AI for mechanical engineering: connecting MCP to ERP, PLM & MES

The real problem: data islands instead of clarity

In mechanical engineering, the system landscape grows over the years. The ERP (SAP, proALPHA, abas or Microsoft Dynamics 365 Business Central) manages orders, materials management and planning. The PLM holds design and variants, the MES knows the real production progress, and the CRM holds enquiries and service cases. Each system makes sense on its own — the friction arises at the interfaces in between.

Every plant manager knows the consequence: sales promises a delivery date without knowing the real production status. Service looks for a spare part number across several generations of BOMs. Production planning maintains the same information three times. That is not a tool problem — it is a connection problem.

This is exactly where the Model Context Protocol (MCP) comes in. It is an open standard through which an AI agent accesses your specialist systems in a structured, traceable way — without needing a new script or an export to Excel for every question.

Use cases: where AI really saves time in mechanical engineering

1. Order & production status on demand “How far along is production order 4711, and will the delivery date hold?” The agent combines the order header from the ERP with the confirmation status from the MES and names bottleneck operations — without anyone jumping between screens.
2. BOM research “Which assemblies contain bearing type X, and in which variants?” Multi-level BOMs and variant configuration are the complexity trap of mechanical engineering. The agent searches the BOM in a structured way and returns the hits with quantities and positions.
3. Spare-part & service requests A service email mentions a machine serial number. The agent finds the matching BOM, identifies the right spare part, checks stock and replenishment lead time in the ERP and creates a quote or order draft — the planner approves.
4. Quote drafts from enquiries From a technical enquiry, the agent pulls similar past orders, takes over configuration and guide prices and creates a quote draft in the ERP/CRM. Sales and engineering review once — the first draft is ready in minutes instead of hours.
5. Planning & material availability “Which open orders are at risk because of missing material?” The agent reconciles requirements, stock and purchase orders and prioritises — preparatory work that today is often done manually in Excel.
6. Service reporting & customer history Before the customer meeting, the agent summarises open service cases, delivered machines and recent maintenance from the CRM and ERP — field service goes into the meeting prepared.

The solution story: one server, many systems

The decisive lever is not yet another island tool, but bundling. Instead of building a separate AI hack for each system, we place an MCP server in front of your landscape that provides the individual specialist systems as clean, clearly bounded tools:

• ERP connection: Orders, materials management, stock and documents. For SAP we use the growing MCP support — in 2026 SAP is releasing an ABAP MCP server and has integrated Claude into its AI assistant Joule via the Model Context Protocol. Details on our page MCP server for SAP. For Microsoft Dynamics 365 Business Central, the connection runs through the OData/API layer — see MCP server for Business Central. For proALPHA and abas (Forterro is building MCP support via the MyForterro platform) we rely on the documented interfaces.
• PLM/MES bridge: BOMs, variants and production confirmations are connected read-only, so the agent can link design data with production data.
• CRM layer: Enquiries, quotes and service cases. For cloud ERP/CRM such as weclapp there are separate routes — see MCP server for weclapp.
• Read before write: By default, the agent reads and creates drafts. Status changes, orders or quotes are created as a proposal and approved by a human.

The appeal of MCP: where an official server already exists, we use it. Where not, we build a lean adapter on the REST, OData or SOAP interface that is already there. The AI model stays interchangeable — you are not tied to one provider.

GDPR, data sovereignty & security in an industrial context

Design data, calculations and customer orders are core assets of a machine builder. Accordingly, we keep the entire processing chain in Europe and give you back control:

• EU hosting of the MCP server (e.g. Hetzner or the Azure Germany West Central region), no US edge. The server component can also run on-premises in your own network if IT requires it.
• AI model with EU data residency: Claude via AWS Bedrock in the EU region Frankfurt, Google Vertex AI (EU), OpenAI Enterprise EU or Azure OpenAI in an EU region — with a data processing agreement.
• Least privilege: The agent sees only the fields and tenants it needs for the task. No blanket access to the whole ERP.
• Complete logging: Every query and every write action is logged with timestamp, user and data diff — important for audit-readiness and IT audits.
• No training use: Enterprise contracts rule out your order or design data flowing into the training of third-party models.

If you want to go deeper: we have covered the topic separately in GDPR-compliant AI.

Approach & effort

We start small and provable. A single, sharply defined use case — such as the order status query or spare-part research — is usually productive within a few weeks. After that, we expand in increments to further systems and tools.

• Audit first: We review your ERP version, available interfaces and system landscape, and prioritise the use cases by benefit and feasibility.
• First use case as a pilot — deliberately narrow, to show real value quickly instead of integrating blindly for months.

We give a reliable effort estimate after the initial conversation — depending on your system landscape, the desired use cases and the maturity of your interfaces.

Frequently asked questions

What does AI actually bring to mechanical engineering?

Above all, speed on routine questions. Instead of clicking through ERP, PLM and MES, planning, sales or service ask in plain language — “Where is order 4711?”, “Which parts are missing?” — and get the answer from live data. That does not save machines, but it saves many small search and waiting times every day that add up.

Does MCP work with SAP, proALPHA and abas?

Yes. From 2026, SAP offers an ABAP MCP server and has integrated Claude into the Joule assistant via the Model Context Protocol. abas is building MCP connectivity via the Forterro platform MyForterro. Where no official server exists yet, we build an adapter on the documented interfaces — and the data stays in your hands.

Does the AI agent write to production data?

Not by default. The agent reads and creates proposals: a quote draft, a purchase suggestion, a prepared status change. Only after human approval does anything become binding. That keeps responsibility with your employees, and the agent is an assistant, not an uncontrolled actor.

Is this GDPR-compliant for an industrial company?

Yes, if the chain is right: MCP server on EU hosting (or on-premises), AI model with EU data residency and a data processing agreement, least-privilege access and full logging. Design and order data do not leave Europe and do not flow into third-party model training.

Do we have to replace or switch our ERP?

No. The MCP server sits in front of your existing systems and uses their interfaces. You keep SAP, proALPHA, abas or Business Central — the AI layer is added on top, not in their place.

How quickly is a first use case productive?

A narrowly defined use case is typically usable within a few weeks. We recommend starting with one system and one clear question, proving the benefit and then expanding in increments — instead of integrating everything at once from the start.