At Plaid, they collect data from over 9,600 financial institutions, creating a massive challenge in classifying ATM deposits, gas station refills, and other transactions into consistent categories. This consistency is crucial for customers like Coinbase, Lyft, or American Express to offer great products to all consumers, regardless of where they bank. The company's system uses machine learning to standardize categorization across thousands of institutions, tackling issues such as transaction categorization, ambiguous merchant names, and varying levels of accuracy across banks. A naive keyword-based approach was initially considered but proved ineffective due to the complexity of the task and the need for constant rule updates. Instead, Plaid employs a combination of machine learning techniques, including word embeddings and neural networks, to learn complex patterns in transaction descriptions and improve categorization accuracy. The system has achieved best-in-class accuracy and coverage despite the evolving nature of the underlying data, and Plaid continues to refine its approach through continuous retraining and experimentation with new model architectures.