Ferdinand Dabitz (Augustus): "The expensive thing in banking is the people"

The Big Whale: What's the difference between an agentic workflow and regular AI?

Ferdinand : Where you want to go with AI inside a financial institution is to make processes either faster or less error-prone, and to take over the manual work humans used to do, executing it faster and with fewer mistakes. That frees the humans to focus on reviewing, approving, and making the decisions. The agentic part is the system carrying out the multi-step work itself, rather than a human clicking through every stage of it.

"The average American bank spends over fifty percent of its revenue on opex."

You frame AI as a margin story before a productivity story. Why?

Look at the ten largest industries in the world by revenue. You've got oil and gas, you've got auto, you've got banking. Banking is the one with the highest opex ratio. It spends the most of all of them on people and manual labour. The other industries are cost-of-goods-driven: the expensive thing in auto is the car parts, the expensive thing in oil is the crude. In banking, the expensive thing is the people. The average American bank spends over fifty percent of its revenue on opex. That's the addressable cost base, and it's why deploying AI here can turn banking into a far more profitable kind of business.

Which workflows get automated first?

There's a lot of opportunity in transaction monitoring. Take BSA/AML alert triage: the AI surfaces the candidate alerts, then a human compliance officer investigates and decides. You let the machines take over the busy work in the monitoring workflow, and the humans focus on the investigations and the calls. Same pattern for sanctions and OFAC match adjudication, where the machine flags and the human clears. And KYC document review, where the AI extracts the fields and a human verifies them.

And what shouldn't be automated?

The decision-making. Especially in a regulated institution, the human has to stay in the driver's seat. That's a question of the safety and soundness of the bank, and of compliance. Data collection and administration around the decision, that's where automation already played a big role and will play a much bigger one. The decision itself stays with a person.

"Fedwire clears payments twenty-two hours a day. What about the other two?"

You argue legacy banks can't simply bolt AI on. Why not?

What holds legacy institutions back is often the legacy technology infrastructure itself. Old bank cores place a real constraint on embedding AI into operations in a safe and sound manner. If you want an agent to take over part of a transaction-monitoring workflow, the agent needs the data infrastructure to talk to the bank core, pull the right pieces of data, and contextualise them correctly. If you run on one of the old cores, an FIS or a Jack Henry, it gets very hard for an agent to talk to the core and get what it needs. If you've built the core from scratch around agentic primitives and agent interactions, it actually becomes possible. That's what we've spent the last years doing.

Describe Augustus in plain terms.

We're in the clearing business. We clear dollar and euro payments for international financial institutions: think a fintech in Latin America, a bank in Southeast Asia, or other global players. We give them dollar and euro accounts, and dollar and euro rates. Where we outperform legacy clearing banks is on settlement speed and availability. Instead of T+2, we can settle same-day or faster, and the bank is available around the clock. The technology platform is what lets those back-office processes run that fast.

Where do stablecoins fit?

A big part of why legacy banks are slow isn't even the rail, it's the processes and the operations, and that's where most of the opportunity to improve sits. But part of it is the rail. Fedwire clears payments twenty-two hours a day. What about the other two? That's where you deploy stablecoins, as a default, instant, 24/7 global rail to fill the gaps wherever there's a real use-case need. By default we run on the traditional fiat rails; after hours, stablecoins cover what Fedwire can't.

"We hope for it to be the case at one point in the future. But today, it isn't."

The word "non-deterministic" comes up a lot around Augustus. What does it mean for a finance reader?

Pre-AI, automation in banks was constrained to deterministic workflows. You had a rule: if this happens, yes; if that happens, no. Decision trees and rules. AI lets you automate even when there's no preset rule or tree, and it captures the fuzzier situations far better. That's where agents really outshine old-school workflow automation, in the ability to handle non-deterministic, fuzzy cases that were never previously defined as a rule. And the important thing to say is that the decisioning always stays with the human. The AI can prepare and administer; in the end, the human makes the call.

The Big Whale: You're around fifty people. How does that change who you hire?

Two profiles. On one hand we're continuously hiring top-tier technical talent, AI-native, high-quality engineering and product people. On the other, we keep investing in bank talent, executives with decades of experience in regulated banking who have done this before. People like that are a crucial part of the proposition, because we can bring those two worlds together.

A fresh OCC charter is extremely rare. How did you get it?

It was a tremendous effort, and it shouldn't be easy. A couple of things mattered. First, the team: Greg Quarles joined early as president of the bank, twenty years at the OCC, then CEO of several US banks, Benjamin Alexander on the compliance side. The team was a really important pillar.

This week, two of the biggest AI adopters, Microsoft and Uber, found their AI bills running past what they'd pay humans. Nvidia said the same about one of its teams. Is Augustus one of those companies where compute costs more than salaries?

No, that's not the case for us. We hope for it to be the case at one point in the future. But today, it isn't.

Last one. What's your advice to a head of AI pushing this inside a big institution?

A good mental model is to focus on tasks, not roles. Sometimes you fall into the wrong frame: "I've got a risk analyst, I want to automate that role away." That's the wrong model. The better one is to look at it on a task level. For each person on your team, what are the individual jobs to be done, and which of them are cumbersome, slow, and require no human judgment or decisioning? You pick those and automate them one by one. It shifts the framing, because it actually lets your people spend their time on the tasks where their judgment is best used, the ones where they have to make decisions, and less time on the busy work a machine probably does better anyway.