AI Agents in Enterprise Workflows: What the Case Studies Get Right (And What's Missing for Warehouse Ops)

There's a pattern in enterprise AI coverage that I keep running into: broad, impressive-sounding case studies that stop just short of being useful for anyone running warehouse operations on SAP.

Nasscom recently published a roundup of AI agent deployments across global enterprises. The case-study format is exactly what practitioners need — specific problems, specific outcomes, no vague promises. The global scope is genuinely interesting. But reading through it as someone who spends most of his time in SAP EWM configuration and warehouse rollouts, something is obviously missing: any example that touches physical operations, warehousing, or supply chain execution.

That gap is worth talking about.

What the Case Studies Actually Show

The Nasscom piece covers AI agent deployments across industries — finance, HR, customer service, IT operations. The common thread is that agents are being used to automate multi-step processes that previously needed human coordination: routing requests, synthesising data from multiple systems, triggering downstream actions.

That framing is legitimate. AI agents aren't just chatbots with a better UI. The defining characteristic is that they can act — call an API, update a record, trigger a workflow — not just respond.

For enterprise IT practitioners, this matters because it shifts the question from "can AI answer my question?" to "can AI close the loop without me?"

That's a meaningful shift. But the examples stay firmly in knowledge work territory.

Where Warehouse Operations Fit — And Where They Don't Yet

Physical warehouse execution is a different problem. In SAP EWM, a lot of what looks like "automation" is already happening through deterministic logic: PPF (Post Processing Framework) actions, warehouse task creation rules, wave management, queue assignment in the RF framework. The system is already making decisions — just rigid, rule-based ones.

The honest question is: where would an AI agent actually add value on top of that?

A few places come to mind:

Exception handling. When a handling unit scan fails, or a transfer order can't be confirmed because of a quantity discrepancy, someone usually has to make a judgment call. That's exactly the kind of multi-step, context-dependent decision AI agents are being trained to handle in other domains.
Labour and resource allocation. Labour management in EWM generates data. Deciding what to do with that data — whether to reassign a worker, adjust a wave, escalate a bottleneck — is still largely manual. An agent that could observe patterns and suggest (or trigger) adjustments would be genuinely useful.
Cross-system coordination. EWM doesn't operate in isolation. It talks to TM (Transportation Management), ERP, sometimes MES. A lot of operational friction lives in the handoffs between these systems. Agents that could monitor and act across system boundaries are a natural fit — in theory.

Why It's Not Happening Yet (Or Isn't Visible)

Part of the reason warehouse-specific AI agent case studies don't show up in pieces like the Nasscom roundup is probably that the implementations that exist are proprietary and early-stage. Nobody rolls out an AI agent that touches live goods receipt flows and then writes it up immediately.

Another part is that SAP EWM environments are genuinely complex to integrate with. The EWM data model, the way warehouse orders and warehouse tasks relate, the role of PPF, the RF interaction layer — these aren't plug-and-play for an AI agent framework built by a team that's never seen a warehouse floor. The learning curve is steep and the risk of a bad action (wrong bin, wrong HU, wrong movement type) is immediate and physical.

That's not an argument against it. It's an argument for why practitioners in this space need to be the ones pushing it forward, not waiting for generic enterprise AI vendors to figure it out.

What I'm Taking From This

Reading case studies like this one is useful even when the examples don't match your domain. The underlying patterns — agents that observe, decide, and act across system boundaries — translate. The warehouse execution layer is just a harder, more consequential environment to apply them in.

Three things worth sitting with:

Deterministic EWM logic covers a lot, but exception handling is still a human tax. That's where AI agents have the clearest case.
The integration complexity is real. Anyone claiming a quick AI agent layer over SAP EWM hasn't thought through the data model.
The case studies will come. They're just not public yet.

If you're working in this space and have seen implementations I haven't — or have questions about where EWM's existing automation actually ends — I'd be interested in the conversation. More of my thinking on EWM and modern tooling is at markus-mildner.com.