Centralized self-serve AI analytics platform connecting Claude to governed Snowflake data

The Challenge

Care Advantage, Inc., a trusted family of in-home care brands operating across six US states, had business users across Operations, Revenue Cycle Management, Value-Based Care, Accounting, and Sales all dependent on IT and data teams for analytics.

Every ad-hoc question, whether a business query like quarter-over-quarter revenue growth or a technical one like how a fulfilment ratio is calculated, required manual effort, created bottlenecks, and risked unanswered or inconsistent answers.

Accordion designed and deployed a centralized self-serve AI analytics platform that connects Anthropic’s Claude Sonnet 5.5 and Opus 5.0 directly to Care Advantage, Inc.’s Snowflake environment via MCP, routing queries intelligently across five business domains and governing access through organizational-level skill deployment.

Value Created

5x adoption growth

in two months, from 50+ to 250+ daily queries, with 5-10x faster ad-hoc analysis.

60+ hrs saved weekly

across the Morning Operations Digest and After-Hours report, previously built by hand.

Hours to minutes

in data lineage traceback across ETL transformations, from source to Gold layer.

Self-serve query answering

for business and technical questions, backed by Snowflake RBAC.

“Questions that used to wait in a queue for IT now get answered five to ten times faster. Our teams feel the difference every day.”

– Mark Zawodny, Chief Information Officer, Care Advantage, Inc.

Our approach

Connected Claude to Snowflake via an MCP server, mirroring Matillion and Tableau business logic in Snowflake, versioned in Git with CI/CD pipelines, to create a single governed data layer
Grounded every query in Care Advantage, Inc.’s own definitions by loading domain-specific KPI reference files and business-context markdown before Claude reasons over the data
Routed each question to one of five domains via few-shot classification, deployed an organizational-level hierarchical skill via Claude plugins to designated users, and verified live Snowflake schema before writing SQL
Executed governed SQL via the MCP connector, logged user-reported issues to a tracked Snowflake backlog, and traced ETL transformations end to end for rapid upstream and downstream lineage
Reduced hallucinations and token use via a centralized KPI skill that exposes only the definitions relevant to each domain
Roadmap in motion toward agentic, proactive workflows that flag anomalies before being asked

Phased roll-out

Phase 1:

Connect to Snowflake & ground in KPI context

Linked via MCP; Matillion and Tableau logic mirrored with Git CI/CD.
Reference files and business-context markdown loaded before reasoning over any query.

Phase 2:

Route by domain

Each query classified and matched to one of five domains via few-shot examples.

Phase 3

Verify against schema, answer & improve

Live schema checked before writing SQL, with undefined KPIs logged to a backlog.
Governed SQL executed, answers formatted to fit the data, gaps logged to a tracked backlog.