AI Agents Are Moving Money Faster Than Compliance Can Follow
Fintech is deploying autonomous agents with guardrails bolted on afterward. That's a recipe for spectacular failure.
I've spent 25 years in mortgage and fintech. I've watched regulators move at glacial speed. I've also watched what happens when you ship financial plumbing with holes in it.
Right now, across lending, crypto, and procurement, AI agents are moving real money with minimal enforcement. And the industry is shipping guardrails as an afterthought, not a prerequisite. That's backwards.
Binance lets AI agents trade, but "keeping them in check is largely up to users." Translation: Binance has abdicated responsibility. Users don't have the instrumentation, the audit trails, or the real-time enforcement to control an autonomous actor moving their capital. That's not a feature. That's negligence wrapped in a press release.
Coupa moved $20M in payments across its agent-powered procurement platform in five weeks with almost no human friction. Twenty million dollars. Five weeks. Think about the failure modes. A logic error in an agent's reasoning loop. A prompt injection. A hallucination about an invoice total. A misconfigured approval threshold. Any of those could become a financial incident—and the first time it happens at scale, we'll all know it went wrong only when accounts payable raises hell.
The mortgage industry is racing to integrate agent frameworks into origination. Origination. That's loan decisions. That's underwriting. That's the gate between consumer and lender. And the speed of deployment vastly outpaces the risk architecture that has to sit underneath it.
Here's what I know works: you don't bolt guardrails on after the fact. You design them in from the start. You define what an agent can and cannot do before you write the first action. You instrument every decision with attribution—which data was used, why the agent chose path A over path B, what human reviewer owns the exception. You set hard limits on transaction size, velocity, and scope. You test failure modes before production. You have an off switch that works in milliseconds, not minutes.
Mortgage has been doing this for two decades. KYC. AML. Anti-steering. Qualified Mortgage rules. Fair lending audits. Fraud scoring with human override. None of it is perfect, but the discipline is there because the cost of getting it wrong is ruinous—regulatory action, massive fines, reputation damage that doesn't recover.
Fintechs and crypto platforms don't have that muscle memory. They're moving fast. They're shipping. And they're hoping compliance catches up.
It won't. Compliance never catches up with velocity. It only catches up with incidents.
The agents aren't the problem. The problem is the people deploying them without the infrastructure that every regulated lender learned to build a long time ago. If you're moving money—whether it's a mortgage, a trade, or a payment—you need to know where it's going, why it's going there, and be able to stop it instantly if something looks wrong.
Ship that first. Then ship the agent. Anything else is gambling with other people's money.
From my toolbox — something I actually ship, not just write about:
vasperamemory-sdk — The TypeScript/JavaScript SDK for VasperaMemory — drop the universal AI memory layer into any app. · ~142/wk on npm