Kyc aml · Production

Adyen AI Applied Research Engineering: Integrity Risk Agents, Uplift, and Data Analysis

The problem

Traditional Integrity Risk operations required large human teams to manually research, analyze, and label cases; companies broadly struggle to apply AI meaningfully in critical large-scale production flows; and current SOTA AI agents achieve only around 16% accuracy on complex multi-step financial data reasoning tasks.

First attempt

Pre-LLM ML models surface Integrity Risk cases to human analysts but the framework still requires a human team to spend time on tasks that have the potential to be automated; current SOTA reasoning agents score only around 16% on Adyen's multi-step data analysis benchmark.

Workflow diagram · grounded in source
1
GNN surfaces Integrity Risk cases
Ai action
ML models, like Graph Neural Networks, surface cases to human analysts.
▾ source quote
“ML models, like Graph Neural Networks, surface cases to human analysts”
2
Human analysts review surfaced cases
Human review
A human team reviews cases surfaced by ML models, spending time on tasks that have the potential to be automated.
▾ source quote
“still requires the deployment of a human team that would eventually spend some time on tasks that have the potential to be automated”
3
LLM agents automate human-heavy tasks
Ai action
AI agents with access to tools like database queries and web search automate human-heavy Integrity Risk workflows.
▾ source quote
“AI agents with access to tools like database queries and web search presents a great opportunity to automate those human-heavy workflows even further”
4
RL optimizes agent trajectories
Feedback loop
Reinforcement Learning using PPO or GRPO optimizes agent trajectories in highly regulated environments.
▾ source quote
“Reinforcement Learning (e.g. using PPO orGRPO) to impact the optimization of agent trajectories in highly regulated environments”
5
Uplift AI decides transaction course
Ai action
An AI trained on a massive dataset decides the course of every transaction by balancing cost, risk and conversion.
▾ source quote
“an AI trained on a massive dataset decides the course of every transaction by balancing cost, risk and conversion in an optimal outcome”
6
Agents reconcile structured and unstructured data
Ai action
AI agents reconcile structured and unstructured data to generate insights, leads, recommendations, and spot mistakes.
▾ source quote
“trustworthy AI that can reconcile structured and unstructured data, generate insights, leads, recommendations, and spot mistakes”
Reported outcome

Adyen has deployed Uplift across every transaction on the platform and co-built the DABStep benchmark with Hugging Face using over 450 real-world tasks; ongoing AI agent research for Integrity Risk and foundational models carries potential for substantial productivity gains.

Reported metrics
SOTA agent accuracy on DABSteparound 16%
real-world tasks in DABStep benchmarkover 450
Payment volume processed (2024)more than USD 1.4 trillion
Year-over-year payment volume growth~25% YOY
Show all 5 reported metrics
SOTA agent accuracy on DABSteparound 16%
real-world tasks in DABStep benchmarkover 450
payment volume processed (2024)more than USD 1.4 trillion
year-over-year payment volume growth~25% YOY
productivity gains from AI advancementssubstantial productivity gains
Reported stack
Graph Neural NetworksLLMsPPODABStepAdyen UpliftGenAI platformdenoising autoencodersTransformer-based architecturesconference companionscode companionsHugging Face
◆ Does this fit your context?

Compare to your context

Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.

Compare to your context →
~30 seconds · free
Source
https://www.adyen.com/knowledge-hub/unlocking-value-through-ai-applied-research-engineering
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Adyen has deployed Uplift across every transaction on the platform and co-built the DABStep benchmark with Hugging Face using over 450 real-world tasks; ongoing AI agent research for Integrity Risk and foundational mo…

What tools did this team use?

Graph Neural Networks, LLMs, PPO, DABStep, Adyen Uplift, GenAI platform, denoising autoencoders, Transformer-based architectures, conference companions, code companions.

What results were reported?

SOTA agent accuracy on DABStep: around 16%; real-world tasks in DABStep benchmark: over 450; Payment volume processed (2024): more than USD 1.4 trillion; Year-over-year payment volume growth: ~25% YOY (source-reported, not independently verified).

What failed first in this deployment?

Pre-LLM ML models surface Integrity Risk cases to human analysts but the framework still requires a human team to spend time on tasks that have the potential to be automated; current SOTA reasoning agents score only a…

How is this kyc aml AI workflow structured?

GNN surfaces Integrity Risk cases → Human analysts review surfaced cases → LLM agents automate human-heavy tasks → RL optimizes agent trajectories → Uplift AI decides transaction course → Agents reconcile structured and unstructured data.

Related kyc aml cases

WHAT TO DO WITH THIS

Now compare it to your context

This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.

Browse the Kyc aml cases →