Automation works best where the work is high-volume, low-judgment, and well-instrumented. It struggles where context is ambiguous and the cost of a wrong call is high.
The clearest wins we have seen recently: invoice and document processing, customer-support triage, internal search and onboarding copilots, and code-assisted operations. The clearest disappointments: agentic systems that try to take 12 steps when the org doesn’t even understand the first one.
Measure. Every automation project should ship with an evaluation harness, accuracy, latency, cost, and a business KPI. If you cannot measure it, do not deploy it.