--- title: "Semantic analytics in production — case study | Jake Williams" canonical: https://jakeawilliams.com/work/semantic-analytics/ description: "A production Claude + Snowflake Cortex semantic layer delivered through an enterprise plugin marketplace: 12 plugins across six business domains, quarterly reserve review 3 days → under 30 minutes, QBR prep 4 days → under an hour, ~200 monthly active users at 30% lower token consumption." --- Home / Work / Semantic analytics Case study · AI & semantic analytics # Semantic analytics in production. Outcome first: a production Claude + Snowflake Cortex semantic layer, delivered through an enterprise plugin marketplace I created — 12 plugins across six business domains. Claims, Policy & Underwriting, Finance, and Legal ask governed questions in plain English; executives assemble QBRs, reserve reviews, and financial reports from it. The quarterly reserve review went from 3 days to under 30 minutes and QBR prep from 4 days to under an hour; ~200 monthly active users (40% of the company) run 2K+ queries a month at 30% lower token consumption; regulatory reporting stays consistent; and cross-domain metrics like policy erosion went from impractical to routine. Most enterprise GenAI pilots never reach production; this one is company-wide. claude snowflake cortex mcp dbt dagster Role Lead engineer — created & run it Scope Enterprise-wide · 12 plugins · 6 domains Status In production Stack Claude · Cortex · dbt · Dagster · MCP 01 · Problem Insurance runs on questions: how is this book performing, what's eroding, does this quarter's regulatory report match last quarter's definitions? Answering them meant analyst queues, hand-built extracts, and — the quiet killer — subtly different metric definitions across domains, which is how a company ends up arguing with its own numbers. The naive fix, "point an LLM at the warehouse," fails in a regulated industry for a reason that deserves respect: an answer you can't govern is a liability with good grammar. The real problem was never can a model write SQL — it's can the business trust what comes back, every time, within the access each person already has. 02 · Approach The constraint set came first: answers had to draw on governed, curated data — not raw tables; metric definitions had to be consistent across domains and with regulatory reporting; and the system had to serve non-technical users, not just analysts. That pointed at a semantic layer rather than a chatbot — and at a marketplace rather than a monolith . The intelligence lives in authored domain skills that encode how each domain's data actually works and bridge the data catalogs into agent-ready data products. Each plugin packages the skills a domain needs; the marketplace makes them discoverable, governable, and supportable at enterprise scale. Writing skills is closer to authoring documentation for a very literal colleague than to prompt engineering — and it's where the leverage is: once claims and policy could be read together, cross-domain metrics like policy erosion went from impractical to routine. 03 · The system Business users analyst → C-suite, plain English Plugin marketplace 12 Claude plugins · authored domain skills Snowflake Cortex executes against the governed gold layer Lakehouse medallion on Snowflake · dbt + Dagster · fed by Kafka CDC The governance rail isn't a feature of the system — it is the system. The chat interface is just the last mile. 04 · Results Result Detail 3 d → <30 min Quarterly reserve review — cross-domain plugin knowledge, deck templates, period-over-period point-in-time analysis; QBR prep fell from 4 days to under an hour 12 plugins · 6 domains Enterprise marketplace, enabled company-wide — ~200 monthly active users, 2K+ queries/month C-suite in production Five C-suite regulars among users building QBRs, reserve reviews, financial reports, and canonical dashboards; token consumption 30% lower than unassisted LLM use Trusted & compliant Snowflake Cortex evaluations were the breakthrough — one set of governed definitions, the same number in the dashboard, the filing, and the chatbot New metrics unlocked Cross-domain measures (e.g., policy erosion) previously impractical to produce 05 · Learnings The semantic layer is the product; the model is a component. Domain skills age like documentation, not like prompts — budget for their upkeep. A marketplace is an organizational design as much as a technical one: it needs a roadmap, evaluations, governance, and support, and someone accountable for all four. The speedup lives in the workflows — measured per workflow, not claimed as a multiplier on people. The pattern that survived contact with a regulated business is the one the industry converged on: AI that works inside the governance you already built, not around it. Building governed AI analytics in a regulated business? I've made the mistakes already — happy to compare notes. Email info@jakeawilliams.com ← All work