Human approval for AI agent payments

Human-in-the-loop is most useful for exceptions rather than every purchase. The goal is bounded autonomy.

The practical model

• Routine low-risk spend can stay autonomous.
• New vendors can trigger review.
• Large transactions can trigger review.
• Rejected requests remain visible in the ledger.
• Approval and settlement should be tracked separately.

How Alem approaches it

In Alem, a policy can require approval above a monetary threshold, for new vendors or for subscriptions. An approval is an authorization decision; it does not by itself move money.

What is live today

Alem's sandbox includes company workspaces, agent registry, budgets, spend policies, human approvals, economic ledger, ROI, HTTP/OpenAPI and MCP. Live payment execution remains private beta / integration-stage.