A hard rejection is a blunt instrument. An agent reaches a limit, the operation fails, and the business is left to work out why. Escalation is a better answer. When an operation crosses a threshold, the system should not simply refuse it — it should turn it into a request a human can approve, with everything needed to decide attached. Done well, this keeps flow alive without weakening the control.

From rejection to request

The difference is small in design and large in effect. A rejection tells an agent "no" and stops. An approval request tells a human "this operation exceeded the mandate — here is the invoice and the counterparty". The operation pauses at a decision point rather than dying at a wall. The limit still holds; the work still has a path forward.

What good escalation looks like

Escalation works best when the context arrives with the request:

  • the operation itself, with amount and counterparty;
  • the supporting documents, such as the invoice;
  • the limit or threshold that was crossed;
  • the mandate the agent was acting under.

A human with these facts can approve, decline or amend in seconds. A human without them has to go and reconstruct the situation, which defeats the purpose. The reviewer should be able to act on the request alone, without opening a second system or chasing a colleague for context.

Why this keeps flow alive

In agentic commerce, operations arrive continuously. A system that rejects anything unusual will stall often; a system that escalates will stall only where judgement is genuinely required. That is the point: autonomy handles the routine, and thresholds route the consequential to people. Every escalation is also a record — a trace of where the boundary was and who decided.

In short

  • Escalation beats rejection for operations that cross a threshold.
  • Requests should carry the invoice, counterparty and the limit crossed.
  • Human judgement is reserved for consequential operations.
  • Each escalation becomes part of the audit trail.

See how escalation fits agent controls on the agent layer.