Precision LLMs in Risk Assessment, from Anti-Money Laundering to Vendor Vetting
A practical framework for deciding where large language models belong in a risk assessment stack, and where they don’t
Large language models (LLMs) promise major gains for risk assessment, but in regulated workflows, hallucinations, information leakage, cost and opaque results carry real consequences. In this whitepaper, Quantifind Co-Founder John Stockton offers a framework for responsibly integrating LLMs into anti-money laundering (AML), Know Your Customer (KYC) and vendor vetting technology. The framework sorts integration opportunities by who they help and when they run. It covers modeler productivity at model time, model upgrades at pipeline time, and investigator productivity at query time, with a human in the loop throughout. A focused use case shows how Quantifind applies LLMs to convert Department of Justice press releases into structured risk data, including the prompting challenges involved and the role of provenance in keeping results trustworthy.
- Where an LLM sits matters as much as which LLM you choose. Placement in the stack can mean the difference between a differentiated capability and a distracting “me too” feature.
- LLMs can improve production models without running in production. Using LLMs to generate training data and embeddings yields faster, more controllable models without LLM-specific risks.
- Focused domains are the best starting point. Applied to semi-structured Department of Justice press releases, LLM-based extraction exceeded 90% accuracy on people entities.
- Provenance is the backstop for imperfect prompting. Giving investigators direct access to source data lets them validate critical findings and catch hallucinations.
A Framework for Responsible LLM Integration
The Promise and Peril of LLMs for Risk Assessment
Why accuracy, speed, cost, scalability and compliance all shape whether an LLM implementation succeeds.
Quantifind and Precision LLMs
Weighing economic cost against model performance in a regulated industry.
Categorizing LLM Integration Opportunities
Modeler productivity, model upgrades and investigator productivity, and when each makes sense.
Focused Example, Structuring the Unstructured
How a Text-to-Table workflow turns DOJ press releases into structured risk data.
Learning to Learn with LLMs
Why starting with a pragmatic, low-risk use case builds the capacity to scale.
John co-founded Quantifind in 2009 and now leads Quantifind’s prototyping efforts to extract signals, and value, from vast amounts of diverse, unstructured data. John’s prior academic work included theoretical and experimental research in precision metrology, quantum entanglement, and atomic physics. In 2012, John was awarded the Peter Strauss Award by the Fannie and John Hertz Foundation for demonstrating exceptional leadership in his entrepreneurial endeavors. He holds a Ph.D. in Physics from the California Institute of Technology.
