Agentic AI on MES: The Executive Playbook for the Next Era of Manufacturing Leadership

Manufacturing execution systems (MES) were built to bring control, traceability, and discipline to production. They digitized work instructions, managed quality, tracked genealogy, and connected operations to enterprise systems. In ISA-95 terms, MES became the operational bridge between business planning and manufacturing control. But that model is no longer enough.

The next wave is not more dashboards, alerts, or copilots layered onto workflows. It is agentic AI: systems that can interpret goals, reason over context, plan multi-step actions, and act within defined constraints.

This changes the role of MES entirely. The opportunity is not to improve MES as a system of record. It is to evolve it into a system of coordinated action across frontline operations. Not autonomy in the sense of removing humans, but agentic execution—where software can detect, decide, and act across quality, maintenance, scheduling, compliance, and operator support. This is not a feature upgrade. It is an operating model shift.

Most factories do not suffer from a lack of data. They suffer from latency between signal and action. A deviation occurs and someone investigates. A supervisor interprets, a planner reschedules, a quality engineer opens a CAPA and a technician is dispatched. This chain is still too slow, too manual, and too fragmented.

Agentic AI compresses that latency. It connects detection directly to coordinated response, turning disconnected workflows into closed-loop execution systems. That is where the real value lies. At the same time, the broader manufacturing landscape is already moving in this direction. NIST’s smart manufacturing work highlights that advances in computing, connectivity, and data are creating major opportunities to improve productivity—but only if they are applied in ways that are trusted, interoperable, and grounded in operational reality.

The challenge is that most organizations are not there yet. OECD research shows that while AI has strong potential in manufacturing, particularly in process optimization, efficiency, and resilience, adoption remains uneven and fragmented across companies and sectors. That gap matters.

It means many manufacturers are investing in AI use cases without changing how execution actually works. They are adding intelligence at the edges while leaving the core operating model untouched. This results in incremental improvement, not transformation. Executives should focus instead on the operating system of execution itself—how decisions are made, coordinated, and acted on in real time. This is why the conversation must move beyond AI as a productivity tool and toward AI as an execution model.

Most transformation programs still operate in waves: digitize a process, deploy a workflow, stabilize it, and move on. That model no longer holds. Manufacturing today is defined by volatility, labor constraints, product complexity, and pressure for resilience. Static systems cannot keep up. What leaders need instead is continuous transformation.

This means building environments where workflows, decisions, and improvement loops evolve constantly—not through yearly projects, but as part of daily operations. This is where agentic AI fits naturally. Unlike rule-based systems, agentic systems can respond to changing conditions, synthesize context across systems and choose among multiple actions.

The future MES stack should not be judged only by how well it standardizes today’s process, but by how well it enables adaptation tomorrow. That is the difference between digitization and continuous transformation

The phrase that matters here is composable agentic architecture.

Composable means the system is modular rather than monolithic. NIST has explicitly linked smart manufacturing progress to service-oriented approaches and to standards that support composability, interoperability, and correct behavior across distributed manufacturing services. The historical models remain relevant because they provide the structural logic for how enterprise and operational systems relate, but agentic AI requires more than layer definitions. It requires a runtime model for coordinated action.

The defining concept is composable agentic architecture. Composable systems are modular, not monolithic. They can be assembled, reconfigured, and extended without breaking the whole. This direction is not new—but it is now becoming essential.

NIST’s work on service-oriented architectures for smart manufacturing makes this explicit: the next generation of manufacturing systems will be built from loosely coupled, on-demand services that can be composed dynamically, rather than fixed applications deployed as monoliths. These service-based systems enable greater flexibility, interoperability, and scalability, allowing manufacturers to integrate data, intelligence, and capabilities as needed across operations.

Agentic AI builds directly on this foundation, but adds something critical: a runtime model for coordinated decision-making and action. It is not enough to expose services. The system must be able to decide how and when to use them. That is where agentic architecture comes in.

In practice, a composable agentic MES operates across five layers:

1. Context layer
   Unifies machine states, operator inputs, quality data, work orders, maintenance history, and enterprise constraints
2. Reasoning layer
   Interprets goals and evaluates options using planning, memory, and tool access
3. Agent layer
   Specialized agents for domains such as scheduling, quality, maintenance, and documentation
4. Orchestration layer
   Manages coordination, conflict resolution, escalation, and policy enforcement
5. Execution layer
   Takes action, updating records, triggering workflows, assigning tasks, and interacting with connected systems

This layered model reflects the evolution of modern manufacturing research. Multi-agent system architectures are increasingly distributing decision-making across specialised agents, such as product and resource agents, with each agent responsible for different parts of the production system. These agents can plan, request actions and coordinate with one another in order to influence the performance of the entire system.

The advantage is not just decentralization. It is adaptability. Because agents share context and operate within a coordinated system, they can respond dynamically to changing conditions rather than relying on fixed workflows. This is especially important in environments where centralized logic struggles with scale, variability, and real-time constraints.

For example: A quality deviation is detected → A quality agent evaluates severity → A scheduling agent assesses production impact → The orchestration layer determines containment strategy → The system holds material, updates instructions, and assigns corrective actions

All within defined guardrails. That is operational AI in practice. The leadership takeaway is simple: If your architecture is still built around tightly coupled applications, it will limit how far AI can go. If it is built as a composable system of services coordinated by agents, it creates the foundation for continuous, adaptive execution.

The right starting points are not the most advanced use cases. They are the most operationally constrained decisions. These are decisions that happen frequently, cut across functions, are time-sensitive, and become expensive when delayed. This is where coordination, not prediction alone, drives value.

