Brex

## How does Brex use AI to categorize expenses and detect fraud?

## Overview Brex employs a multi-layered artificial intelligence (AI) and machine learning (ML) system to automate the categorization of expenses and to detect potentially fraudulent transactions. This system is designed to enhance operational efficiency, enforce compliance with company policies, and improve security. The AI architecture integrates advanced models, including Large Language Models (LLMs), with a human-in-the-loop review process to ensure both speed and accuracy. For expense categorization, the system automatically identifies the merchant and classifies the transaction into one of 48 business-specific categories, with a claimed accuracy rate of 95%. This process has evolved from using traditional ML models like Random Forest to leveraging LLMs for real-time, dynamic auto-population of expense fields such as category, memo, and attendees. ## Key Features Brex's AI architecture is composed of several specialized 'agents' that work together to streamline the expense management process. The 'Brex Assistant' helps employees with tasks like writing memos and filing reimbursement claims. The 'Audit Agent' continuously monitors all expenses against the company's internal policies and categorizes any violations by their level of risk. The 'Review Agent' automates the approval of low-risk expenses that are clearly within policy, while escalating any exceptions or suspicious activities for human review by the finance team. This division of labor allows the AI to handle the high volume of routine transactions, freeing up finance professionals to focus on strategic oversight and exception management. The entire system is designed to learn and adapt over time through a continuous feedback loop, where user corrections and feedback help to refine the models' precision. ## Technical Specifications The AI models analyze a rich variety of data sources to make their predictions. These sources include merchant information, historical company transaction data, detailed Level 3 (L3) payment data which contains line-item details from receipts, and data synchronized from a company's Enterprise Resource Planning (ERP) and Human Resources Information System (HRIS) tools. By embedding a company's specific expense policies and Standard Operating Procedures (SOPs) into the transaction analysis, the AI can automatically enforce rules and flag potential violations. To maintain high data quality, the system uses probability thresholds; if a prediction does not meet a desired level of precision, the transaction is routed to a human review queue. This human-in-the-loop process may involve an internal support team or crowdsourced services like Amazon Mechanical Turk to manually classify the expense, ensuring that the system's data remains clean and reliable. ## How It Works For fraud and anomaly detection, Brex utilizes proprietary machine learning models that operate 24/7 to provide real-time monitoring. This system is designed to identify unusual spending patterns and sophisticated fraud attempts that might be missed by traditional, static rules-based systems. The detection models analyze a wide range of variables, including transaction amounts, merchant locations, Merchant Category Codes (MCCs), and the velocity of transactions. For example, the system can flag unauthorized purchases in high-risk categories or transactions that deviate significantly from an employee's normal spending behavior. This proactive monitoring helps businesses mitigate risk and prevent financial losses. ## Limitations and Requirements Brex places a strong emphasis on security and privacy in its AI systems. The platform uses industry-standard security measures, including AES-256 bit encryption for data at rest and TLS 1.2 or better for data in transit. It also supports security features like two-factor authentication (2FA), mobile biometrics, and Single Sign-On (SSO) with major identity providers. The use of customer data for these processes is governed by a Platform Agreement, a Privacy Policy, and a Data Processing Addendum (DPA), which outlines the use of proprietary fraud and risk modeling. While the system is highly automated, Brex acknowledges that it is not infallible and that occasional manual correction by finance teams is necessary, reinforcing the importance of the hybrid AI-plus-human approach. ## Summary In conclusion, Brex's use of AI is central to its expense management and fraud detection capabilities. By combining LLMs, proprietary ML models, and a robust human-in-the-loop review process, the platform automates a significant portion of the manual work associated with expense reporting and compliance. The system of specialized AI agents helps to enforce policies in real-time and escalate exceptions for human judgment. While this technology provides significant efficiency gains and security enhancements, it is designed to augment, not entirely replace, the critical oversight role of a company's finance team.