💬 NLP

An Agentic LLM Framework for Adverse Media Screening in AML Compliance

This paper presents an agentic LLM framework that utilizes Retrieval-Augmented Generation to automate adverse media screening for AML compliance, effectively distinguishing high-risk individuals from clean names by generating an Adverse Media Index score to reduce false positives and manual review efforts.

Original authors: Pavel Chernakov, Sasan Jafarnejad, Raphaël Frank

Published 2026-03-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pavel Chernakov, Sasan Jafarnejad, Raphaël Frank

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a bank manager trying to decide whether to lend money to a new customer. Before you say "yes," you have to do a background check to make sure they aren't involved in shady business, corruption, or money laundering. This is called Adverse Media Screening.

Traditionally, banks did this like a detective with a magnifying glass and a list of keywords. They would search for words like "fraud," "arrest," or "scandal." But this method is flawed. It's like searching for the word "apple" and getting results about fruit, the tech company, and a song by The Beatles all mixed together. You get too many false alarms (false positives), and human staff have to spend hours manually sorting through the noise to find the real danger.

This paper introduces a smarter, faster solution: The AMI Agent. Think of it as hiring a super-intelligent, tireless AI detective that doesn't just search for keywords, but actually reads and understands the news.

Here is how the system works, broken down into simple steps:

1. The Search Engine (The Librarian)

First, the AI acts like a librarian. You give it a name (e.g., "John Smith"). Instead of just guessing, it goes to the internet (using tools like Google Search) and grabs the top 10 news articles or web pages about that person.

2. The Web Crawler & Processor (The Sifter)

Next, the AI visits those links. It's like a robot vacuum cleaner that sucks up the text from the websites but ignores the ads, pop-ups, and navigation menus. It chops the long articles into smaller, manageable chunks (like cutting a long novel into chapters) so it can focus on the details.

3. The LLM Agent (The Detective)

This is the brain of the operation. The AI uses a technique called RAG (Retrieval-Augmented Generation). Imagine the AI has a giant library of facts in its head, but for this specific case, it pulls out the exact chapters from the news articles it just found.

It then asks itself three critical questions for every piece of text it finds:

Identity Check: "Is this article really about the John Smith I'm looking for, or just a different John Smith?" (This solves the "apple vs. Apple" problem).
Negativity Check: "Is this news actually bad, or just neutral?" (It understands that being "arrested" in a story about a movie plot is different from being "arrested for fraud").
Risk Check: "If this is true, how dangerous is it for the bank?"

4. The Verdict Generator (The Judge)

Finally, the AI acts like a judge. It takes all the answers from the detective phase and writes a final report. It gives the person a Risk Score (from 0 to 1) and, most importantly, writes a plain English explanation of why it gave that score.

Why is this a big deal?

It understands context: Traditional systems scream "ALARM!" if they see the word "bomb." The AI knows the difference between "a bomb threat" and "a bomb squad movie review."
It explains itself: If the AI flags someone as high-risk, it doesn't just give a number; it says, "I flagged this because Article #3 mentions a fraud investigation in 2022." This makes it easy for humans to double-check the work.
It's flexible: You can change the "playbook" (the questions the AI asks) without rewriting the whole code. If new laws come out, you just tweak the questions.
It's open: The creators made the code open-source, meaning anyone can use it, test it, or improve it.

The Results

The researchers tested this "AI Detective" on four groups of people:

Clean Names: Famous scientists who have never done anything wrong. (The AI gave them near-zero risk scores).
PEPs: Politicians and public officials (who are higher risk). (The AI gave them moderate scores).
Watchlists: People with regulatory warnings. (The AI gave them higher scores).
Sanctioned Criminals: People officially banned for crimes. (The AI gave them very high scores).

The system successfully separated the "good" people from the "bad" people, proving that AI can do this job faster and more accurately than old keyword methods.

The Catch

The paper admits the system isn't perfect.

It relies on the internet: If the bad news isn't online, or if the search engine hides it, the AI won't find it.
Name confusion: If two people have the exact same name and similar jobs, the AI might still get confused.
Hallucinations: Sometimes AI makes things up. The system is designed to prevent this by forcing the AI to stick to the articles it found, but human review is still needed for serious decisions.

The Bottom Line

This paper proposes a new way to handle financial safety checks. Instead of a human sifting through thousands of confusing search results, we now have an AI agent that reads, understands, and explains the news, acting as a powerful assistant to keep banks safe from criminals. It's like upgrading from a flashlight to a night-vision camera with a built-in translator.

1. Problem Statement

Adverse media screening (negative news screening) is a critical component of Anti-Money Laundering (AML) and Know Your Customer (KYC) compliance. Financial institutions must identify public information indicating a customer's involvement in financial crime, corruption, or fraud.

Limitations of Current Approaches:
- Keyword-based systems: Generate high false-positive rates due to the "signal-to-noise" problem (e.g., conflating legitimate mentions with adverse ones).
- Hybrid systems: Combine keywords with human review, offering better precision but suffering from high costs and poor scalability.
- Traditional ML/NLP: Struggle with contextual understanding and entity disambiguation, often requiring large, labeled training datasets which are scarce in this domain.
Regulatory Pressure: Evolving regulations (e.g., EU's 4AMLD through 6AMLD and the upcoming AMLR) demand more automated, adaptable, and rigorous screening processes that cover new predicate offenses like cybercrime.

