Coinbase deployed a fine-tuned AI agent that reduced fraudulent onramp transactions by 30% and fraud value by 22%, outperforming closed-source models.
AI & Agents ·
Coinbase has deployed a fine-tuned AI model built on open-weight architecture to detect fraud in its Onramp payment service, achieving results that exceed those of closed-source competitors. An experimental rollout demonstrated a 30% reduction in fraudulent transactions and a 22% decrease in fraud value, with the system operating at lower latency and cost than frontier models. The company adapted Qwen3.5-9B using its own transaction data, arguing that task-specific models can outperform general-purpose systems when trained on domain-specific information.
The AI agent reviews transaction sequences to spot patterns that single-payment analysis might miss, assigning risk scores but unable to override blocks already in place from earlier checks. In a comparative benchmark across 16,140 transactions, newer versions of closed-source models including GPT and Anthropic variants showed mixed results—GPT-5.6 improved precision but detected fewer fraudulent cases overall, while Opus and Sonnet versions performed worse on multiple measures than their predecessors.
The deployment reflects broader industry momentum toward AI agents in financial-crime detection, though questions persist around autonomous decision-making in high-stakes environments. Coinbase's implementation incorporates safeguards including defined decision limits and application codes that govern agent actions, yet debate continues about oversight and accountability in increasingly autonomous systems.