Ever since we showcased the potential of AI agents and MCP (Model Context Protocol) servers within MES (Manufacturing Execution System) and Data Platform contexts at the MES & Industry 4.0 Summit 2025, things have taken off fast. Everyone’s excited. Rightfully so.
At Critical Manufacturing, we are actively developing a number of MCP-related features that will enter technical preview later this year.
And yes – the potential is enormous. Connecting LLMs and agents directly to real tools, systems, and operations unlocks intelligent, context-aware insights and actions within automation and MES environments. The results so far? Genuinely impressive.
But with that power comes some real pitfalls. And we have already seen what happens when teams rush in.
The Dream of MCP is Real
MCP allows large language models to interact with tools, such as databases, applications, devices, dashboards. They do it through structured APIs and data. For manufacturing, this could mean:
- Asking “Why did Line B slow down yesterday?” and getting real insights
- Logging quality defects with a single sentence
- Auto-generating dashboards with prompts “show me the OEE of my area over the last 7 weeks”
- Getting contextual comparisons: “explain me the reasons for the differences in performance of product 1 vs product 2”
It is game-changing stuff. But like any great tool, it only works when built on a solid foundation
And yet… many teams are simply skipping critical phases
Take a look at what’s already happening:
- MCP servers are being released like crazy, sometimes with little thought given to what problem they solve, whether the data is ready, or how they will be safely used.
- Our internal tests show huge performance and behavior differences across LLMs, especially in system prompting and RAG (retrieval-augmented generation) strategies
- Thousands of MCP servers are publicly exposed, some with no authentication, enabling access to powerful tools that can leak sensitive data — or even execute remote commands. (SCWorld analysis)
- Even well-meaning deployments, like Microsoft’s MCP tool for SQL Server, are running into performance issues, confusing behavior, and limited workflows.
Four Reasons to Strategize Before Deploying MCP in Manufacturing (especially in regulated spaces):
1. Don’t “expose everything.” Be specific.
Just because an LLM can use 50 tools doesn’t mean it should. Tool overload confuses agents and users alike. Start with a small, meaningful set: e.g., “log a defect,” “query production rate,” “create maintenance task.”
2. Security isn’t optional. It’s survival.
MCPs expose real capabilities. Not just data, but also actions. Without proper isolation, auth, or input validation, you are opening up risks ranging from data leakage, or even real data loss, to different vulnerabilities.
3. If your data is messy, the AI’s answers will be too.
LLMs hallucinate. If you feed them unstructured, unlabeled, or ambiguous data via MCP, they’ll confidently give wrong answers. Strategize your data architecture and content. Create data models, curate your data, add descriptions, add context. Build data tools that are solid, holistic and explain themselves.
4. Action tools = big power, big risk.
Reading from a table is one thing. But when an LLM can perform actions and modify configurations, or run jobs, things get serious fast. Use approval flows. Add human-in-the-loop. Build rollback paths. Log everything. MCP for control and write-access is not just a tech problem, but an operational risk.
Our Philosophy?
Start small. Focus on real user value. Build with tight constraints and auditable systems. The endgame of intelligent, autonomous operations is worth the effort. However, especially in manufacturing and regulated industries, cool tech needs mature architecture and well-thought-out strategies.
We will share more as our previews roll out. In the meantime, if you’re exploring MCP in manufacturing, build it like it is going into production. Because it will soon.
