Brex’s AI Hail Mary — With CTO James Reggio (acquired for $5B by Capital One!)
Latent Space PodcastKey Moments
Brex CTO James Reggio discusses their three-pillar AI strategy, operational AI, and multi-agent systems for financial institution efficiency and innovation.
Key Insights
Brex employs a three-pillar AI strategy: Corporate AI (adopting AI tools), Operational AI (reducing operational costs), and Product AI (creating AI-powered customer features).
The company fosters a "Quitters Welcome" culture, celebrating employees who leave to become founders, viewing them as valuable assets who bring unique perspectives.
Brex's engineering team is structured around product domains, with a dedicated AI center of excellence focusing on LLM applications and agentic systems.
Operational AI has seen significant impact, automating tasks like customer onboarding, underwriting, and fraud detection through SOP-driven agents rather than complex RL.
Brex prioritizes building its own infrastructure and frameworks for AI, especially for multi-agent systems and agent-to-agent communication, viewing it as a competitive advantage.
The company aims to train employees in AI fluency across all departments, not just engineering, encouraging experimentation and rewarding novel AI use cases.
BREX'S THREE-PILLAR AI STRATEGY
Brex has implemented a structured AI strategy comprising three key pillars. The Corporate AI pillar focuses on adopting and procuring AI tools across all business functions to significantly enhance workflow efficiency. The Operational AI pillar is dedicated to building and acquiring solutions aimed at reducing the company's operational costs, a crucial aspect for a financial institution. Finally, the Product AI pillar involves developing new features that position Brex as an integral part of their customers' own corporate AI strategies, thereby enabling Brex to be a valuable solution for other businesses.
TRANSITIONING FROM MOBILE ENGINEERING TO CTO: JOURNEY OF LEADERSHIP
James Reggio's path to CTO at Brex, originating from a mobile engineering background, is noted as a less common trajectory. He highlights that his experience as a multiple-time founder was more instrumental in his appointment than his technical specialization. This founder mentality, emphasizing leadership and business acumen, proved crucial for the CTO role, which extends beyond purely technical responsibilities. This perspective also informs Brex's approach to hiring, particularly their "Quitters Welcome" initiative, which embraces individuals who may eventually become founders themselves.
ENGINEERING STRUCTURE AND THE AI CENTER OF EXCELLENCE
Brex's engineering organization, encompassing around 300-350 individuals, is primarily structured around product domains like corporate cards, banking, expense management, and accounting. A distinct AI center of excellence, comprising about 10 people, operates somewhat separately to focus intensely on LLM applications and agentic systems. This dedicated team was formed by envisioning a company with AI at its core, built from scratch today, which then informed their internal development approach. This structure allows for concentrated innovation in AI while integrating its benefits across the broader engineering department.
OPERATIONAL AI: SOP-DRIVEN AGENTS AND AUTOMATION
Operational AI at Brex has yielded immediate and significant business impact, particularly in automating core processes. Initially exploring Reinforcement Learning for credit decisions, the team found that simpler web research agents outperformed it. This emphasis on breaking down problems into granular, SOP-driven steps, which are auditable and repeatable, translates effectively to LLMs. Key applications include automating customer onboarding, KYC processes, fraud detection, and dispute resolution, significantly reducing human intervention and operational costs while maintaining compliance and customer experience quality. This approach allows operations staff to shift from execution to prompt engineering and AI evaluation.
BUILDING AND MANAGING AI PLATFORMS AND AGENT FRAMEWORKS
Brex prioritizes building its own AI infrastructure, including an LLM gateway for prompt management, data egress, and model routing. For agentic systems, they leverage Mastra for certain workflows but have developed their own internal framework for multi-agent networks, especially for orchestrating complex employee-facing assistants. This custom solution allows for more natural agent-to-agent communication and better management of tasks spanning multiple product lines. The strategy involves not picking winners among foundational model providers but offering employees choice, fostering experimentation and providing valuable usage data for contract renewals.
THE EVOLVING CRAFT OF ENGINEERING AND AI FLUENCY
The advent of AI and agentic coding is reshaping the engineering landscape. Brex is adapting its interview process to assess candidates' ability to use and understand AI-generated code. Internally, they encourage AI fluency across all departments, not just engineering, through training programs and spotlights on innovative AI use cases. While AI accelerates code generation, a key challenge is maintaining code quality and reducing 'slop.' Brex focuses on maturing its AI usage, balancing efficiency gains with robust code reviews and architectural rigor, ensuring that AI amplifies 'good' outcomes without solely relying on AI to fix AI-introduced problems.
