Manufacturers have spent years connecting machines, applications and business processes. The next challenge is not simply collecting more data, but using that information to make manufacturing operations more responsive and intelligent.
Artificial intelligence is opening new possibilities. AI agents could investigate production deviations, coordinate workflows and interact with enterprise applications. At the same time, advances in robotics are creating opportunities to connect digital intelligence with physical operations.
But introducing these capabilities raises an important question:
How can manufacturers enable collaboration between people, AI agents and intelligent machines without compromising operational reliability, security or control?
The answer lies not only in the technology itself, but in how it is integrated into existing manufacturing processes and systems.
Traditional manufacturing systems largely operate through predefined processes. Applications exchange information, workflows follow established rules, and people make decisions based on the available data.
AI agents introduce a more dynamic approach.
Instead of simply executing a predefined workflow, an agent could identify a problem, retrieve relevant information, select appropriate tools and recommend or initiate actions within defined boundaries.
For example, an agent might detect a production deviation, investigate historical data and suggest corrective measures. Another could support production scheduling by considering material availability, machine capacity and current priorities.
This creates opportunities for more connected and responsive manufacturing operations. However, it also introduces new dependencies.
An AI agent is only as reliable as the information, tools and permissions available to it.
In an SAP-based manufacturing environment, that information may come from SAP S/4HANA, SAP Digital Manufacturing, planning applications or shop-floor systems.
A well-integrated landscape with standardized processes, governed data and clear system responsibilities therefore provides an important foundation for AI.
Rather than replacing existing manufacturing architecture, AI builds on it.
One development that could support this evolution is the Model Context Protocol (MCP).
MCP provides a standardized way for AI applications to connect with external tools and information sources.
In manufacturing, MCP-based connections could potentially allow agents to interact with enterprise capabilities such as:
The important difference from conventional integrations is how these capabilities may be used.
While traditional applications typically follow predefined interactions, an AI agent can potentially select different tools depending on the task it needs to complete.
However, MCP does not replace existing integration platforms, APIs or security mechanisms. It provides a way to make tools and information accessible to AI applications, while access control and operational safeguards must still be implemented.
For manufacturers, this introduces practical questions:
These questions become particularly important when agents move beyond retrieving information and begin interacting with operational processes.
AI agents should operate within the enterprise architecture, not bypass it.
Consider a production environment where a machine begins producing parts that deviate from the required quality specifications.
In a conventional process, an operator might detect the deviation, a quality engineer would investigate possible causes, maintenance would inspect the equipment, and a production planner would assess the impact on the schedule.
An AI-enabled process could support these activities in a more coordinated way.
A manufacturing agent could detect the anomaly using production and quality data. It could then retrieve historical information, examine relevant machine parameters and check the current production order in connected systems such as SAP Digital Manufacturing and SAP S/4HANA.
Based on this information, the agent could identify possible causes and suggest corrective actions. Depending on its permissions, it might also initiate an approved quality workflow or request a maintenance inspection.
The responsible operator or engineer would remain involved in evaluating and approving decisions affecting production.
In a more advanced environment, an appropriately designed robotic system could potentially perform certain physical inspections or adjustments.
This illustrates the potential of human–agent–machine collaboration:
The goal is not to remove people from manufacturing processes, but to help them make better-informed decisions and respond more effectively to operational challenges.
Not every manufacturing process requires the same degree of AI autonomy.
The transition toward AI-enabled operations can be understood through several levels of capability:
1. Digital processes: Business processes are standardized and supported by connected applications.
2. Automated processes: Rules and predefined workflows automate repetitive activities and predictable decisions.
3. AI-assisted processes: AI analyzes information, detects patterns and recommends actions to support human decisions.
4. Agentic processes: AI agents can select tools and execute defined sequences of actions within established permissions and approval requirements.
5. Physical AI: Intelligent physical systems can perceive their environment and perform actions within appropriate operational and safety constraints.
These levels are not necessarily a fixed implementation sequence. Different capabilities may coexist depending on the manufacturing process and its requirements.
For manufacturers, the objective should be to identify where additional intelligence creates measurable value and determine how much autonomy is appropriate.
As manufacturing environments become more intelligent, organizations need to manage more than applications, integrations and conventional business processes.
AI introduces additional elements such as models, agents, tools, data sources, permissions and monitoring requirements.
This makes enterprise architecture and lifecycle management increasingly important.
The SAP Integrated Toolchain provides capabilities that can support this foundation:
These capabilities can help manufacturers maintain transparency across the systems and processes into which AI is introduced.
However, AI-specific governance is also required to manage agent behavior, tool permissions and operational boundaries.
Organizations need to understand not only which applications support a manufacturing process, but also which AI agents can interact with those applications and what actions they are authorized to perform.
This creates a new governance question:
Which AI agent is allowed to do what, in which process, using which data and tools?
Answering it requires coordination between enterprise architecture, business process management, security and AI governance.
Traditional automated testing typically verifies whether a system produces the expected result under defined conditions.
AI-enabled processes introduce additional complexity because an agent's behavior may vary depending on the available information, instructions and tools.
For example, an AI agent may reach the same objective through different sequences of actions or encounter unexpected situations that were not explicitly anticipated during development.
Manufacturers therefore need to test more than whether an agent can successfully complete a task.
They also need to evaluate:
Consider an agent that can create or modify production orders.
Testing should verify not only whether the agent can perform the transaction, but also whether it refuses or escalates the action when required authorization or business conditions are missing.
Existing approaches such as process-based testing, automated regression testing and centralized test management in SAP Cloud ALM can help validate the underlying processes and integrations.
Additional AI-specific evaluations are needed to assess agent behavior, permissions and failure scenarios.
In manufacturing, this is particularly important because errors may affect production availability, product quality and delivery performance.
Testing therefore becomes both a quality assurance measure and an important part of AI governance.
AI-enabled manufacturing is not limited to software applications.
Advances in robotics, sensors and computer vision are creating new possibilities for intelligent physical systems that can interpret their surroundings and perform tasks.
In the longer term, humanoid robots may become relevant for selected industrial environments, particularly where workplaces have been designed around human movement and interaction.
However, physical AI introduces additional risks.
While software agents interact primarily with digital systems, robots can directly affect equipment, materials and people.
Increasing autonomy must therefore be accompanied by appropriate safety concepts, cybersecurity controls, operational boundaries and human oversight.
Safety-critical functions should remain governed by suitable industrial control and safety mechanisms rather than relying solely on AI-generated decisions.
The objective is to allow intelligent systems to contribute within controlled operating conditions, not to give them unrestricted authority over physical manufacturing processes.
As AI moves closer to the shop floor, the integration of IT, operational technology and safety governance becomes increasingly important.
AI agents, MCP-based tool connections and physical AI could significantly change how manufacturing operations are managed.
However, the introduction of these technologies does not make existing digital manufacturing strategies obsolete.
Integrated systems, standardized processes, reliable data and transparent architecture remain essential. What changes is how intelligence can interact with these foundations.
Manufacturers do not necessarily need to begin with complex autonomous scenarios. A practical starting point is to identify a clearly defined business process, establish reliable access to the required information and determine which activities can be safely supported by AI.
From there, new capabilities can be tested and introduced systematically.
The goal is not to make manufacturing autonomous for its own sake. It is to make operations more responsive, transparent and adaptable while maintaining reliability, safety and human accountability.
The manufacturers best prepared for this development will not necessarily be those with the largest number of AI experiments, but those with a digital foundation that allows new capabilities to be introduced responsibly and at scale.