Will OSI (Open Semantic Interchange) be to business data what Iceberg is to technical data?
The initiative is moving in that direction, at least. Snowflake is leading it within a consortium that brings together around fifteen other companies, including a French one: Mistral AI.
The promise of a common language, especially for AI
We are essentially at the stage of a statement of intent. The promise: to develop a standardized semantic layer for data analysis. In sight are business intelligence tools… but above all AI agents. According to Accenture, these are expected to become the primary “users” of enterprise software by 2030.
Salesforce mirrors this forecast to frame its involvement in OSI. Data can be unified without their meaning being aligned, the American company explains. A customer-satisfaction score can be defined on very variable scales by different lines of business. This problem is exacerbated by logical gaps, even minor ones, that can exist between tools—for example, weeks starting on different days (typically Monday for one, Sunday for the other).
Salesforce – or more precisely Tableau – will focus its initial contributions on bidirectional exchange of semantic metadata, the propagation of governance information, and the native query logics (using the source platform’s runtime to preserve integrity).
A project arriving too late?
OSI will be an open project, under an Apache license. Supposed to rest on YAML, it will include metadata specific to large language models, such as notions of custom instructions and synonymy. The models developed will be paired with translation modules for the platforms of each participating vendor.
Among these participants is dbt Labs, which has already managed to rally a large portion of the ecosystem to its semantic layer launched in 2022. There is also Cube, which has chosen a headless approach independent of any BI tool.
As Select Star, also participating in the initiative, summarizes, a truly transversal semantic layer would yield gains in explainability and portability. And ultimately, in the relevance of analysis between dashboards and AI assistants.
A former Cube veteran in the data space is less enthusiastic. For him, OSI arrives too late. The initiative would have been valuable as long as humans continued to manage the semantic layer. Or, it is likely that this task will be entrusted, in the short term, to AIs (he cites Claude Code and Codex CLI).
One should note that Microsoft is not in the loop. After all, with DAX and MDX formats, the Redmond company already captures a substantial share of semantic-layer usage.
