AI Agents in the Warehouse: Why "Analyze and Report" Is No Longer Enough

There's a gap between what AI promises in manufacturing and what it actually does in most warehouses today — and it's not a technology gap, it's an architecture gap.

A recent Nasscom piece on manufacturing AI agents made the distinction clearly: the next step isn't smarter analytics, it's agents that can act. That framing hit close to home. Because in EWM rollouts, I keep running into exactly this ceiling — systems that flag problems brilliantly and then wait for a human to do something about it.

The Dashboard Problem

Most warehouses I've worked in have more visibility than ever. WM dashboards, EWM monitors, exception queues — the data is there. What's missing is the closed loop.

An AI tool that tells you "putaway efficiency dropped 12% in the last two hours" is useful. An agent that detects the drop, identifies the bottleneck (say, a blocked warehouse task queue in a specific storage section), and either reroutes work assignments or escalates with context — that's a different category of tool entirely.

The Nasscom framing calls this the shift from analytical AI to agentic AI. In warehouse terms: from monitoring to orchestrating.

What EWM Already Does (And Where It Stops)

To be fair, SAP EWM isn't standing still. The warehouse order creation rules, activity area sequencing, and labor management integration already give you a lot of levers for automated orchestration. The system can prioritize tasks, consolidate picks, and route RF users dynamically based on workload and proximity.

But those rules are static. They were configured during the project, signed off, and they run the same way every day. The warehouse changes — throughput spikes, staffing gaps, equipment goes down — and the rules don't adapt. Someone has to go back into the configuration and adjust, or create a workaround via ABAP or a custom PPF action.

That's the gap agentic AI targets: real-time adaptation without manual intervention.

Where Cobot and Robotics Integration Gets Complicated

The Nasscom article focuses partly on robotics and cobot integration, which is where things get genuinely complex in practice. Cobots on the floor don't fail in clean, predictable ways. A sensor miscalibrates. A pallet type changes. A pick location is blocked by a late GR.

In an EWM context, you're often dealing with handling unit management at a granular level — the robot needs to know the HU structure, the weight tolerance, the destination storage bin. If any of that is off, the task fails. Today, that failure usually surfaces as an exception that parks in a queue until someone clears it.

An agentic system would detect the failure condition, assess whether it can reroute the task to a manual RF user or a different robotic unit, update the warehouse order accordingly, and log the exception with enough context for root cause analysis later. That's not science fiction — the components to build it exist. The integration work to wire them together in a real EWM environment is just not trivial.

What "Acting" Actually Requires in an EWM Architecture

For an AI agent to act inside an EWM landscape, it needs more than read access to data. It needs:

Reliable APIs or BAPIs to create, modify, or cancel warehouse tasks and warehouse orders without breaking downstream processes
Event triggers — ideally via SAP Event Mesh or a middleware layer — so the agent is reacting in near real-time, not polling
Guardrails — clear rules about what the agent can and cannot do autonomously (rerouting a pick is low risk; cancelling a transfer order mid-execution is not)
Audit trail — everything the agent does needs to be traceable, both for operations and for compliance in regulated industries

The last point is underrated. In automotive warehouse environments, traceability isn't optional. An agent that acts but doesn't document why it acted creates more problems than it solves.

The Honest State of Things

I'll say it plainly: most EWM environments I've seen aren't ready for agentic AI today. Not because the technology doesn't exist, but because the data quality, integration architecture, and organizational readiness aren't there yet.

What I think is worth doing now is designing with this direction in mind. When you're configuring warehouse order creation rules or building a custom RF transaction, ask whether the logic you're encoding could eventually be handed to an agent. Keep the rules externalized. Keep the APIs clean. Don't bake everything into hard-coded ABAP.

The warehouses that will absorb agentic AI fastest are the ones that already have disciplined EWM configurations and clean integration patterns — not the ones that jumped straight to a robotics POC.

If you're thinking through how agentic patterns could fit into an EWM architecture you're working on, I'd be interested in the conversation. What I'm seeing and building around this is also showing up on markus-mildner.com.