Finance / Payments
Adyen: Optimising authorisation, fraud and fees at the same time
- Company
- Adyen
- Country
- Netherlands
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Checked a source excerpt through a search tool
The work problem
A single payment has several stages tangled together. The payment experience given to the customer depends on predicted risk, that risk is affected by the authentication result and the choice of payment route, and the route in turn changes the cost. In the past a separate model ran at each stage. The fraud model blocked a transaction when chargeback probability crossed a threshold, and the authentication model decided the route without knowing that.
Technology and data
Uplift's AI ties together several machine learning models of different kinds so they share each other's judgements. The models exchange conditions through message passing and are optimised with Reinforcement Learning towards a single objective that balances fraud, cost and conversion together. Inputs are supplied by the Feature Platform. Slow, complex features are computed with Spark and fast features with Apache Flink, and stored in Cassandra across several regional data centres. The inference service Alfred keeps only one model as principal and promotes new models through ghost and challenger stages. A ghost does not affect the result and only leaves records, while a challenger affects the result on a set share of traffic; when it is statistically confirmed to be better than the principal, they swap places.
Results
Adyen said it processed 670 million transactions over the four days of Black Friday and Cyber Monday in 2024. During this period every transaction passed through two to five AI endpoints, with a median latency budget of 20ms at each point. It said applying Weak Supervision in production raised recall by 22 percent, cut authorisation loss by 46 percent and delivered a 13 percent improvement in issuer decline rate. With Off-Policy Evaluation it said it saved 20 weeks a year previously lost to AB tests that produced nothing, and estimated it gained an additional 9 million to 54 million transactions over half a year.
Limits and open questions
This article is Adyen explaining its own technology, and it does not disclose the authorisation improvement for individual merchants. The data platform scale figures are as of early 2025.
Sources
- The AI behind Uplift - Adyenadyen.com, Accessed
Compiled from public sources. These are not results from ATF Works customers.