Brex

Brex combines corporate cards, banking and spend-management software with AI agents that automate financial work. Its agents serve both customers—handling expenses, policy questions and finance workflows—and internal teams responsible for onboarding, risk and support. Capital One completed its acquisition of Brex on April 7, 2026, in a stock-and-cash transaction; the closing announcement said Pedro Franceschi would continue as CEO.[1][9]

Brex organized its AI work into three pillars: corporate AI, the tools employees use; operational AI, automation of the company's own financial-services processes; and product AI, customer-facing features. In a September 2025 interview, CTO James Reggio owned product AI and COO Camilla Matias owned operational AI, with both collaborating on corporate AI. A shared platform supported the product and operational work.[3]

The platform grew from early LLM experiments in 2023. Reggio recalled January 2023 infrastructure for deploying, versioning and evaluating prompts, routing models, managing data egress, and monitoring costs; in the September 2025 interview, he dated platform experimentation to March 2023. Its Retool interface exposed prompt management, model testing, evaluations and workflow integrations, allowing operations staff to refine prompts and test new models themselves. The September interview credited a roughly 25-person systems engineering team with building and maintaining it. Shared internal and external MCP tools made new product capabilities available to internal agents as well.[1][3]

Brex's product agents address two audiences: finance teams with role-specific work, and employees who use their company's Brex card. For employees, Reggio described an executive-assistant model: an agent with enough context to handle travel booking and expense documentation. His stated goal was to remove the software chores surrounding a purchase: “The best user interface is just the card that lets you swipe and go.”[1][3]

A single assistant with many tools and retrieval-augmented context struggled to cover expense management, travel, procurement and policy questions. Reggio said that overloading it with tools or switching its prompt according to the apparent topic performed less well than specialist subagents behind an orchestrator. Brex built a framework in which agents could hold multi-turn conversations, while product teams could iterate on their own specialists without one team owning every employee action.[1]

Reggio illustrated the mechanism with a dinner-expense question. An employee asks how much they may spend per person; the assistant consults a policy agent, which may need to know whether the dinner is a customer event, a team event or part of travel. The policy agent asks the assistant to obtain that clarification, and the conversation continues. This back-and-forth supplies missing context that a single tool call and response may not resolve.[1]

In the interview published January 17, 2026, Reggio described a TypeScript agent layer using Mastra, pgvector and Pinecone alongside existing Kotlin and Elixir backend code. About half of its agent-layer applications ran on Mastra; the other half used an internal framework focused on multi-agent networks. Mastra offered ergonomics similar to Brex's earlier internal LLM framework, while the custom network framework better suited specialist conversations and their evaluations.[1]

The audit workflow separates detection, judgment and follow-up. An audit agent searches for potential waste, fraud or policy avoidance, favoring fewer missed violations. A review agent decides whether a finding warrants a case, considering its significance and the employee's broader compliance behavior. A case then triggers the employee's assistant to collect a business justification or use context already present in its conversation history. Reggio's illustrative pattern was repeated $74 purchases where receipts become mandatory at $75—a pattern requiring more context than one expense alone.[1]

Brex tested multi-turn behavior with simulated-user agents given objectives to accomplish. Some evaluations began with handwritten conversation turns to isolate particular behaviors. Reggio also described accuracy evaluations that blocked unacceptable regressions and more subjective measures, such as tone and coherence, tracked over time.[1]

Operational automation relied on breaking standard operating procedures into discrete decisions and actions. In September 2025, Reggio described high confidence in automating eight to ten steps of Brex's 14-step know-your-customer workflow, with human review retained. Fraud automation similarly handled high-confidence cases while analysts reviewed the remainder with AI-generated findings. Matias expanded managers' responsibilities to include overseeing agents, improving prompts and providing feedback, alongside managing people and escalations.[3]

One more elaborate approach proved disappointing. Reggio said Brex invested in reinforcement learning for credit decisions and underwriting with an unnamed outside specialist company, but the resulting performance was inferior to a simpler web-research agent. He connected the operational approach to the need for repeatable, auditable procedures: the difficult work was expressing the SOP and unwritten institutional knowledge clearly enough for an agent to follow.[1]

Payment Operations Manager Kyle Martin built dispute automation by putting a 100-plus-page dispute guide and the outsourced team's SOP into Brex's platform, then iterating on prompts and results. Matias reported that the described preparation process fell from three hours to three seconds and produced stronger submissions, including additional arguments drawn from Mastercard rules. Those timings concern preparing dispute submissions, rather than the card network's eventual resolution.[3]

For adverse-media recognition during KYC, Reggio reported an accuracy improvement from 85% to 88%. He emphasized that performance remained below 90%, reflecting the difficulty of deciding which news meets the definition of adverse media. The result was specific to that workflow, rather than a general accuracy measure for Brex's agents.[3]

Brex let employees assemble their own AI tool stack instead of selecting one provider for everyone. Through ConductorOne in Slack, employees could configure tools such as ChatGPT, Claude or Gemini and coding tools including Cursor and Windsurf; access was provisioned through Okta. Reggio described this as both a way to reduce adoption friction and a source of usage data for contract renewals.[3][1]

Operations employees assessed their fluency at four levels—User, Advocate, Builder and Native—supporting their choice with examples and discussing progress at quarterly manager check-ins. Matias assigned projects to help employees build fluency. Engineering changed its interview to a project expecting agentic coding, evaluating how candidates worked and whether they understood the generated code. Reggio said existing engineers and managers also completed the exercise, without retained pass/fail scores.[3][1]

In the interview published in January 2026, Reggio put Brex's scale at roughly 40,000 customers, about 300 engineers and 350 people across engineering, product and design, including roughly ten specialists focused on LLM applications. He said engineering headcount had remained flat during his CTO tenure while the business grew. He also identified a cost of faster code generation: teammates could lose familiarity with rapidly changing services, making code review, maintainability and incident response harder.[1]

Before those agent investments, Brex built its business around startup corporate cards. That earlier product combined payments with expense software—the financial workflows its later agents would address.[6][1]

Founded in 2017 by Henrique Dubugras and Pedro Franceschi, Brex followed the founders' earlier payments company, Pagar.me. In an October 2018 Y Combinator interview, they recalled beginning YC with a virtual-reality concept before pivoting to corporate cards. Their initial name, Veyond, appeared on some early cards. They also recalled being unable to obtain a startup credit card despite $120,000 in YC funding.[9][6]

The 2018 offering provided startup corporate cards without personal guarantees and with built-in expense management. The founders explained that underwriting businesses rather than individuals enabled those terms, and reported 48 employees in October 2018. Brex announced general availability of its Empower platform on October 19, 2022, reporting that it had crossed $3 billion in annualized processing volume within less than three months of going live. By Reggio's later interview, product domains included corporate cards, corporate bank accounts, expense management, travel and accounting.[6][8][1]

Brex cut 282 people, roughly 20% of the company, in January 2024. In his March operating-model letter, Franceschi said Brex had eliminated pure people-manager roles, arguing that managing people could not be divorced from their work. Allen Park's January 2026 account reported that Brex passed $500 million in annualized revenue during 2025; that figure describes annualized revenue, rather than annual recurring revenue.[4][5][1]

Brex announced a $5.15 billion transaction with Capital One on January 22, 2026. Capital One announced completion on April 7, describing Brex's combination of corporate cards, expense automation, real-time payments and AI agents. The closing announcement confirmed a stock-and-cash transaction and Franceschi's continued leadership of Brex.[10][9]