From Calculations to Context: How AWS’s Semantic Layer Powers Agents to Understand Business Data

On July 31, AWS open‑sourced the Context Ontology Accelerator (COA), a platform that links structured and unstructured data, uses AI to draft enterprise ontologies, and stores the vetted results in a knowledge graph accessible to agents through the Managed Catalog Provider (MCP). The official goal is to make agent decisions more accurate, consistent, explainable, and auditable.

Evolution of the Semantic Layer

Traditional semantic layers focus on metrics, dimensions, relationships, and business definitions for reporting. In the agent era, the layer must also answer how entities relate, which rules apply, the provenance of conclusions, and access permissions. Thus, the semantic layer is shifting from a static analysis tool to the foundational business‑context infrastructure required at agent runtime.

Limits of Retrieval‑Augmented Generation (RAG)

RAG excels at locating answers that exist in documents, but many enterprise decisions span structured data, entity relationships, and business rules. An AWS Builder Center supply‑chain example illustrates a query that must combine supplier data, risk scores, and procurement rules—information spread across multiple sources and requiring logical joins that pure document retrieval cannot resolve.

COA Workflow

COA breaks its workflow into three steps: Scan (connects databases, catalogs, and documents to extract business metadata), Model (AI‑assisted generation of an ontology draft, reviewed and approved by experts, with optional governed metrics), and Serve (exposes the assets via MCP, REST, and SPARQL interfaces for agents).

The key innovation is not a single knowledge‑graph component but the integration of two previously separate asset types: deterministic business metrics and the ontology of entities, relationships, rules, and constraints. Agents can therefore invoke pre‑governed semantics instead of re‑inferring business logic from raw schemas.

Layered Resolution and Governance

COA employs layered resolution: queries that can be answered by governed metrics follow a deterministic path; those requiring structured relationship queries invoke the ontology/virtual knowledge graph; more complex scenarios fall back to an agentic approach. The system also incorporates Cedar authorization and SQL firewall to control who can access which operations.

From Semantic Layer to Context Layer

In June, AWS announced the AWS Context service, describing a "shared, governed context layer" that automatically maps relationships, business rules, and domain knowledge into a knowledge graph for runtime agent consumption. COA’s custom ontology capability is slated to become a native feature of AWS Context, extending the semantic layer’s boundary beyond metrics to a full business‑context layer.

As model capabilities grow, the need for robust semantic foundations becomes more critical: models can generate complex SQL or orchestrate workflows, but they cannot decide which metric is official, which relationship is real, or which rule must not be violated without a stable, governed semantic model.

Conclusion

The semantic layer is evolving from a BI‑centric metric model to an agent‑centric business‑context infrastructure that combines metrics, ontology, knowledge graphs, MCP interfaces, and governance mechanisms, enabling agents to reason over enterprise data safely and effectively.

Sources: AWS What’s New (COA release), AWS COA Documentation, AWS Builder Center (Beyond RAG), AWS Prescriptive Guidance (Semantic Layer for Agentic AI), AWS ML Blog (AWS Context).