An AI agent receives an expense request. Which conditions should trigger human approval?

When an AI agent evaluates an expense request, should approval depend only on the amount, or also on risk, policy exceptions, employee role, and business context? Consider scenarios such as unusually high expenses, missing documentation, policy violations, low-confidence decisions, or requests involving sensitive categories. This discussion explores how organisations can define approval thresholds and escalation rules while allowing low-risk expenses to be processed automatically. The goal is to design AI workflows that balance automation, control, and accountability.