This week, we’re exploring how Brex is rebuilding its operations org around AI — changing roles, workflows and the skills needed to succeed.

Agent + Human Ops: How Brex is Changing Roles and Workflows with AI 

Many companies think about applying AI to their existing workflows. But this feels reminiscent of Canadian philosopher Marshall McLuhan’s now-famous quote: “We look at the present through a rear-view mirror; we march backwards into the future.”

Brex is taking a different approach.

“If I was starting the company today in this new AI era, I’d build operations completely different,” says Camilla Matias, Brex’s COO. “We can’t just use ChatGPT. The roles must be different. We’re changing what we expect from people.”

This thinking sent Matias and CTO James Reggio down a path to reimagine roles, workflows and skill requirements within the operations team. But this isn’t a light switch-flicking problem, it’s one that requires fundamental change to not only the work being done, but also the tools and processes enabling that work.

Brex’s journey has been full of learnings — ones that companies also trying to build their own AI-centered workflows can apply:

• Treat internal operations with external rigor. Brex created an internal agent builder using underlying components from its external product. Employees can build, test, refine and deploy AI agents with the same rigor as customer-facing products. This keeps experiments safe, scalable and connected, while also creating a product loop where these components get better simultaneously.
• Create pathways for legal and procurement to approve tools. Instead of endless RFPs, Brex pre-approved categories of AI tools (based on data handling, not vendor identity) and automated provisioning through Slack.
• Roles changed to agent management, not task management. BPOs, with early defined SOPs, are now largely automated by agents. Managers need to analyze workflows, improve prompts and design human + agent systems.
• Rebuild workflows around what AI does best. Workflows requiring subjective judgement (surprisingly), understanding complex rules and following clearly laid-out steps at scale were perfect for AI.
• Hire more generalists, not specialists. Because knowledge is more accessible, in the hiring process, Brex screens for AI fluency. They’ve also created an internal program where employees receive AI training and rank themselves on one-of-four fluency levels, which are used in performance reviews.
• The goal shouldn’t be 100% automation. Reggio says that when you set the goal at 100% automation instead of 40% automation, you end up making investments that oftentimes yield zero value because they have a binary outcome.

Reggio leaves us with one key insight: “Don’t overthink it. If you can break down a workflow into an SOP that you can describe to a human, then an LLM can probably perform it pretty well.”

Thanks, as always, for reading and sharing!

-The Review Editors