Agentic AI in Manufacturing: Best Practices for Quality, Speed, and Safety
Answering the audience’s questions
At the AI Frontiers 2025 event, one audience member posed a question that goes to the heart of modern manufacturing challenges:
“When introducing agentic AI into domain-specific workflows, what best practices help manufacturers improve product quality and decision-making speed while minimizing operational risk, especially in highly regulated or safety-critical industries?”
A panel of industry leaders—including Andrew Scheuermann, CEO of Arch Systems; Rajkumar Mylvaganan, Platform Product Manager and Head of Digital Manufacturing Platform Core at ZF Group; Jonathan Wise, Chief Technology Architect at CESMII; and Vatsal Shah, Founder & CEO of Litmus—took on this question. Their answers cut through the hype and offered practical, experience-based guidance. Here’s how they broke it down.
Building on solid foundations: the four layers
Rajkumar Mylvaganan explained that manufacturers need to think in layers. At the base are connectivity and data collection (PLCs, SCADA, MES). Above that comes data management and contextualization—using ontologies and knowledge graphs so that raw signals make sense. Only then can you add AI/analytics, and finally, agentic AI: the layer that allows systems to plan, reason, and act alongside humans.
Across all four layers runs a non-negotiable thread: governance, security, and observability. In safety-critical environments, every AI decision must be explainable, trackable, and reviewable.
Cleaning up the data mess
Several panelists pointed out that bad data remains the biggest bottleneck. Many plants are still drowning in siloed, low-quality information.
Vatsal Shah of Litmus stressed that AI cannot “magically” clean dirty data. Manufacturers must liberate and structure what they already have—using common formats like JSON, adopting standard protocols such as OPC UA, and applying ontologies that describe assets consistently. Without this groundwork, agentic AI agents are set up to fail.
Where to deploy agentic AI first
According to Andrew Scheuermann of Arch Systems, the best early use cases are where humans are already overstretched:
• Root cause analysis (RCA/CAPA): Agents can scan logs, dashboards, and manuals (even across languages) to draft a structured report, leaving the operator to verify and approve.
• Downtime classification: AI can label events quickly, reducing the repetitive work engineers dislike.
• Unstructured data processing: From HMIs to test logs, agents can turn previously unusable artifacts into usable insights.
In short: focus on dull, repetitive, or skipped tasks where AI improves accuracy rather than introducing risk.
Human in the loop: the ultimate safety net
Reliability came up repeatedly. The consensus: in regulated industries, agentic AI must serve as a co-pilot, not an autopilot.
Andrew Scheuermann described their approach as giving the operator a “pre-written report” generated by AI. Humans validate it before action is taken. This keeps accountability clear while still speeding up workflows. Rajkumar reinforced the point: evaluation datasets are essential to test agents’ reliability and throw out outputs that don’t meet quality thresholds.
Tackling interoperability and legacy systems
Jonathan Wise of CESMII emphasized that brownfield environments are the rule, not the exception. Machines of different vintages, formats, and vendors all speak different “dialects.”
The solution isn’t to hope LLMs figure it out—it’s to build consistency through shared semantic models and interface classes. Only with standardized interpretations can AI deliver safe, repeatable results. Otherwise, you’re just moving the problem “up a layer.”
From pilots to sustainable value
The panel urged manufacturers not to get stuck in pilot purgatory. Quick wins are valuable for building confidence, but long-term success requires a platform approach: shared services, scalable data layers, and lifecycle management of AI agents.
Change management is equally critical. Operators should be involved from the start, both to build trust and to help shape how agents learn and reason. Leadership buy-in is essential, too.
Takeaway: cautious optimism with a roadmap
Agentic AI is already helping manufacturers boost quality and accelerate decisions—but only when introduced carefully, with strong governance and operator involvement.
Essential Building Blocks:
• Standardized data and semantics
• Layered architecture with governance at every level
• Use cases targeting dull, repetitive, or skipped tasks
• Human-in-the-loop for safety and compliance
• Evaluation sets to measure and control reliability
• A platform strategy for scaling across plants
In safety-critical industries, the payoff is significant: faster, more consistent decision-making and improved product quality, delivered without sacrificing trust or control.
Related articles:
GenAI in Manufacturing: Start with Data Quality, Governance, and Quick Wins
Answering the AI Frontiers Audience’s Question “Generative AI isn’t just about documents. Factories generate huge amounts of sensor data, but turning that into real value with AI is challenging. What should companies tackle first—improving data quality, linking older systems, training teams,
AI on the Shop Floor: Turning Predictive Maintenance into Real Results
For manufacturers, predictive maintenance powered by IIoT and AI is no longer just about advanced analytics in the cloud. The real breakthrough comes when frontline workers can act on those insights in real time. At the IIoT World Manufacturing Day
Agentic AI Isn’t a Product – It’s an Integrated Business Strategy
Everyone’s talking about AI. From pilot projects and co-pilots to promising dashboards, artificial intelligence is cropping up left and right in manufacturing. What’s rare, however, are wide-scale implementations that deliver measurable results on the factory floor. In fact, IDC estimates