Architecting the Semantic Layer for AI Agents

The recent unveiling of enterprise-grade AI agents, exemplified by Databricks' Genie One, marks a definitive shift. The focus is no longer on passive data analysis but on autonomous action within the enterprise. For data architects, this shift forces a critical re-evaluation of a core component we've curated for decades: the semantic layer. Its purpose is about to be radically redefined, moving from a descriptive model for human analysts to a prescriptive, machine-executable contract for AI agents.

Simply connecting an LLM to a BI tool's API is a fragile, short-term tactic. To productionise AI agents that can safely and reliably interact with your business, we must architect a new kind of semantic layer—one built for action, not just insight.

Why is the traditional BI semantic layer insufficient for AI agents?

A traditional BI semantic layer is a read-only abstraction designed for human interpretation. It excels at defining business metrics, hierarchies, and join logic for analytics, but it is fundamentally incapable of telling an AI what it can *do*. An AI agent requires a machine-readable catalogue of business entities, their relationships, and, most critically, the permissible operations or actions that can be performed upon them.

Consider a BI semantic layer definition for "customer churn". It might define churn as the count of customers who have not made a purchase in 90 days. An analyst can see this, understand it, and build a report. An AI agent, however, needs to know more. It needs to know it can execute an action like `trigger_customer_retention_campaign` for a specific `customer_id`, that this action requires a `campaign_budget` parameter, and that it is governed by a policy preventing contact with customers who have opted out. The classic semantic layer has no construct for this functional knowledge.

What are the core components of an "Actionable Semantic Layer"?

An actionable semantic layer is a deliberate architectural pattern comprising three tightly integrated components. It provides the central nervous system that allows autonomous agents to perceive, reason, and act upon the business environment in a governed way.

First is the **Governed Entity & Metric Store**. This is the evolution of the classic semantic layer, defining the core nouns and numbers of the business—customers, products, revenue, inventory levels. Tools like the dbt Semantic Layer or Cube, underpinned by a central governance catalogue like Unity Catalog, form this foundation. Consistency here is non-negotiable.

Second is the **Action & Tool Catalogue**. This is the critical new element. It is a machine-readable, semantically rich registry of available business operations, typically exposed as internal APIs. Each registered action includes its function signature, required parameters, pre-conditions, and expected outcomes. To make this discoverable via natural language, the descriptions of these tools are embedded and stored in a vector database, allowing an agent to find the right tool for a task like "find and escalate overdue invoices for high-value clients".

Third is the **Dynamic Policy & Entitlement Engine**. This component serves as a real-time checkpoint. Before an agent executes a discovered action, it must query this endpoint to verify that it has the explicit permission to do so. This engine evaluates the agent's identity, the user context it's operating under, the specific entity it's targeting (e.g., `customer_id=4512`), and the requested action against a set of centrally managed business rules and access policies. This transforms governance from a static report to an active, real-time control.

How do open table formats and governance platforms enable this?

This advanced architecture is only viable when built upon a modern, reliable data foundation. The data lakehouse, powered by open table formats like Apache Iceberg or Delta Lake 3.1, provides the transactional integrity and data quality guarantees required by the Entity & Metric Store. Without ACID transactions and schema enforcement at the storage layer, the semantic definitions become unreliable, and agentic actions could be based on inconsistent or corrupt data.

Furthermore, the time-travel capabilities inherent in these formats are essential for auditability. When an AI agent executes a transaction, you must be able to reconstruct the exact state of the data before and after the action. This is a foundational requirement for debugging, accountability, and regulatory compliance.

Centralised governance platforms like Databricks Unity Catalog become the unifying control plane. Unity Catalog's ability to manage access not just to tables but also to functions (UDFs), volumes (unstructured data), and AI models allows you to implement the Policy & Entitlement Engine within a single, consistent framework. It provides the technical substrate to define that a specific AI agent `role` can execute a registered `function` against a specific `table`—the core of actionable governance.

What does this mean for Australian organisations?

For Australian organisations, adopting an actionable semantic layer is a direct mechanism for implementing trustworthy and compliant AI. This architecture provides the technical guardrails necessary to align with local requirements like the Privacy Act and adhere to the principles outlined in frameworks such as the NSW AI Assessment Framework (AIAF). The AIAF's focus on transparency, fairness, and accountability is not just a policy goal; it becomes an architectural outcome.

By explicitly defining and cataloguing permissible actions and enforcing entitlements at runtime, you create an inherent audit trail. Every action an AI agent takes is preceded by a verifiable policy check and results in an immutable transaction log. This level of technical assurance is critical for organisations in highly regulated sectors like finance and healthcare, giving boards and regulators confidence that AI systems are operating within defined, safe boundaries. This approach is central to how we at Precision Data Partners help clients across Sydney and NSW de-risk their AI investments and build systems that are both powerful and responsible. A well-architected, actionable semantic layer is a cornerstone of any serious responsible AI strategy.