Uplifting conversion across the acquisition funnel with personalization using contextual bandits on AWS

This article describes how Amazon Payments applied contextual multi-armed bandits to personalize content across a product acquisition funnel using Amazon SageMaker AI.

• Multi-armed bandits balance exploration and exploitation to identify best-performing content variations in real-time
• Linear UCB (LinUCB) algorithm conditions decisions on customer context vectors for personalized recommendations
• Multi-objective approach optimizes all three funnel stages (start, submit, approve) simultaneously to avoid degrading downstream metrics
• Batch architecture with weekly SageMaker Processing jobs updates model and publishes recommendations to DynamoDB
• Seven-week A/B test showed high single-digit relative lift for one population; content quality, not algorithm, was the limiting factor
• Building blocks are vetted individually then combined into large arm pools, enabling oversight at scale

Contextual bandits provide an efficient optimization layer for personalized content selection when generative AI expands the content pool, enabling continuous learning without waiting for A/B tests to conclude.