Deep learning expands risk capability when signals are sequential, relational, or unstructured-conditions increasingly common in large e-commerce platforms. This chapter explains when neural architectures justify their complexity and how they complement strong tabular baselines. It introduces embeddings for high-cardinality identifiers, sequence models that learn temporal patterns from behavioral telemetry, and multitask learning setups aligned with platform objectives. The chapter emphasizes that deployment feasibility determines value: Feature freshness, batching and latency, monitoring for silent failures, and robustness to label noise and drift are central design constraints. It also discusses practical integration with existing decision stacks, including how deep models can provide representations consumed by lighter-weight scorers and policies. Rather than presenting deep learning as a universal replacement, the chapter positions it as a strategic tool when representation learning unlocks signal beyond manual feature engineering, and when the organization can support the engineering, monitoring, and governance required to keep high-capacity models trustworthy over time. We discuss how to evaluate deep models under delayed outcomes, how to avoid overfitting to short-lived patterns, and how to keep explanations and monitoring credible when representations evolve. The chapter closes with criteria for deciding when deep learning is warranted, and when simpler models yield better operational return.