Autonomous payout reviews at scale: How Deriv implements agentic guardrails

Deriv built Opra, a payout-review agent that autonomously evaluates withdrawal requests in real time, resolving 96.3% of routine cases without human intervention while routing ambiguous transactions to payment officers with full context. The system operates under four guardrails: scoped authority with earned autonomy, a human-in-the-loop fallback, idempotency and transactional safety, and an immutable, reconstructable audit log for regulatory compliance.

We’ve built a decision layer at Deriv that determines how quickly a payout can proceed and when it needs a person to review it first. For an online broker processing high-volume daily withdrawals worldwide, every payout a client requests is a small promise: getting their funds back into their hands quickly and safely. Keeping that promise sounds simple. At scale, it’s one of fintech's hardest balancing acts: speed for the client on one side, diligence to keep everyone safe on the other. For a long time, we balanced speed and diligence by relying on people to do much of the manual work. So we built something to change the problem's shape. We set out to help our Payment Officers review withdrawals faster. But the more we automated, the clearer it became: the real opportunity was creating a system that evaluates every request, resolves routine transactions autonomously, and escalates to a person only when human judgment is strictly required. That system is Opra . Here’s how it thinks, and why that’s different from simply automating a checklist. How Opra works: Reasoning beyond scripted If-Else rules Opra goes far beyond following a fixed list of scripted steps. It’s a payout-review agent that watches for withdrawal requests in real time. The moment one appears, it gathers everything relevant into a single view and starts reasoning about it, weighing the pattern the request fits, its risk profile, and whether similar requests have previously been trusted to resolve on their own. From there, Opra makes a decision. If a request fits a pattern that Opra is trusted to handle end-to-end, it acts independently, without waiting for a person. If it doesn’t—or if it only partially matches— Opra doesn’t guess or force it through. Instead, it routes the case to the team best placed to make the judgment, recording the decision instead of leaving it buried in a thread. Whether Opra decides or a person does, the outcome returns to the same coordinated workflow. That’s the real shift: in every case, Opra first decides whether it’s a decision it should be making at all, rather than simply executing a predefined set of rules. Opra by the numbers - 96.3% autonomous resolution rate: Over nine out of ten routine requests are verified and approved without manual friction. - Zero blind approvals: 100% of ambiguous transactions are flagged and routed to payment officers with full context preserved. - Full auditability: Every decision—autonomous or manual—is appended to an immutable, reconstructable audit trail. Four core guardrails for financial AI agents An agent making real decisions about real funds only works if the boundaries around it are as deliberate as the agent itself. These are some of the guardrails we built in from day one: - Scoped authority and earned autonomy: Opra acts autonomously only within explicitly defined and regularly reviewed limits. Nothing is handled autonomously by default; it must first earn the trust to operate within those boundaries. - Human-in-the-loop HITL fallback: Anything ambiguous, unusual, or outside those limits is routed to a person, never pushed through automatically. - Idempotency and transactional safety: Actions that can’t be safely undone receive additional safeguards, preventing issues such as duplicate submissions. - Immutable audit logging and reconstructability: Every action, whether taken by Opra or a person, is written to an append-only, immutable audit log, ensuring the complete decision trail is always reconstructable for regulatory compliance. The result is an agent that’s given genuine autonomy—but only as much as it has earned, and never without a complete record of every decision it makes. Expanding autonomous operations safely We’re continuing to expand what Opra can take on, guided by the same priority that shaped it from day one: getting clients their funds quickly while keeping the guardrails firmly in place. Key takeaway: Autonomy must be earned, not assumed The goal hasn’t changed since the first line of code: every client should get their funds back quickly and safely, and every Payment Officer should be able to focus their attention where it genuinely matters—not on manual legwork, and not on double-checking an agent that hasn’t yet earned the trust to operate independently. Opra lets us deliver that at scale. If you’re already building agents that need to earn autonomy rather than assume it by default, I’d love to hear how you’re approaching it in the comments. Benedicta Moudy Wiyono https://www.linkedin.com/in/moudywiyono/?lipi=urn%3Ali%3Apage%3Ad flagship3 pulse read%3BfTCp0U4cRE%2Ba8JPM3pKfPg%3D%3D is an AI Engineer at Deriv. Follow our official LinkedIn page https://www.linkedin.com/company/derivdotcom/posts/ and our tech account on X https://x.com/derivdottech for more updates.