What Did the Human Actually Approve? Binding Approval to AI Agent Actions

An AI agent prepares a $5,000 vendor payment. A finance manager reviews the recipient and amount, then approves it. Hours later, after a retry or workflow regeneration, the agent attempts to submit a payment for $50,000.

The manager approved a payment. But did they approve this payment?

Human-in-the-loop is often implemented as a simple sequence: pause, ask for approval, resume. But when an AI agent regenerates tool calls or a workflow resumes hours later, the action that eventually executes may differ from the one the human reviewed.

In this talk, we’ll explore how to bind approval to a specific operation and validate that authorization at execution time. We’ll examine how to distinguish safe retries from changes that require renewed approval, how to detect changes to critical fields such as the recipient and amount, and how to revalidate preconditions when execution resumes after a delay.

We’ll also examine approval expiry, change detection, and execution-time checks that preserve human intent without unnecessarily interrupting autonomous workflows.