EVALUATION STRATEGIES AND MITIGATING HALLUCINATIONS
Evaluating AI applications involves different approaches for operational and product AI. For operational AI, evaluations are integrated into the platform, co-developed by subject matter experts and engineers, with mistakes feeding back as regression tests. Product AI, particularly multi-agent systems, is more challenging to evaluate. Brex uses simulated user agents for multi-turn conversations and LLM-judged assessments, sometimes incorporating handwritten preambles to isolate specific behaviors. Addressing hallucinations, especially in the Brex assistant, is critical. This involves grounding agents with accurate product documentation and implementing system prompts and circuit breakers to prevent unsupported claims or actions.
THE FUTURE OF HEADCOUNT AND AI'S IMPACT ON ROLES
The conversation around AI's impact on headcount, particularly the junior vs. senior mix, remains nuanced. Brex has achieved significant business growth without increasing engineering headcount by focusing on efficiency gains through AI and improved execution strategies. The CTO emphasizes that AI amplifies both positive and negative aspects of development, leading to more complex capacity planning than a simple headcount reduction. The company's approach is to enhance existing engineers' leverage rather than seeing AI as a direct replacement for personnel, focusing on evolving roles and skills to meet future demands.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
Common Questions
Brex's AI strategy is built on three pillars: Corporate AI (adopting AI tools across the business), Operational AI (building solutions to lower operational costs for financial institutions), and Product AI (introducing new features to integrate Brex into customers' corporate AI strategies).
Topics
Mentioned in this video
A code review tool Brex uses and highly recommends for its high signal-to-noise ratio and ability to catch numerous issues.
An AI technique Brex initially invested in for credit decisions but found inferior to simpler web research agents.
Experienced engineers at Brex who pair with younger AI natives to leverage their deep understanding of code bases and products.
Payment Card Industry Data Security Standard, implicitly relevant to operational AI in finance, ensuring compliant processes.
Part of the flexible tool choices offered to Brex employees via Conductor, allowing them to select development environments.
A framework used by Brex for accelerating development on their agent layer, chosen for its ergonomics and similarities to their internal LLM framework.
The practice of using AI agents to assist in the coding process, a key focus for Brex's engineering team.
The platform used for building the UI/UX of Brex's internal AI tools, including prompt and tool managers, making them accessible to the ops team.
The company where the guest, Jio, is the CTO. The discussion centers around Brex's AI strategy, engineering, and operations.
Another company where the CTO, Jio, gained experience prior to Brex.
The future of finance that Brex aims to build, focusing on AI-driven automation and intelligent workflows.
Brex's approach to building complex AI systems, involving multiple agents communicating and collaborating to achieve objectives.
The third AI strategy pillar at Brex, focused on developing new features that enable customers to integrate Brex into their own corporate AI strategies.
Mentioned as a previous employer of Brex's CTO, Jio, contributing to his diverse experience.
Part of Brex's technology stack, used in conjunction with Pinecone for specific AI/ML functionalities.
A firm with which Brex published a piece discussing their AI strategy and operational AI investments.
A key pillar at Brex aimed at leveraging AI to reduce operational costs for financial institutions through custom solutions.
Auditable and repeatable processes that are key for operations and translate well to LLM automation.
A role within Brex's engineering department, identified as a significant user of Cursor, indicating leadership adoption of AI tools.
One of Brex's three main AI pillars, focusing on adopting and integrating AI tooling across the entire business to enhance workflows.
A model developed by Brex to categorize and train employees on their AI capabilities, from user to native levels.
Experienced engineers who are expected to spend more time on high-level tasks like teaching AI and developing complex patterns, rather than just coding.
Less experienced engineers who can be taught and guided by senior engineers and AI models.
Internal infrastructure built by Brex to deploy, manage, version, and evaluate prompts, handling data egress and model routing.
Know Your Customer, a process automated by Brex's AI agents for evaluating customer applications.
Brex's criteria for businesses they serve, with a focus on commercial segments and specific revenue/transaction thresholds.
Specialized engineers focused on developing AI solutions, a key hiring group for Brex, often including former or future founders.
Mentioned as a tool where non-engineering employees showcase their AI agent building skills.
A metric Brex uses to track the extent to which AI is involved in code commits.
A backend programming language used at Brex, contrasted with TypeScript for their newer agent codebase.
A critical process in software development, which is being impacted by AI-generated code and requires increased human attention and rigor.
Refers to the availability and reliability of systems, a key consideration in operational AI.
More from Latent Space
View all 61 summaries
86 minNVIDIA's AI Engineers: Brev, Dynamo and Agent Inference at Planetary Scale and "Speed of Light"
72 minCursor's Third Era: Cloud Agents — ft. Sam Whitmore, Jonas Nelle, Cursor
77 minWhy Every Agent Needs a Box — Aaron Levie, Box
42 min⚡️ Polsia: Solo Founder Tiny Team from 0 to 1m ARR in 1 month & the future of Self-Running Companies
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Try Summify free