AI Agent-Powered Compliance Review Automation for Financial Services

stripe · Finance · 2024 · source Agentic: yes · promotional bias: med · summary-only (no fetched source) · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Financial-services compliance review

Stripe needed to scale Enhanced Due Diligence reviews without linearly increasing compliance headcount or losing auditability.

Constraints:

• regulated workflows
• human decision authority
• fragmented data sources
• audit trail

Architecture

Stripe’s compliance automation, as described in the summary, uses Amazon Bedrock-powered ReAct agents to automate the investigative and research portions of Enhanced Due Diligence reviews. Complex compliance work is decomposed into bite-sized tasks in a DAG, agents investigate multiple data sources and jurisdictions, an LLM proxy/audit layer preserves traceability, and human analysts remain responsible for final compliance decisions.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  edd-review-intake[EDD review intake] --> dag-task-orchestrator[DAG task orchestrator]
  dag-task-orchestrator --> react-bedrock-agents[ReAct Bedrock agents]
  react-bedrock-agents --> compliance-data-sources[Compliance data sources]
  react-bedrock-agents --> human-analyst-review[Human analyst review]
  dag-task-orchestrator --> audit-trail[Audit trail]
  audit-trail --> human-analyst-review

Components

Data flow

An EDD review is decomposed into smaller DAG tasks. ReAct agents use Amazon Bedrock to perform investigation and research over relevant data sources and jurisdictions. The system records reasoning and outputs for auditability, then presents findings to human analysts who remain in the decision-making loop rather than delegating final regulated decisions to the agent.

Why this architecture

The architecture stops at research and investigation because financial-services compliance requires auditability and human accountability. A DAG of smaller tasks makes the workflow more tractable for agents, while the human-in-the-loop boundary prevents autonomous systems from making final regulated decisions.

Models used: Amazon Bedrock models · ReAct agents

Scale

• The problem sits inside an industry spending over $206 billion annually on financial-crime operations.
  “where enterprises spend over $206 billion annually on financial crime operations.”summary, stated
• Stripe handles about $1.4 trillion in annual volume, 1.38% of global GDP.
  “Stripe, a global payment processing platform handling $1.4 trillion in volume annually (representing 1.38% of global GDP), developed a sophisticated AI agent system to transform their compliance review operations.”summary, stated
• The system reached 96% reviewer helpfulness and reduced average handling time by 26%.
  “The solution achieved a 96% helpfulness rating from reviewers and reduced average handling time by 26%”summary, stated

Alternatives considered (roads not taken)

Fully automate compliance decisions

Rejected because: The system keeps analysts in the loop to preserve accountability in regulated reviews.

Manual EDD research by analysts alone

Rejected because: The goal is to reduce research/handling time and scale without proportional headcount growth.

One large monolithic compliance task for the agent

Rejected because: Stripe decomposes the work into smaller DAG-orchestrated tasks.

Patterns

human-in-the-loop-gate (compliance)

Agents automate research and investigation, while analysts retain decision authority.

tool-use-orchestration (workflow)

Complex compliance reviews are decomposed into bite-sized tasks in a DAG for autonomous investigation.

Tradeoffs

What part of compliance to automate.

Chose investigative and research portions over fully automated compliance decisions.

Rationale: The system preserves auditability and human accountability in a regulated financial workflow.

Automation boundary in regulated compliance.

Chose agent research with human analyst decisions over fully automated EDD outcomes.

Rationale: This improves handling time while preserving human accountability and auditability.

Workflow decomposition.

Chose DAG-orchestrated bite-sized tasks over one monolithic compliance prompt.

Rationale: Decomposition makes multi-source/jurisdiction investigation more manageable and auditable.

Gotchas (what bit them)

Manual reviewers were acting as navigators instead of analysts.

Root cause: Relevant compliance information lived across fragmented systems and jurisdictions.

Resolution: Use agents to gather and research evidence so reviewers can focus on final judgment.

Evaluation

Stripe reports reviewer helpfulness and handling-time changes.

Metrics: 96% helpfulness · 26% average handling time reduction

Outcomes

• Compliance teams could scale EDD reviews with less linear headcount growth.

Source transcript fetch failed; extraction is summary-only.