2. Methodology: The AMI Agent Framework

The authors propose an Agentic LLM System called the AMI Agent (Adverse Media Index Agent). It utilizes Retrieval-Augmented Generation (RAG) to automate the screening pipeline.

System Architecture

The pipeline consists of five sequential stages:

Search Engine: Queries web search APIs (e.g., Google Custom Search) using configurable parameters (name, DOB, affiliations) to retrieve top- $k$ URLs.
Web Crawler: Fetches and parses HTML content, stripping boilerplate and ads to extract clean text using LangChain.
Document Processor:
- Chunking: Splits documents into overlapping text chunks (default 1000 tokens for API models, 500 for local).
- Embedding: Converts chunks into dense vectors using embedding models (e.g., Qwen3-Embedding-0.6B, BAAI/bge-large-en-v1.5) and stores them in a FAISS vector store for efficient similarity search.
LLM Agent (Core Reasoning):
- Executes a configurable playbook of assessment questions via RAG.
- Playbook Dimensions:
  - Identity Matching: Confidence that the document refers to the target subject.
  - Negativity Assessment: Severity of negative facts.
  - Risk Evaluation: Risk level of the attributed activity.
- Uses DSPy to enforce typed signatures, ensuring structured outputs (scores 0–1 and natural language justifications).
Verdict Generator: Synthesizes evidence from all documents and questions to produce a final Adverse Media Index (AMI) score (0–1) and a summary justification.

Key Technical Features

Multi-Backend Support: Compatible with local models (Ollama, vLLM) and API services (OpenAI, OpenRouter).
Zero-Shot Learning: Leverages pre-trained world knowledge of foundation models, eliminating the need for labeled training data.
Modular Design: Playbooks are defined in YAML, allowing easy customization of risk criteria without retraining.

3. Experimental Setup

Dataset: Four distinct populations representing a risk spectrum:
1. Clean Names: Academic researchers from DBLP (Low Risk).
2. PEPs (Politically Exposed Persons): From OpenSanctions (Low/Moderate Risk).
3. Regulatory Watchlists (RW): Individuals with regulatory actions/fines (Moderate Risk).
4. Sanctioned Persons (SDN): OFAC Specially Designated Nationals (High Risk).
Models Evaluated: Grok 4.1 Fast, Gemini 2.5 Flash, and GPT-4.1-Mini (via OpenRouter).
Metrics: Score separation (mean difference between populations), distribution analysis (ECDF), justification quality, and efficiency.

4. Key Results

Risk Separation: All models demonstrated clear separation between low-risk (Clean) and high-risk (SDN) populations.
- Clean Names: Mean scores ranged from 0.015 to 0.029 (very low false positives).
- SDN Names: Mean scores ranged from 0.730 to 0.863 (high risk detection).
- Intermediate Groups: PEP and RW populations showed graduated scores (0.06–0.18), correctly reflecting their moderate risk levels.
Model Performance:
- Grok 4.1 Fast achieved the strongest separation, with the highest SDN mean (0.863) and steepest ECDF curves for low-risk individuals.
- GPT-4.1-Mini showed more conservative scoring (lower SDN mean of 0.730), potentially reducing false positives at the cost of sensitivity.
Interpretability: The system successfully generated natural language justifications for every score, allowing human reviewers to verify the logic.

5. Key Contributions

Agentic Architecture: A novel end-to-end pipeline for automated adverse media screening using LLM agents with RAG.
Multi-Dimensional Scoring: A framework that separately evaluates identity matching, sentiment negativity, and activity risk, aggregating them into a single AMI score.
Open-Source Implementation: A flexible system supporting diverse LLM backends (local and API) and embedding models, promoting reproducibility.
Comprehensive Evaluation: A rigorous comparative study across four distinct risk populations and three state-of-the-art LLMs.

6. Significance and Limitations

Significance:

Contextual Understanding: Unlike keyword systems, the LLM agent can distinguish between "arrested for fraud" and "arrested development of a project," drastically reducing false positives.
Regulatory Compliance: The system provides interpretable justifications, creating an audit trail essential for regulatory defense and human-in-the-loop decision-making.
Adaptability: The playbook-based approach allows institutions to adapt to changing regulations (e.g., new AMLR requirements) without retraining models.

Limitations:

Search Engine Dependency: Effectiveness is limited by the search engine's indexing and ranking algorithms; unindexed or low-ranking adverse media may be missed.
Entity Disambiguation: While improved, distinguishing individuals with common names remains challenging without additional metadata (e.g., DOB).
Hallucinations: Although RAG mitigates this, models can still generate plausible but factually incorrect justifications, necessitating human review for high-stakes decisions.
Language Constraints: The current implementation is English-only; multilingual screening requires further development.
Temporal Bias: The system treats all retrieved content equally, lacking a mechanism to weight recent information more heavily than historical data.

Conclusion

The paper demonstrates that agentic LLM frameworks offer a superior alternative to traditional keyword-based screening by combining automation with deep contextual reasoning and interpretability. The AMI Agent successfully distinguishes between risk levels across diverse populations, offering a scalable, transparent, and adaptable solution for modern AML compliance.