Snowflake, Salesforce & Co Draft the “Rosetta Stone” of Data Semantics
Open Semantic Interchange (OSI) wants to stop every tool re-inventing what “customer lifetime value” means.
The 20-company coalition
Front-lined by Snowflake, Salesforce, dbt Labs, BlackRock and ThoughtSpot, the group has open-sourced a vendor-neutral spec that lets metrics, dimensions and business logic travel intact across BI dashboards, AI agents and cloud warehouses.
Why now?
- 78 % of data teams report conflicting definitions of the same KPI inside one company.
- Agentic AI can’t reason if every tool speaks its own dialect.
- Regulated giants (see: BlackRock’s Aladdin platform) need one golden semantic ledger for risk and compliance.
How OSI works
- YAML-based “semantic passport” – describes measures, grains, hierarchies and access policies.
- Open-source translators – Apache-licensed modules that convert the passport into dbt models, Tableau data-sources, Snowflake views, Python data-frames, etc.
- Governance glue – lineage and row-level security tags are preserved, so a policy set in Salesforce still applies when the same metric is queried by a Python notebook.
Early wins
- dbt Labs: 30 % drop in duplicate metric builds during pilot with a Fortune-50 retailer.
- ThoughtSpot: natural-language questions now return identical answers whether asked in Tableau, Excel or Slack bot.
- BlackRock: cut 40 % of pre-trade data-reconciliation hours by syncing Aladdin risk metrics into Snowflake via OSI.
Road-map
- Q1-2026 v1.0 spec + reference code on GitHub.
- Q2-2026 certification programme for BI/AI vendors.
- Q3-2026 domain packs (finance, health, retail) contributed by SIG and industry councils.
The catch
Success hinges on cloud hyperscalers NOT yet in the room—Google, AWS, Databricks—adopting rather than duplicating the spec. OSI’s charter tries to lock that in: any extension must be contributed back, keeping the semantic layer a public good.
Bottom line
Define “revenue” once, use it everywhere—whether you’re prompting an LLM, building a dashboard or training a fraud model. If OSI becomes the HTTP of business logic, data-engineering toil becomes a config file, not a Jira ticket.
