AI agents aren't replacing customer-facing work. They're replacing its variance.

The hottest job in customer-facing operations is the gap between the best agent and the average one. AI agents are being deployed to close that gap. The companies reporting the largest gains picked routine, repetitive work where the variance was already visible, and the question is no longer whether the technology works. It is which front-office processes in any given operation are shaped like that, and which are shaped like a decision.

Forrester's recent synthesis of customer deployments (Siemens qualifying over 12,000 monthly B2B inbound leads in minutes with a 2% lift on previously ignored opportunities, an unnamed public-sector firm cutting routine invoice-processing labor by 90%, Engine offloading more than half of its travel-management cases) all share the same shape. The work absorbed is high-volume, low-variance, previously under-served. What is replaced is the gap between top and average performance, not the work itself (this reframe is an inference from the pattern of gains across multiple deployments).

That distinction decides everything downstream. A process built on routine variance can be handed to an agent without losing the case. A process built on ambiguous intent, negotiated terms, or escalation loses the case on handoff. The boundary everyone is debating, job replacement, is the wrong one. The boundary that seems to matter, based on which process types showed measurable gains, is variance-shaped versus decision-shaped.