Beyond the first AI breakthrough: How Brex makes onboarding decisions in minutes with an evolving agentic team

Flora Zhang has spent nearly a decade in payments, from managing a charge-card portfolio at American Express to fighting fraud and leading AI-driven onboarding operations at Brex.

As Head of Onboarding at Brex, she runs a team unlike the one she joined five years ago: a mix of human experts and AI agents making KYC and underwriting decisions side by side.

Flora shares what it looks like to operate AI in onboarding beyond initial pilots, from setting the right guardrails to raising the bar on data quality.

When AI agents first landed in financial services, they promised to help teams move faster than ever before. But transforming your operations with AI requires real strategy. Teams should pair AI experts with practitioners who deeply understand the workflow end to end.

To do this, Brex built what Flora describes as “a startup within a startup.” It’s a dedicated team that brings engineers and operations experts together to redesign processes with AI at the core.

Across KYC and underwriting, Flora's team deployed AI agents to support decisions that had previously run on deterministic, rule-based logic—leading to massive gains in efficiency and decision quality. Automated decisions flow through agents, while complex cases route to expert analysts.

According to Flora, the review time for a single case dropped from over an hour to under ten minutes. “It was really impressive for me to actually see how big of an improvement we saw from AI.”

But the most interesting thing for Flora isn’t the initial win. It’s learning how to manage and tune the agents’ performance, apply human expertise in critical cases, and build a feedback loop in which human judgment continuously improves the system.

When organizations begin deploying AI, it’s natural to evaluate progress by the first big launch. Flora encourages leaders to look beyond that milestone.

At Brex, for example, the initial working batch of automation came together quickly. Flora estimates a couple months, driven by a visionary mentor, Brex’s former COO Camilla Morais, who was eager to embrace the technology early. But that was only the beginning.

“For AI, people always cheer for how big of a first win it is, which is very true,” Flora says. “But equally important is all the maintenance: keeping the program evolving on itself…and constantly delivering the ideal production result.”

Maintenance becomes more sophisticated in complex areas like customer onboarding, which can involve close to fifty different checks. The Brex team mapped out the optimal sequence of these checks, whether humans or agents should be involved, and how to build the architecture that allows everything to run smoothly. For enterprise use cases, Flora notes, that ongoing cross-functional work is difficult to imagine AI replacing entirely.

Flora’s takeaway echoes a pattern we hear often from leaders deploying AI well: the system is never finished. A first win can earn attention, but durable value comes from treating AI as something you continuously tune and govern.

Flora thinks financial services leaders can also deliver value from AI by building the right boundaries.

The first question to ask is how far to push automation. As Flora puts it, today’s AI models could technically automate decisions well beyond where Brex draws the line. But aggressive automation isn’t always the right approach, particularly in a regulated environment.

Given this, the Brex team intentionally trades automation against its risk appetite. Flora says, “Sometimes I have to mandate manual review on a certain size of limit to make sure that we’re actually really confident with the decision rather than relying fully on AI.”

The second question is around data quality. In Flora’s words, AI is only as good as its inputs, akin to cooking a great dish with the right ingredients.

For instance, human reviewers could smooth over the occasional messy input, but at scale, automated decisions need cleaner, more consistent data. A questionable signal that a human might have worked around, like a business address that uses “St.” versus “Street,” can now ripple across many decisions with AI.

“We’re no longer always going to have the human to catch it,” Flora notes. “That puts a higher bar in terms of how we manage the data, how we manage the ingredients we make the dish with.”

That higher bar, she argues, is actually pushing teams to do valuable work: pay close attention to data quality, backtest logic thoroughly, and monitor automated decisions with rigor.

Above all, Flora cares most about what AI changes for people and teams, and how to build and sustain domain expertise.

Today, Brex can confidently automate routine, lower-stakes decisions because the agents used were built and trained by financial services experts.

The question now is: when those experts move on, what does the next generation of talent actually look like? Or, as Flora puts it, do we have the kind of talent pipeline that can jump directly into the complex end?

It’s a forward-looking perspective that understands human expertise as something to be nurtured, and skill sets as something that must evolve alongside technological innovation.