Research increasingly supports this shift. Work on multi-agent maintenance scheduling shows that decisions cannot be optimized in isolation. Maintenance timing must balance asset health, production demand, and system-wide constraints. Distributed agents can generate local recommendations, but those recommendations only become valuable when they are coordinated through a system that evaluates reliability, cost, and operational impact together.

At the same time, broader manufacturing research shows that AI delivers the most value when it connects maintenance, production scheduling, and quality into a single decision loop rather than optimizing each function independently. This is the key insight. The highest ROI use cases are not isolated predictions. They are coordinated operational decisions.

Two areas illustrate this clearly.

Maintenance is one. The shift is from predictive alerts to coordinated action. AI-driven maintenance already enables real-time monitoring and failure prediction, moving beyond rigid schedules toward data-driven intervention. But prediction alone does not solve the problem. The real value comes when maintenance decisions are aligned with production constraints, labor availability, and system-level tradeoffs. Multi-agent approaches demonstrate how distributed decision-making, combined with system-level coordination, can improve both reliability and throughput at the same time.

Quality is another. The opportunity is not better detection alone. It is faster and more connected response. AI is increasingly used for defect detection and real-time monitoring, but its broader impact comes when quality signals are linked directly to scheduling, maintenance, and operational decisions. A deviation should not trigger a standalone workflow. It should trigger coordinated action across the operation, including containment, schedule adjustment, instruction updates, and corrective execution.

These domains sit at the center of MES value. They are where coordination matters most and where delays are most costly. The takeaway for executives is straightforward. Do not start with isolated AI use cases. Start with decisions that require synchronization across the system. That is where agentic architecture delivers outsized returns and where it begins to reshape how the factory actually operates.

This is where many initiatives will fail. Agentic systems do not just generate insights. They take action. That changes the risk profile entirely.

In physical operations, failures cannot simply be patched after the fact. When AI is coordinating labor, moving material, or influencing production, the consequences are immediate. As the World Economic Forum puts it, once AI becomes embodied in physical systems, the limiting factor is no longer what the technology can do, but how responsibility, authority, and intervention are governed.

This is why governance is not overhead. It is infrastructure. As agentic systems scale, risk scales with them. Capabilities can now be deployed, updated, and distributed faster than organizations can adapt. Without strong governance, operational risk grows faster than control systems. This is exactly the problem frameworks like NIST’s AI Risk Management Framework are designed to address. The AI RMF provides a structured approach to managing risk across the full lifecycle of AI systems, built around continuous processes to govern, map, measure, and manage AI behavior in context.

The key idea is that governance is not a one-time control. It is an ongoing system. For manufacturing, that translates directly into how agentic MES must be designed and operated. Leaders must define which decisions can be automated, where human approval is required, what safe states look like, and how monitoring and intervention work in real time. These are not policy questions alone. They are system design decisions.

An operational system should include clear permission structures, traceable actions, policy constraints, approval thresholds, and explicit stop conditions. It should continuously monitor behavior, measure outcomes, and adapt controls as conditions change. This is the shift.

Governance is no longer something you layer on after deployment. It must be built into the execution system itself, just as reliability, safety, and quality controls are built into physical production systems. If you cannot explain what the system was allowed to do, what it did, and why, you do not have an operational system. You have a demo.

This is not a labor replacement story. Framing it that way is a mistake. Manufacturing has always depended on frontline expertise, and that does not change. In fact, it becomes more important. Research from the World Economic Forum reinforces this point clearly. Fully automated operations are rarely feasible or even optimal. Human workers remain a critical differentiator in manufacturing performance, especially as systems become more complex and technology adoption accelerates.

At the same time, frontline workers are often the least involved in how new technologies are introduced, even though they are the primary users. That gap creates friction, slows adoption, and limits value. The research shows that successful technology deployment depends on engaging workers directly, incorporating their feedback, and designing systems that are human-centric from the start. This aligns directly with the shift toward agentic systems. In an agentic environment, the role of the human does not disappear. It moves up the stack.

Operators no longer spend most of their time executing predefined tasks. They supervise, intervene, and guide system behavior. Engineers define workflows, policies, and exceptions. Managers move from chasing status to managing performance systems. The system executes. The human orchestrates. This is also why leading organizations are beginning to treat frontline workers as knowledge workers whose expertise is amplified by better tools and contextual information, rather than replaced by them.

The implication is important. Technology adoption is no longer just a technical challenge. It is an organizational one. If systems are introduced without worker involvement, they will underperform. If they are designed to augment human judgment and incorporate frontline insight, they become significantly more effective and durable over time. That is the real workforce shift. Not fewer humans in the loop, but better use of human attention.

There is real hype in the market and real risk. Gartner estimates that more than 40% of agentic AI projects will be scrapped by 2027, largely due to unclear business value, rising costs, and weak risk controls. That is not a failure of the technology. It is a failure of approach.

Most companies are still treating agentic AI as an add-on to existing systems or as a collection of isolated experiments. They are chasing capability without redesigning how execution actually works. That is why they will fail. The companies that succeed will start from a different question. Not “How do we add AI to MES?” But “Which operational decisions should become semi-autonomous, under what constraints, and on what architecture?”

That is the shift. Agentic AI is not about making MES smarter. It is about making execution faster, coordinated, and adaptive by design. The companies that lead will not treat this as a feature roadmap. They will treat it as a new operating system for the factory. That is where the gap will open.