Require human approval before an AI agent makes a payment

Give a real person the approve button on every refund, payout and limit change your agent tries to execute — and keep proof that survives an audit.

The problem

An AI agent that can move money is only as safe as its worst prompt. A hallucinated amount, a manipulated instruction, or a duplicated tool call turns into a real transfer the moment the agent fires it. Blocking every payment defeats the point of automation; letting them all through is how a single bad decision becomes an unrecoverable one. What finance and payments teams actually need is a gate: routine actions flow, high-stakes ones stop for a named human, and every outcome is provable afterwards.

How Raposa solves it

Raposa is a human-in-the-loop approval layer for AI agents. Before the payment executes, your agent makes one HTTP call — request_human_approval() — describing the action. Raposa pauses the agent and routes the request to the authorized approver, who approves or rejects with one click. Only an explicit approval lets the agent continue; silence is never treated as consent. You can require more than one signature — for example two people from a named finance group — and set reminders, escalation to a second approver, and an audited expiry for anything left unanswered.

Every decision is sealed in a verifiable, hash-chained audit trail: who approved, when, and on what grounds — signed, exportable and tamper-evident. HMAC-signed webhooks notify your systems of each outcome. Data stays in the EU, and a DPA is available, so the same control that protects you operationally also stands up to auditors and regulators.

What it looks like

Because the approve button lives in a signed email link, a Telegram card or a Slack message, opening the message never decides anything — only the button does. The approver does not need access to your agent runtime.