Amazon Payments used a contextual bandit to select personalised content across a three-stage customer acquisition funnel, AWS says. In a seven-week online test, one customer population saw a high single-digit relative lift in final-funnel conversion, while another saw no improvement.
AWS AI Watch analysis
What happened
AWS describes a system that learns which content works for different visitors, using behavioural signals rather than assigning each person to a fixed segment. It used Linear UCB, a method that balances trying less-tested options with choosing those that currently look strongest.
The funnel had three stages: application start, submission and approval. AWS combined a separate model score for each stage, using roughly equal weights. That matters because content that attracts more starts may not bring more suitable applicants through to approval. The company says the system uses an opaque customer identifier only to route recommendations, not as a model input.
In the seven-week online test, AWS reports a high single-digit relative lift in final-funnel conversion for one customer population, but no improvement for another. It also says approval outcomes can take days to arrive, so the system updates start and submission signals immediately and processes approval feedback in a later batch.
The AWS account of the approach includes a code example and a repository for testing the method on synthetic data.
Why it matters
The useful point is not simply that a bandit can personalise content. It is that the system optimises across several stages instead of chasing an easy early metric that might undermine later approvals. AWS’s results also show why a single success story would be too neat: the reported lift applied to one population, while another saw no gain.
The approach keeps some exploration running as new options arrive, rather than waiting for a conventional test to finish. AWS says its content variations were built from reviewed components, with generative AI assisting production; the individual components were vetted before being combined.
Our read
This is a concrete account of AI-driven personalisation with a more useful scoreboard than clicks alone. The strongest lesson is to measure the whole journey and be willing to report when a result does not travel across customer groups. The reported lift is promising, but it is AWS’s account of its own test, not a guarantee that the same method will work for other businesses.
What to watch
- Whether AWS shares more detail about the customer populations and the test’s conversion measurements.
- Whether the reported lift holds across later tests and additional content options.
- How businesses balance exploration with the need to keep unsuitable offers out of the funnel. Activity teaser Amazon Payments used a contextual bandit to personalise content across application starts, submissions and approvals. AWS says a seven-week test found a high single-digit relative lift in final-funnel conversion for one customer population, but no improvement for another. The interesting choice was to optimise the whole funnel rather than chase starts alone. Should businesses judge personalisation by early engagement, or by outcomes further down the line?
Discussion spark: Should businesses judge AI personalisation by early engagement, or by outcomes further down the funnel even when those outcomes take longer to measure?
Sources and evidence
- Uplifting conversion across the acquisition funnel with personalization using contextual bandits on AWS (1 October 2026, 16:51 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)