Amazon Payments bandit lifts funnel conversion high single digits

The post frames this work as the second half of a two-part problem. An earlier post covered using generative AI on Amazon Bedrock to produce personalized content at scale within brand guidelines. Once that production constraint is removed, the harder question becomes selection: among many candidate variations, which one to show each visitor, and how quickly can the system learn that. Amazon Payments chose a contextual bandit because it conditions on behavioral signals rather than hand-defined segments, so a pattern learned in one context transfers to similar visits without needing separate traffic for each group.

The design choice the team emphasizes most is going multi-objective. The customer journey has three stages — application start, submission, and approval — and optimizing one in isolation can degrade another, a trade-off they call the seesaw problem. Running one LinUCB model per stage and combining their UCB scores with roughly equal weights was the only formulation they found that kept estimates non-negative across all three stages simultaneously. The batch architecture, a weekly SageMaker AI Processing job reading from Amazon S3 and publishing recommendations to a key-value store, fits the delayed nature of approval feedback; approvals lag by days and are held back until a later cycle.

The A/B test produced a split verdict: a high single-digit percentage relative lift on the final stage for one population, and no improvement, with a statistically significant approval regression, for the other. The team's reading is that the model explored broadly and still found no winning arm, which points at the content pool as the binding constraint rather than the algorithm. The practical stakes therefore sit with whoever owns the content pipeline. If a modest set of vetted images and taglines is too narrow, the bandit can only confirm that no combination beats the static page. The approach may pay off most for teams that can widen and refresh that pool, for instance with generative pipelines, while keeping the existing experience as an arm so the system cannot regress far.