Finance ops · Production

Adyen Uplift: AI-driven global optimization across every payment transaction

The problem

Adyen's earlier ML approach optimized each payment step in isolation, preventing globally optimal decisions across the full transaction journey. Rule-based systems offered apparent control but could not scale with payment complexity and volume.

First attempt

Attempts to combine multiple payment decisions into larger deep learning models failed to meet real-time engineering requirements for latency and uptime in the critical payment flow.

Workflow diagram · grounded in source
1
Payment transaction arrives
trigger
“Every single transaction touched between 2 and 5 different AI endpoints where a machine-learning model was making a decision and had an allocated latency of 20 ms (median)”
2
Feature Platform serves input vectors
integration
“The AI is connected to a Feature Platform that provides low-latency, high-cardinality, high-volume, multi-geography input vectors to both the training and inference services. For slow, complex features we use distributed compute through …”
3
ML models make payment decisions
ai_action
“consists of a collection of machine learning models of various natures that share awareness and knowledge. These models are optimized globally through Reinforcement Learning and share the same objective: to balance fraud, cost, and conve…”
4
Entity resolution via transaction graph
ai_action
“a graph that links together transaction attributes in order to recognize entities. This graph represents a powerful source of information that we can model to extract features as well as to train on”
5
Authentication rail selection
routing
“Uplift uses decisioning on which is the best authentication "rail" to choose to balance fraud, conversion, and cost, and can choose among several available actions such as an exemption, a version of 3DS or a passkey”
6
Challenger model testing and promotion
feedback_loop
“When a challenger proves stronger performance than the principal model with statistical confidence, we sunset the principal (to "retired") and we promote the challenger to be the new "principal"”
7
Drift detection and observability
feedback_loop
“we constantly run diagnostics to ensure the performance stays in place. We run classic drift detection (classically under the umbrella of MLOps) as well as more complex algorithms to detect business performance drifts and biases such as …”
Reported outcome

Adyen Uplift drove all transactions through the Adyen platform during Black Friday/Cyber Monday 2024, processing 670 million transactions with 99.9999% API uptime.
Weak Supervision in production increased recall by 22%, reduced auth rate loss by 46%, and improved issuer refusal rate by 13%. Off-Policy Evaluation saved an estimated 20 weeks per year in A/B test overhead and contributed 9–54 million incremental transactions over six months.

Reported metrics
recall improvement (Weak Supervision)+22%
auth rate loss reduction (Weak Supervision)-46%
issuer refusal rate gain (Weak Supervision)+13%
AB test time saved per year (Off-Policy Evaluation)20 weeks/year
Show all 13 reported metrics
recall improvement (Weak Supervision)+22%
auth rate loss reduction (Weak Supervision)-46%
issuer refusal rate gain (Weak Supervision)+13%
AB test time saved per year (Off-Policy Evaluation)20 weeks/year
incremental transactions from Off-Policy Evaluation (6 months)9–54 million
transactions processed Black Friday/Cyber Monday 2024 (4 days)670 million
peak transaction rate163K per minute
peak API requests25K per second
API uptime99.9999%
off-policy to on-policy estimate correlation+80%
processed volume 20231 trillion USD
year-on-year growth 202326%
AI inference latency median per endpoint20 ms
Reported stack
SparkApache FlinkCassandraAirflowPostgresAlfredSHAP values
Source
https://medium.com/adyen/the-ai-behind-uplift-c1d36abe527d
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Adyen Uplift drove all transactions through the Adyen platform during Black Friday/Cyber Monday 2024, processing 670 million transactions with 99.9999% API uptime.

What tools did this team use?

Spark, Apache Flink, Cassandra, Airflow, Postgres, Alfred, SHAP values.

What results were reported?

recall improvement (Weak Supervision): +22%; auth rate loss reduction (Weak Supervision): -46%; issuer refusal rate gain (Weak Supervision): +13%; AB test time saved per year (Off-Policy Evaluation): 20 weeks/year (source-reported, not independently verified).

What failed first in this deployment?

Attempts to combine multiple payment decisions into larger deep learning models failed to meet real-time engineering requirements for latency and uptime in the critical payment flow.

How is this finance ops AI workflow structured?

Payment transaction arrives → Feature Platform serves input vectors → ML models make payment decisions → Entity resolution via transaction graph → Authentication rail selection → Challenger model testing and promotion → Drift detection and observability.

Related finance ops cases