A P&C insurer went from stored procedures to a warehouse, then to AI across the company

The engagement started as a fractional chief data officer role and became a standing data and AI leadership engagement: stack selection, hiring, data-quality tests the business signs off on, a Claude rollout and now a skills platform. Each layer made the next one possible.

Spring: foundations and people

We selected the stack with the data leader (Fivetran into Snowflake, dbt core inside Snowflake, Power BI on top) and stayed on Power BI rather than migrating. The first hire was made with an AI-first technical assessment we designed: a data engineering problem broad enough that pasting it into Claude does not finish it. Candidates reported spending 80% of their time understanding the problem and 20% generating.

Data-quality tests on the legacy tables catalogued what could not be trusted, then stayed in place as regression guards. The data leader adjudicates findings with the business, which is where the real bottleneck was.

Autumn: above the dashboard

We pushed the roadmap above the dashboard, building conversational AI alongside the first dashboard in each business area rather than waiting for modeling to finish.

Claude was rolled out at the end of April with code access first and data access later, which cleared approval months sooner. A company-wide skills platform followed: one library across Claude, Copilot and Gemini, an eval gate, usage by department, workshops and office hours.