Exception-shaped agents: what the surviving AI projects have in common

The forecasts for AI agents currently point in opposite directions. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027. Klarna's AI assistant took over the workload of around 700 support agents — and the company later hired humans again. At the same time, a growing share of enterprises report at least one agent in production. Both are true: many projects fail, and some are running. We looked at what has actually shipped in procurement and logistics — and at what the survivors have in common.

What the surviving projects have in common

The working examples — the exception agents at DHL, Microsoft's Procurement Agent for Dynamics 365, the case studies from manufacturers — share one shape: narrow scope, built in exception form. Not an agent that "does procurement." An agent that handles one defined exception: an unconfirmed purchase order, a failed IDoc, an invoice that doesn't match its PO.

The pattern behind it is the same every time: the agent detects, diagnoses, drafts — a human decides. It watches the exception queue, establishes the cause, proposes the next step or words the message to the supplier. Sending, posting, changing is done by a person.

Why this shape works

Exceptions are the part of the work that cannot be standardized — which is exactly where a large share of manual time sits. At the same time, the task is narrow enough to verify reliably: a diagnosis is right or wrong, a draft fits or it doesn't. And because the agent writes nothing to the system, the risk stays bounded — no wrong posting run, no accidental order. The projects that fail usually start with the opposite ambition: a general assistant that should "handle everything in procurement," with an open task space and open system access.

What this means for an SAP estate

In SAP landscapes the exception queues already exist: the IDoc errors in WE02, unconfirmed purchase orders, blocked deliveries, invoice mismatches. Diagnosis needs no write access — an agent can work on the exports your users already pull every day. That makes this shape work on ECC 6.0 exactly as it does on S/4HANA, without BTP and without touching the running system.

This is the shape we build at aplexio: read-only, draft-based, on the estate you have. The SAP Logistics Copilot starts with the IDoc queue — a worked example is in Anatomy of an IDoc analysis. If you want to check which exception queue in your landscape should go first: book a 30-minute fit call.