Refund approval task
The refund workflow reaches the threshold. A task lands in the supervisor’s queue with the case, the agent’s recommendation, the policy applied. Two clicks, the workflow resumes.
Each task is a human-in-the-loop checkpoint with context, evidence and an SLA. A workflow pauses for a task when judgment is required; a human works it; the workflow resumes with the decision recorded.
A task is not a notification. It is a contract. When a workflow reaches a moment that needs human judgment, it creates a task — with the assignment, the queue, the priority, the SLA and the evidence already attached. The work doesn’t live in someone’s inbox; it lives in the cognitive layer, waiting for a human to commit a decision.
Tasks get created in three ways. Workflows pause for them when a HITL gate is reached. Agents escalate to them when something doesn’t fit their constraints. Operators create them manually from Command Center when they want to route a piece of judgment to a teammate.
Where humans work them: inside Operational Apps (the workspace for the decision), dispatched and tracked from Command Center (the cockpit), and found through Omnisearch (the cross-cutting find). When the human commits, the task closes; the workflow resumes; the evidence chain captures who decided, with what context, applying which rule.
The refund workflow reaches the threshold. A task lands in the supervisor’s queue with the case, the agent’s recommendation, the policy applied. Two clicks, the workflow resumes.
An agent extracts fields from an invoice; a task opens for a human to verify deviations from the standard clauses. The reviewer focuses on judgment; the data entry is already done.
A workflow hits an unexpected case. A task lands in the right team’s queue with the full context, the original input, and the rule that flagged it. The team decides; the workflow continues.