TLDR Tech / Crypto
Building an AI-Native Coinbase Compliance System
Coinbase rebuilt its compliance operations around AI by recognizing that existing procedures encode constraints of the old operating model (fragmented systems, queue pressure, and human working memory limits) rather than optimal investigation logic, so feeding those procedures to a model replicates bottlenecks rather than eliminating them. The resulting four-layer architecture pairs classical ML for alert clustering with a multi-agent investigation system where specialized sub-agents covering alert context, identity signals, transaction behavior, source-of-funds, onchain activity, and adverse media each write into shared case memory, a coordinator agent reconciles disagreements, a narrative agent drafts the final report, and a self-validation guardrail routes uncertain cases to manual review. The system handles ~55% of Coinbase's US fraud case volume with less analyst time per case, and represents the first pod of a multi-month roadmap to rebuild compliance from the ground up.
Read the full story