The Legacy Integration Problem
The world's largest enterprises run on ERP systems built in the 1990s and 2000s—SAP R/3, Oracle E-Business Suite, JD Edwards, and their successors. These systems contain decades of transactional data, encode complex business logic across thousands of configuration parameters, and process millions of transactions daily. They are also fundamentally incompatible with the APIs that modern AI systems expect.
ERP systems expose functionality through proprietary protocols: BAPI calls for SAP, SOAP web services for Oracle, custom RFC interfaces for legacy integrations. Modern AI agents, built to consume RESTful JSON APIs, cannot natively communicate with these systems. The result is a hard barrier: AI can analyze data exported from ERP, but it cannot act on ERP data in real time—the critical gap between insight and action.
What is Model Context Protocol?
Model Context Protocol (MCP) is an open standard that defines how AI models exchange context with external data sources and tools. Rather than requiring each tool to implement a custom integration with each AI model, MCP specifies a universal protocol: a tool exposes a list of capabilities (what it can do) and a schema for each capability (what inputs it expects and outputs it produces). Any MCP-compatible AI model can then discover and invoke any MCP-compatible tool without custom integration work.
For enterprise AI teams, MCP is transformative because it decouples the AI model layer from the tool integration layer. Instead of maintaining a bespoke integration between, say, Claude and SAP, the team maintains an MCP server for SAP (which exposes SAP capabilities via the standard protocol) and an MCP-compatible AI orchestrator (which can consume any MCP server). New AI models can be swapped in without re-implementing integrations; new tools can be added without modifying the AI layer.
MCP for SAP: A Practical Architecture
Implementing MCP for SAP involves building an MCP server that wraps the SAP BAPI layer. The server exposes capabilities such as "retrieve_material_master," "create_purchase_order," "get_invoice_status," and "update_work_order"—each with a JSON schema defining the required parameters and the expected response structure. The MCP server handles authentication (SAP RFC credentials), protocol translation (converting JSON requests to RFC calls), and response normalization (converting SAP responses to JSON).
The AI orchestrator calls the MCP server's tool list on startup, dynamically building a registry of available SAP capabilities. When an agent task requires accessing SAP data or triggering an SAP action, the orchestrator selects the appropriate capability from the registry, constructs the required parameters (using the LLM to interpret the user's intent into structured API parameters), and invokes the capability via MCP. The MCP server handles the SAP interaction and returns a structured JSON result.
Security and Governance Considerations
Connecting AI agents to production ERP systems introduces serious security and governance concerns. The MCP server must enforce least-privilege access: each capability should be backed by an SAP role with only the permissions required for that specific operation. Read capabilities should never have write access, and write capabilities should have explicit approval workflows before they execute destructive changes.
Every MCP call should be logged with the full context: which agent made the call, for which user task, with what parameters, and what result was returned. This audit trail is essential for compliance in regulated industries and for diagnosing errors when an agent-initiated ERP action produces unexpected results. Many enterprises require a separate approval workflow for agent-initiated write operations—an MCP middleware layer that intercepts write calls, routes them through a human approval queue, and only forwards them to SAP after approval is received.
Measured Impact
Enterprises that have successfully deployed MCP-based ERP integrations report three categories of impact. First, cycle time reduction: tasks that previously required a human to log into SAP, locate the relevant record, and manually update a field—typically taking 15-30 minutes per instance—are completed by the agent in seconds. Second, error reduction: agents following strict JSON contracts make fewer data entry errors than humans working in complex SAP screens. Third, capacity liberation: the human time freed from routine ERP interactions is redirected to higher-value analytical work.
The investment in building and maintaining MCP servers for ERP systems is substantial, but it is a one-time investment that benefits every AI agent deployed subsequently. Organizations that complete this integration layer first report dramatically faster deployment cycles for subsequent AI use cases, as each new agent can immediately access the full range of enterprise data and actions.