The bottleneck after the bottleneck
A modern lakehouse fixes the data platform. It does nothing about the queue of people waiting on analysts. At BlueSprig, stakeholders in Finance and Clinical Ops had questions the data could answer, but every question still needed someone fluent in SQL and the semantics of clinical operations data. The replatform removed the infrastructure bottleneck. This layer removed the human one.
The design
- Claude reaches the lakehouse only through a custom MCP (Model Context Protocol) server, a governed interface with the access rules built in
- Semantic access layer: canonical metric definitions, standard filters, and curated context, so the model answers with the organization's definitions rather than improvising its own
- A text-to-SQL eval harness, built in Python, scores generated queries against known-good answers, so we know the accuracy number before stakeholders bet on it
- Guardrails from the company's AI governance framework, which I wrote, along with the vendor-assessment framework beside it
The outcome
Non-technical stakeholders in Finance and Clinical Ops query the lakehouse directly, in plain language, and get answers grounded in governed definitions. Analyst time moved from servicing report requests to improving the semantic layer itself — every definition added makes every future answer better.
Why it matters beyond one company
Most enterprise AI analytics fails in one of two ways: ungoverned access that produces confident nonsense, or governance so heavy nobody uses it. The pattern here is repeatable: MCP as the governed gateway, a semantic layer as the source of truth, an eval harness as the running proof. That's the playbook for putting AI on top of enterprise data in a regulated environment.