Everyone at Money20/20 in Las Vegas wanted the same thing: faster, cheaper, and better, all at once. There were several ways to get there. AI agents were one. Stablecoins and other digital currencies were another. Real-time rails like FedNow were a third.

Every thread raised the same follow-up:

Can we trust it, and how do we reduce the risk?

I spent three days at our booth and in the sessions, and that question came up constantly.

Re-engineering, not just adding agents

Agent vendors have multiplied since last year, and most are aimed at a single step. Wells Fargo’s head of AI products, Faraz Shafiq, said redesigning workflows with AI in mind is his biggest area of focus for 2027, and that the bank wants to move fast with guardrails.

One example stuck with me. A vendor was handling lending. It improved time to yes, but it was essentially a guided journey, with an AI agent sometimes reaching out by phone or text to help.

It wasn’t handling the complex, high-stakes work happening in the background. I couldn’t tell what was running there, because it was never mentioned even when I dug a bit deeper.

Payments: new rails, new rules, same show

In payments, the speed comes from the rails. In early September, 21 banks and asset managers, including Goldman Sachs, Bank of America, and Citi, announced plans for a joint dollar stablecoin company. FedNow handled about 5.0 million payments in the second quarter, up 83% in volume from the first.

Then the rules moved while we were still on the floor. As the show came to a close on Wednesday, the OCC finalized its GENIUS Act rule for payment stablecoin issuers, with requirements for reserves, redemption, custody, and capital. It takes effect January 18.

A change like that doesn’t stay with the team that reads it. It reaches core processes across a bank. That is why orchestration matters. When the rules change, you need to change the process once, see everything it touches from end-to-end, and keep the risk of the change under control.

Fraud: threats evolve fast, and liability is unsettled

Deepfakes, synthetic IDs, and fraud-as-a-service keep getting more sophisticated every day. And bad actors have the same AI engineering tools we do. Fraud defenses that stand still are a huge risk and need to evolve ahead of these threats.

Fortunately, defenses are moving too. Mastercard used the show to launch a risk score for stablecoin payments, and its security chief said trust is what lets innovation scale. The job is changing as well: agents can gather the evidence and make calls at certain severity levels, so analysts spend their time on the truly hard calls to make. NatWest says its Digital Fraud Agents save 21 minutes per case.

The harder question is who is responsible when something goes wrong. If an agent buys something its owner never authorized, who is holding the bag? A Mastercard executive raised it on stage, and a Citi executive noted that the dispute rights a cardholder takes for granted, the kind chargebacks provide, have no stablecoin equivalent yet.

Nobody had a clean answer.

Orchestration: a word that covers a lot of ground

Plenty of vendors claimed orchestration. At one end, a product guides an applicant through a journey and calls it end to end. At the other end, orchestration handles complex, high-stakes work across systems, people, and agents, with every step visible.

That second kind is where asymmetric value is. The hard work gets done faster, with fewer handoffs and fewer surprises. Leaders can see cost and performance as it happens, and change the process when conditions change. A good test: ask to see one loan from first request to last, across everything that touched it, and every decision. If they can’t pull up a single pane of glass to prove the journey, you’re likely seeing a slick demo.

What bank leaders asked us

Four questions came up again and again at our booth.

Visibility: can we see it? They want to see the whole process, end to end, across systems, people, and agents. It didn’t matter if it was a real-time process where AI agents were nearly non-existent. Or a long, arduous process with AI agents working alongside humans as equal players. Visibility mattered.

Flexibility: can we change it? They want to control their own destiny, in strategy and in execution. That means changing processes quickly and using whatever technology makes sense, without being locked into one agent, model, or vendor.

Control: can we keep it in bounds? They want dynamic AI agents and processes kept in bounds. Optimize gives leadership a high-level view of what agents are achieving and costing. Operate goes granular: what an agent did, the logic behind it, and its token usage, tool calls, and loops.

Proof: can we prove it? They want an audit trail for compliance, and proof that agents behave as designed and deliver value. Camunda keeps a record of process and agent decision for the entire journey. Plus it provides teams ways to test how AI agents behave in that process before they put them in production.

Agents earn their place inside the process

Most of the agents I saw still summarize, draft, and explain, with a person still handling most of the work. That is starting to change as technology and risk tolerances evolve.

The bigger gap is where AI agents sit. Too many agents live outside the core process, in a silo. That limits what they can do, hurts results, and adds complexity at a large enterprise or even a small fintech.

Put an agent inside the process, with controls and visibility, and it can move the numbers banks will be judged on at their next board meeting.