How AI-Native Organizations Share Domain Knowledge

To enable AI agents to share standardized industry domain knowledge, a two-layer semantic layer structure was introduced. It consists of a general-purpose Core based on industry standards and an Overlay containing organization-specific processes, with the Overlay referencing the Core only by ID without modifying it.

This structure was validated for effectiveness through 150 domain questions. When the semantic layer was applied, the number of standard concept citations (grounding) surged from 1.6 to 7.6, and the expected keyword inclusion rate (specificity) rose from 67% to 81%. Notably, the value of organization-specific knowledge became more apparent with harder questions, as evidenced by an observed reversal in LLM judge scores.

However, the semantic layer is merely a compass that provides direction, not an answer key that confirms the actual behavior of the current system. Therefore, a verification loop that searches the actual codebase to verify the direction indicated by the layer is essential. This approach is effective for building a common knowledge base in environments where multiple AI agents handle the same code, but its applicability is limited in domains where no standards exist.