AI-driven value creation occurs where the work takes place

The AI discussion in enterprise IT visibly began at the user interface. Assistants draft texts, summarize information, or help users navigate complex processes. That’s useful. For many SAP users, the bigger strategic challenge lies where orders are fulfilled, goods are moved, equipment is maintained, and malfunctions are resolved. The decisive progress is operational—and thus closer to physical value creation. In the production floor, warehouse, and plant, AI is not just another chat channel. It faces real-world constraints: safety regulations, cycle times, material availability, skill levels, and spatial conditions. A recommendation that sounds plausible in the system must be executable, traceable, and safe on the shop floor. That is precisely where the opportunity lies. AI can combine signals from processes, devices, and workflows in such a way that employees can more quickly identify which decision matters most at that moment.

For SAP managers, the key question is shifting. It’s no longer: “Which function can be enhanced with AI?” but rather: “Which operational decisions yield measurable benefits, and what information is currently missing that would allow us to make those decisions reliably?” ERP remains indispensable in this context. It provides master data, orders, inventory, rules, and transactions. But it is not necessarily the place where every decision is made. Operational reality continuously generates context that is only partially reflected in the traditional process model: the status of a machine, the actual location of a material, a deviation in the workflow, or an employee’s experience. Future-proof architectures connect these levels, rather than treating the production floor merely as a downstream execution zone.

The path leads toward physical AI

When software does more than just analyze and make recommendations—when it interacts with sensor technology, image processing, autonomous systems, or robotics—a digital response becomes an action in the real world. This expands the potential—and the stakes. An error in a text summary is quite different from a misprioritization on a conveyor belt or an inappropriate intervention in a plant process. That is why robotics should not be viewed as a mere extension of automation. It combines perception, decision-making, and movement. The questions here are: Which tasks can systems handle independently? Where is human approval required? How is a safe exception defined? How can a human understand why a system suggests or executes an action?

Governance Takes Shape in This World

Governance begins with responsibilities and does not end with an AI policy. For each use case, the business unit, IT, operations, and—if applicable—security personnel must clarify who is responsible for objectives, data quality, changes, and escalations. Models require a business mandate, boundaries for their use, and criteria by which their results are evaluated. Permissions deserve special attention. The principles that apply to transactions and data access in the ERP world must be extended to AI-driven actions. A system can view information without being allowed to modify it. It can recommend an action without triggering it. It can operate within a narrow framework without having free rein over processes or assets. These distinctions are not a hindrance; they are what make automation responsibly scalable in the first place. The temptation is great to treat AI as a one-size-fits-all platform decision. A different approach is often more successful: start with clearly defined operational scenarios, jointly assess impacts and risks, and feed the insights gained back into the architecture, data model, and governance. This does not create a parallel universe alongside SAP, but rather a robust connection between business processes and their physical execution.

To the partner entry: