How KOHO built a Claude-powered self-serve analytics tool on years of dbt governance

In a dbt Summit 2026 session, Célia Bru, senior director of data at KOHO, and Gabriel Gambacorta, technical product manager at KOHO (and formerly a member of Bru's data team), argued that the hard part of self-serve analytics has never really been the AI. KOHO is a Canadian fintech with about 250 employees, currently in phase two of three toward a Canadian banking license, and more than 10 percent of its staff are data people.

Bru spent most of her time on what she called the "fuzzy stuff" (data team culture) and the "boring stuff" (data foundations), because those two things, in her telling, are what made it possible to build KoDa, KOHO's in-house self-serve analytics tool, at all.

KoDa lets anyone at KOHO ask a business question in Claude and get an answer back from a governed dbt data mart running on Redshift, with Bedrock and Claude providing the LLM layer. Bru said the individual pieces (a clean dbt mart, a production-grade warehouse, and an LLM already available companywide) had existed since a diagram from KOHO's principal data architect back in March; the missing piece was simply connecting them securely and governably.

The team held itself to firm design principles while building it: KoDa is scoped to analytics use cases only (no debugging or investigating individual transactions), it only ever touches certified dbt mart models with no raw data or PII, it is read-only with no ability to write back to the warehouse, every employee gets the same single experience, every answer traces back to the same company-wide dbt definitions, and the whole thing was built on infrastructure KOHO already had rather than new vendor tooling.

None of that would have been possible, Bru said, without work that came years earlier. She described building buy-in through what she called the "referee position": the data team owns the inputs into a decision but not the decision itself, and it owns metric definitions so that no individual team gets to decide what counts as a win.

When something breaks, the team says so before anyone has to discover it, which is what led KOHO to scrap its entire pipeline in 2022 and rebuild it in dbt from the ground up. That rebuild is now roughly 1,400 models with tests and documentation, all PII stripped out of the models, and Bru pointed out that KoDa's restriction to governed dbt marts has become one of the strongest arguments for getting the rest of the company onto that same foundation, since teams without a governed mart simply can't use the tool.

Gambacorta then walked through how KoDa is actually built: a "KoDa skill," which is markdown files plus pre- and post-hook MCP calls, surfaced through Claude Code, connected to KOHO's dbt marts through a Redshift MCP server. What took real effort wasn't the skill itself but the evaluation layer underneath it: roughly 11 commits from three people over three months went into building context libraries and what Gambacorta called an "eval council," a growing library of about 375 real analyst questions used for both offline evaluation (does a change to the style guide improve or regress answer quality) and online evaluation (every question KoDa answers gets logged, with an analyst on a weekly rotation reviewing new or high-priority ones).

Gambacorta's framing for where the industry should focus next: roughly a third of dbt Summit sessions this year touched on context or the semantic layer, but only about one in 33 talked about evaluation, which he thinks is backwards given how much of KOHO's own effort went into evals rather than the model or the skill itself.

Bru closed by returning to trust: the KOHO data team built its credibility over years by being in the room for every answer, and self-serve tools take the team out of that room. Their interim answer is to log every question KoDa is asked and to still personally review any self-serve result headed to the board.

Longer term, Bru and Gambacorta both argued the bottleneck is shifting from the tooling to the people: teaching a whole company when to trust a self-serve answer and when to come back to the data team, and building data literacy broadly enough that people ask better questions in the first place.

Watch every keynote and breakout from dbt Summit 2026 on demand. Watch dbt Summit 2026 on demand →