Many conversations about AI agents focus on models and frameworks. But when organizations attempt to deploy agents in real operational environments, a different challenge quickly emerges. How agents reliably and securely access enterprise data.

Without reliable access to relevant data, AI agents struggle to support operational decisions. Whether diagnosing equipment failures, assisting customers or monitoring supply chains.

Let’s consider an AI agent diagnosing equipment failures in a manufacturing plant. To recommend the right action, the agent must interpret sensor readings, review maintenance manuals, analyze historical failure reports and check service schedules. Each piece of information lives in a different system and in a different format.

This raises an important question for enterprise leaders and data engineers: how should AI agents access enterprise data reliably and safely?

A complete enterprise data strategy includes governance, pipelines, storage and data quality. But when it comes to how agents retrieve information at runtime, two complementary approaches are emerging:

• Retrieval-Augmented Generation (RAG) for unstructured knowledge.
• Model Context Protocol (MCP) for structured operational data.

These approaches do not replace a broader data strategy. Instead, they represent two important pillars of enterprise AI data access.

What data do enterprise AI agents need?

In operational environments, AI agents rarely rely on a single source of information. Instead, they combine knowledge context with current operational data. If you consider the previous equipment failure example, an AI agent may need to:

• Retrieve troubleshooting procedures from a maintenance manual.
• Review past incident reports describing similar failures.
• Check current vibration level sensors.
• Verify when the equipment was last serviced.

The data needed by the AI agent falls into two categories – unstructured information and structured data. Unstructured information explains how systems behave. Structured data reveals what is happening right now. Both are required for an AI agent to reason and make decisions in operational situations.

RAG for enterprise knowledge

RAG enables AI systems to retrieve relevant enterprise knowledge and incorporate it into the model’s reasoning. Instead of relying on model training data, the AI agent retrieves domain-specific documents and uses them as context when generating responses.

The RAG access pattern pulls from enterprise documents to the retrieval system that the AI agent accesses. From there, the AI agent makes a recommendation. In the equipment failure scenarios, RAG might retrieve:

• Sections of the pump maintenance manual describe vibration thresholds.
• Engineering troubleshooting guides.
• Historical reports describing similar incidents.

This allows the AI agent to interpret abnormal conditions using documented expertise. For example, it may recognize that elevated vibrations combined with increased temperature are a known indicator of bearing wear. However, documentation alone does not confirm that the condition is occurring.

Why RAG by itself is not enough

Operational decisions require current data, not just reference knowledge. Even if the AI agent retrieves documentation describing vibration-related failures, it still needs operational data to assess the situation, such as:

• Current vibration measurements.
• Temperature readings.
• The last maintenance date.

These data points typically reside in structured systems such as telemetry platforms, asset management systems or operational databases. Without access to this information, the AI agent may provide general guidance. It cannot accurately assess the equipment's condition.

MCP for structured data access

MCP provides a standardized interface that allows AI systems to retrieve data from external tools and enterprise systems. Instead of building custom integrations for every system, MCP enables AI agents to interact with structured data sources via predefined connectors.

In the equipment monitoring example, an AI agent could use MCP to retrieve:

• Live vibration and temperature sensor readings.
• Maintenance history from service systems.
• Spare part availability from inventory systems.

This allows the AI agent to combine operational measurements with the knowledge retrieved through RAG.

Combining RAG and MCP

RAG and MCP address different parts of the enterprise data landscape. In the equipment failure example, an AI agent might:

• Use RAG to retrieve troubleshooting procedures describing vibration-related failures.
• Use MCP to retrieve current sensor readings and maintenance history.
• Compare the observed conditions with documented failure patterns before recommending inspection or maintenance.

This combination allows the AI agent to move from generic advice toward data-informed recommendations.

What does this mean for enterprise leaders?

As organizations experiment with AI agents, the focus often begins with the model's capabilities. Over time, a different limitation emerges. The reliability and trust of AI agents accessing enterprise data. Leaders and date engineers evaluating AI agents should consider:

• How agents will retrieve organizational knowledge.
• How will they access operational data?
• Whether those access patterns are governed or observable.

Thinking about these questions early can help prevent many of the challenges organizations encounter when moving from prototype AI agents to operational systems.

• Read more about manufacturing and why predictive maintenance needs more than retrieval in this blog post.

• Discover how SAS® Retrieval Agent Manager can help you get faster insights from unstructured enterprise data.

• Explore how to modernize your data estate for AI success.