Answering the audience’s questions
Integrating agentic AI into legacy manufacturing does not require replacing existing infrastructure. Panelists from Litmus, ZF Group, CESMII, and Arch Systems at AI Frontiers 2025 outlined a practical six-part approach: free trapped data, add semantic context through ontologies, respect brownfield reality using interface classes, target repetitive tasks nobody wants to do, start small with quick wins before scaling, and grow into a platform model.
But for manufacturers, the big question isn’t what agentic AI can do in theory. It’s how to make it work inside plants that weren’t designed for it. At AI Frontiers 2025, an audience member voiced this concern directly:
“What are the most effective strategies for integrating agentic AI into these existing environments to deliver measurable improvements without a full infrastructure overhaul?”
The panel’s answer was pragmatic: don’t start over. Work with what you already have, layer in context, and focus on practical steps that deliver results.
How do you start with existing plant data?
The first step is to free data already trapped in dashboards, HMIs, and siloed databases, then put it into standard formats, clean it enough to be reliable, and apply basic governance so it can be trusted. Without structured inputs, agentic AI cannot deliver dependable results. No new infrastructure is required to begin.
That means putting it into standard formats, cleaning it enough to be reliable, and applying basic governance so it can be trusted. Without structured inputs, agentic AI cannot deliver dependable results.
Why does data need semantic context?
Connecting machines is not enough on its own. To get real value, manufacturers must give their data context using tools such as ontologies, knowledge graphs, or standardized semantic models. These tools create a common language for different systems to communicate. Once data is contextualized, agentic AI can suggest next steps, draft reports, and support decision-making.
Once data is contextualized, traditional analytics can start to reveal patterns, and agentic AI can go further—suggesting next steps, drafting reports, or supporting decision-making.
How do you bridge legacy machines for AI?
Legacy plants run equipment from different decades and vendors on the same production line, and replacing it all is not practical. Interface classes and modular abstractions let very different machines contribute consistent datasets, such as energy consumption or performance metrics, so AI agents can deliver insights without requiring every system to be rebuilt first.
A more effective approach is to use interface classes or modular abstractions that let very different machines contribute consistent datasets—for example, energy consumption or performance metrics. This makes it possible for AI agents to deliver insights without requiring every system to be rebuilt.
What tasks should agentic AI handle first?
Agentic AI is most effective when applied to tedious tasks that engineers and operators dislike: parsing test logs, summarizing maintenance manuals, and processing HMI screenshots. Agents turn this messy material into clean, actionable insights. Engineers still review the output, but instead of spending hours gathering information, they spend minutes validating it.
Here, agents can parse and summarize the material, turning it into clean, actionable insights. This doesn’t replace human expertise—it supports it. Engineers still review the output, but instead of spending hours gathering information, they spend minutes validating it.
How do you scale agentic AI across plants?
The safest adoption strategy starts with quick wins: a contextualized dashboard, an AI-generated downtime report draft, or an agent that classifies common events. Once those wins are established, manufacturers can scale through a platform model, creating shared services across plants, managing agent lifecycles, and building the operational discipline to keep AI reliable at scale.
Once those wins are established, manufacturers can take the next step: scaling through a platform model. This means creating shared services across plants, ensuring lifecycle management of agents, and building the operational discipline to keep AI reliable at scale.
What is the agentic AI roadmap for manufacturers?
The roadmap follows five steps in sequence: free trapped data, add semantic context, bridge legacy machines with modular frameworks, apply AI to repetitive workflows, and scale from small wins to enterprise-wide platforms. Each step builds on the previous one, and manufacturers can begin seeing measurable improvements without waiting for a full infrastructure overhaul.
By freeing trapped data, adding semantic context, bridging legacy machines with modular frameworks, applying AI to repetitive workflows, and scaling from small wins to enterprise-wide platforms, manufacturers can see measurable improvements now—not years from now.
Agentic AI may be new, but it doesn’t require a new factory. It requires a smarter approach to the one you already have.
Special thanks to Vatsal Shah (Litmus), Rajkumar Mylvaganan (ZF Group), Jonathan Wise (CESMII), and Andrew Scheuermann (Arch Systems) for sharing their insights in response to this question during AI Frontiers 2025.
Related articles:
- Managing the Rise of Agentic AI in Manufacturing
- Building the Data Foundations for Agentic AI in Manufacturing
FAQ
1. What is agentic AI in manufacturing?
Agentic AI refers to AI systems that suggest next steps, draft reports, classify events, and support decision-making with limited human input. In manufacturing, as described by panelists at AI Frontiers 2025, these agents parse test logs, summarize maintenance manuals, and process HMI screenshots, turning messy operational data into clean, actionable insights for engineers to validate.
2. How do you integrate agentic AI into legacy plants without a full overhaul?
According to experts from Litmus, ZF Group, CESMII, and Arch Systems at AI Frontiers 2025, the approach has five steps: free data from dashboards and siloed databases, apply semantic models such as ontologies or knowledge graphs, use interface classes to bridge legacy machines, apply agents to repetitive workflows, and scale through a shared-services platform model across plants.
3. What is the difference between connecting machines and contextualizing data?
Connecting machines moves data from one system to another, but contextualizing data gives that data meaning. Tools such as ontologies, knowledge graphs, and standardized semantic models create a common language across different systems. Without this semantic layer, agentic AI cannot reliably interpret the data it receives, and the output of connected systems remains difficult to act on.
4. Why should manufacturers start with small agentic AI use cases?
Starting with small use cases, such as a contextualized dashboard or an AI-generated downtime report, proves value without major disruption. Panelists at AI Frontiers 2025, including representatives from CESMII and Arch Systems, recommended this approach because it builds organizational confidence, identifies integration gaps early, and creates a foundation for scaling through a platform model across multiple plants.
Related Reading
Beyond Predictive Maintenance: How to Protect Your Margins in an Era of Disruption
Beyond Dashboards: Manufacturers Want Predictive Data, Not Pretty Graphs
From APIs to Digital Twins: Warehouse Integration Strategies for Smarter Supply Chains
