Compliance monitoring · Production

How Amazon uses AI agents to support compliance screening of billions of transactions per day

The problem

Amazon must screen billions of transactions daily across its global businesses for sanctions compliance, but the human-intensive review approach creates significant bottlenecks—manual review cycles can take days, directly impacting customer experience through delayed transactions, account holds, and order fulfillment disruptions.

Workflow diagram · grounded in source
1
Transactions enter screening pipeline
trigger
“screens approximately 2 billion transactions daily across 160+ businesses globally to prevent prohibited transactions”
2
Tier 1: Fuzzy matching and vector embedding
ai_action
“advanced fuzzy matching algorithms and custom vector embedding models developed and deployed using Amazon SageMaker to compare each input entity data against entities on sanctions and other government lists. This layer is optimized for h…”
3
Tier 2: ML model noise reduction
ai_action
“Employs traditional machine learning models to filter out low-quality matches, significantly reducing noise. This layer decreases false positives by analyzing match quality signals, allowing compliance teams to focus on genuine risks”
4
High-quality matches routed to AI investigation
routing
“When a potential match passes through Tiers 1 and 2, a case is created and routed to our AI-powered investigation system”
5
Specialized agents gather and analyze evidence
ai_action
“these agents systematically gather relevant information, analyze it holistically following established Investigation Standard Operating Procedures (SOPs) curated by Amazon's Compliance team”
6
Recommendation agent synthesizes findings
ai_action
“The agent classifies cases as false positive or true positive with confidence levels, producing a comprehensive case summary including all evidence gathered, analysis from each agent, risk assessment, and a clear recommendation with supp…”
7
Low-confidence cases escalated to human
human_review
“when agent confidence scores fall below our defined threshold, the case is automatically escalated to a human investigator with a human-in-the-loop task”
8
Investigation summary and final decision
output
“Agents subsequently gather and analyze information and produce a comprehensive investigation summary with risk assessment and final decision”
Reported outcome

The AI-powered investigation system achieves 96% overall accuracy with 96% precision and 100% recall, automates decision-making for over 60% of case volume, and screens approximately 2 billion transactions daily, outperforming humans on historical decisions.

Reported metrics
Daily transactions screenedapproximately 2 billion
Overall accuracy96%
Precision96%
Recall rate100%
Show all 6 reported metrics
daily transactions screenedapproximately 2 billion
overall accuracy96%
precision96%
recall rate100%
automated decision-making rateover 60%
performance vs human investigatorsoutperforming humans
Reported stack
Amazon SageMakerAmazon BedrockAmazon Bedrock AgentCore RuntimeStrandsThinkTool
Source
https://aws.amazon.com/blogs/machine-learning/how-amazon-uses-ai-agents-to-support-compliance-screening-of-billions-of-transactions-per-day?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The AI-powered investigation system achieves 96% overall accuracy with 96% precision and 100% recall, automates decision-making for over 60% of case volume, and screens approximately 2 billion transactions daily, outp…

What tools did this team use?

Amazon SageMaker, Amazon Bedrock, Amazon Bedrock AgentCore Runtime, Strands, ThinkTool.

What results were reported?

Daily transactions screened: approximately 2 billion; Overall accuracy: 96%; Precision: 96%; Recall rate: 100% (source-reported, not independently verified).

How is this compliance monitoring AI workflow structured?

Transactions enter screening pipeline → Tier 1: Fuzzy matching and vector embedding → Tier 2: ML model noise reduction → High-quality matches routed to AI investigation → Specialized agents gather and analyze evidence → Recommendation agent synthesizes findings → Low-confidence cases escalated to human → Investigation summary and final decision.

Related compliance monitoring cases