business question in Claude and get an answer from a governed dbt data mart on Redshift, with Bedrock and Claude supplying the LLM layer. The components existed by March in a diagram from KOHO's principal data architect; the missing piece was connecting them securely and governably.

Design principles were firm: KoDa handles analytics use cases only, touches only certified dbt mart models with no raw data or PII, is read-only, gives every employee the same experience, traces answers to company-wide dbt definitions, and runs on existing infrastructure rather than new vendor tooling.

That rested on earlier work. Bru described the data team's "referee position": it owns the inputs into a decision and the metric definitions, but not the decision. When something breaks, the team says so first, which led KOHO to scrap its entire pipeline in 2022 and rebuild in dbt. That rebuild now spans roughly 1,400 models with tests and documentation and all PII stripped out. Because teams without a governed mart cannot use KoDa, the tool has become a strong argument for moving the rest of the company onto that foundation.

Gambacorta explained the build: a "KoDa skill" of markdown files plus pre- and post-hook MCP calls, surfaced through Claude Code and connected to KOHO's dbt marts via a Redshift MCP server. The real effort went into the evaluation layer: roughly 11 commits from three people over three months built context libraries and an "eval council," a library of about 375 real analyst questions used for offline evaluation of whether changes improve or regress quality, and online evaluation logging every answer, with an analyst on weekly rotation reviewing new or high-priority questions. He noted that about a third of dbt Summit sessions touched on context or the semantic layer, while only about one in 33 covered evaluation, which he considers backwards.

Bru closed on trust: the data team built credibility by being in the room for every answer, and self-serve tools remove them from it. For now, every KoDa question is logged and any self-serve result headed to the board is still personally reviewed. Longer term, both argued the bottleneck is shifting from tooling to people: teaching a company when to trust a self-serve answer and when to return to the data team, and building enough data literacy that people ask better questions.

Source: dbt Labs Blog