MCP Protocol: Connecting AI Agents to Your ERP System

The biggest bottleneck in enterprise AI adoption is not the AI itself. It is connecting AI agents to the systems where business data actually lives: ERP, MES, SCADA, HR, and financial platforms. Every custom integration is expensive, brittle, and creates a maintenance burden that grows exponentially with each new system.

The Model Context Protocol (MCP)^[1], introduced by Anthropic in late 2024, provides a standardized way for AI agents to discover and interact with external systems. Think of it as the USB standard for AI-to-enterprise connectivity: one protocol, many devices.

How MCP Works

MCP follows a client-server architecture built on JSON-RPC 2.0^[1]. The core components are:

MCP Server: Exposes a set of "tools" that represent operations on an external system. An ERP MCP server might expose tools like get_sales_orders, query_inventory, or create_purchase_order. Each tool has a defined input schema (what parameters it accepts) and output schema (what data it returns).

MCP Client: The AI agent (or the orchestration platform running it) that discovers available tools and invokes them. The client calls tools/list to discover what tools a server offers, then calls individual tools as needed to complete its task.

Transport Layer: MCP supports multiple transport mechanisms. For enterprise ERP integration, HTTP with Server-Sent Events (SSE) is the recommended approach, as it works through firewalls and load balancers. Local integrations can use stdio transport for simplicity.

MCP for SAP S/4HANA

SAP's OData APIs provide a natural foundation for MCP server implementation. A SAP MCP server wraps OData service calls as MCP tools. For example, the A_SalesOrder OData entity becomes an MCP tool that AI agents can call to query, create, or update sales orders.

The key advantage over direct OData integration is discoverability. An AI agent does not need to be pre-programmed with SAP's data model. It calls tools/list, receives a description of available operations, and determines which to use based on its task. When SAP is updated with new modules or custom fields, the MCP server's tool list updates automatically.

Authentication flows through the MCP server, not the AI agent. The server holds SAP credentials (typically an OAuth 2.0 client or X.509 certificate) and enforces access control before forwarding requests. The AI agent never sees the SAP password.

MCP for Nordic ERP Systems

Nordic manufacturers often run Monitor ERP, Visma, Jeeves, or IFS alongside or instead of SAP. These systems have varying API maturity:

IFS Cloud provides comprehensive REST APIs with OpenAPI specifications, making MCP server development straightforward. IFS's built-in projection layer translates entity models into API endpoints that map well to MCP tool definitions.

Monitor ERP offers a SOAP-based API that requires a translation layer. The MCP server converts Monitor's XML-based operations into JSON-schema-validated MCP tools. This adds development effort but provides a clean abstraction that isolates the AI agent from SOAP complexity.

Visma products (Visma.net, Visma Business) expose REST APIs of varying completeness. The MCP server approach is particularly valuable here because it can aggregate tools from multiple Visma product APIs into a single discoverable interface.

Microsoft Dynamics 365 uses OData V4, similar to SAP. The MCP server pattern is essentially the same: wrap Dataverse entities as MCP tools with proper input/output schemas.

Security Architecture

Connecting AI agents to ERP systems demands rigorous security controls. The MCP architecture addresses this through several layers:

Tool-level authorization. Each MCP tool can enforce fine-grained permissions. A compliance reporting agent might have read access to financial data but no write access. A procurement agent might create purchase orders but not approve them above a threshold.

Data filtering. The MCP server can filter sensitive fields before returning data to the AI agent. PII, salary data, or trade secrets can be redacted or excluded based on the calling agent's role.

Audit logging. Every tool invocation is logged with timestamp, calling agent identity, parameters, and response summary. This creates the Article 12-compliant audit trail required under the EU AI Act^[2] for AI systems interacting with enterprise data.

Rate limiting and circuit breakers. MCP servers should implement rate limiting to prevent runaway agents from overwhelming ERP systems. Circuit breakers automatically stop forwarding requests if the ERP system becomes unresponsive.

Implementation Roadmap

A practical MCP implementation for ERP integration follows four phases:

Phase 1 (Week 1-2): Read-only tools. Start with query operations: sales orders, inventory levels, financial summaries. These carry zero risk of data corruption and deliver immediate value for reporting and analysis agents.

Phase 2 (Week 3-4): Validated write operations. Add creation tools with strict schema validation: create sales orders, log production data, submit expense reports. Every write operation requires schema validation before reaching the ERP.

Phase 3 (Month 2): Workflow integration. Connect MCP tools to approval workflows. An AI agent creates a purchase order draft; the MCP server routes it through the existing ERP approval chain before execution.

Phase 4 (Month 3+): Multi-system orchestration. Connect multiple MCP servers (ERP, MES, HR, carbon tracking) so agent swarms can orchestrate cross-system workflows like compliance reporting that pulls data from five different sources.

References

[1] Anthropic, Model Context Protocol Specification, November 2024 — open standard for AI-to-enterprise system connectivity via JSON-RPC 2.0.
[2] European Parliament and Council, Regulation (EU) 2024/1689 (EU Artificial Intelligence Act), OJ L, 12.7.2024 — Article 12 on record-keeping for AI systems.
[3] SAP SE, SAP OData V4 API Reference — OData-based API framework for SAP S/4HANA integration.

Next step: DWS IQ provides pre-built MCP servers for SAP, IFS, Dynamics, Monitor, and Visma. Connect your ERP to AI agents in days, not months. See the integration catalog at dws10.com.