UBO Identification in AML: Why LLMs Fail & How Agentic AI Delivers Accuracy

As global regulations tighten, the accurate identification of Ultimate Beneficial Owners (UBOs) has become a critical, yet error-prone compliance task. While many turned to AI, standard Large Language Models (LLMs) have proven inadequate for the analytical rigor required, highlighting a fundamental misuse of the technology. The solution lies in a more sophisticated, multi-layered approach combining graph technologies with Agentic AI to deliver the necessary accuracy, explainability, and automation.

The Regulatory Challenge of UBO Identification

Across the global financial landscape, identifying Ultimate Beneficial Owners (UBOs) is not optional-it's a regulatory imperative. Mandates from bodies like OFAC in the USA, coupled with the European Union's "50% or more" ownership threshold and the UK's stringent sanctions regime, demand not just compliance but full transparency and auditability in the ownership determination process. Despite this pressure, UBO analysis remains one of the most complex and labor-intensive tasks in financial compliance, frequently leading to the question heard at industry events like ACAMS Europe 2024: "How do you handle UBOs?"

Why Standard AI Fails: The Limitations of LLMs

Recent experiments with leading LLMs like ChatGPT, Gemini, and Copilot reveal a consistent pattern of failure when analyzing even moderately complex ownership structures. These models frequently make critical errors, including incorrect multiplication of ownership percentages across nested entities, failure to aggregate shares from sanctioned individuals, and drawing erroneous conclusions about sanction exposure. The root cause is not a failure of AI itself, but its misapplication. LLMs are designed for language generation, not procedural analytical reasoning. They lack the built-in logic to navigate complex ownership hierarchies and perform the precise calculations required for compliance.

Composite AI and AI Agent: The New Paradigm for Analytical Reasoning

A more effective approach is found in Composite AI — a unified architecture that combines multiple AI techniques to deliver intelligent, goal-oriented analysis. Within this architecture, Agentic AI components autonomously plan and execute multi-step reasoning tasks. They work with structured knowledge graphs and verified data, pursue clearly defined objectives (e.g., identify the UBO), and orchestrate workflows across other AI elements like LLMs and ML models. Most importantly, they generate transparent, human-readable insights that trace every step of the analysis, ensuring full auditability and explainability for regulators.

A Three-Step Framework for UBO Automation

Effective UBO automation requires a structured, multi-layered process. This framework breaks the challenge down into three manageable, technology-driven steps.

Step 1: Entity Resolution
Financial institutions often hold fragmented data across legacy systems. For example, one system may have a full company name, another just an email, and a third only a physical address. Before any analysis can occur, this data must be unified. Using an ontology-based knowledge graph, Entity Resolution technology synthesizes these disparate data points, inferring connections based on contextual clues to create a single, reliable view of each entity. This unified view is the non-negotiable foundation for accurate ownership analysis.

Learn More About DataWalk Entity Resolution >>>

Step 2: Graph Inference
With a unified dataset, the next step is to reconstruct and analyze the ownership logic. Graph Inference algorithms traverse the knowledge graph to map ownership chains across multiple levels, calculate aggregated shares (including those from sanctioned entities), and automatically identify individuals or entities that surpass control thresholds. This technology is also uniquely capable of detecting anomalous structures designed to obfuscate ownership, such as circular dependencies or networks of entities holding small shares.

In standard systems built on relational databases, performing such deeply nested calculations is not straightforward — each additional ownership layer requires new joins and complex query logic that quickly becomes unmanageable at scale. The challenge becomes even more complicated when analysts need to answer time-related questions, such as who controlled a company in 2020. This demands the inclusion of temporal data to track when ownership started and ended. Only graph-native architectures handle this complexity efficiently, as they persist relationships and temporal attributes as part of the data model, making it possible to query historical control paths directly.

Step 3: Agentic AI Interpretation
The final layer is the Agentic AI interface, which acts as a decision-support system for compliance analysts. This agent interprets the findings from the graph inference step and allows users to ask natural language questions like, "Why is this entity flagged as high-risk?" The AI can then explain the logic chain ("Entity D, which is sanctioned, owns 26%, and Person X owns 25%; their combined control exceeds the 50% threshold."), visualize the ownership structure, and generate audit-ready documentation compliant with OFAC, EU, and UK standards.

Learn More About DataWalk Generative AI >>>

DataWalk: Putting Composite AI into Practice

The principles of this three-step framework are embodied in the DataWalk platform. DataWalk provides a next-generation intelligence environment that operationalizes the Composite AI approach for complex compliance challenges like UBO identification. By integrating all data sources into a unified knowledge graph, DataWalk enables powerful Entity Resolution. Its advanced analytical engine performs sophisticated Graph Inference to uncover hidden ownership structures and calculate risk. Finally, its intuitive, user-centric interface empowers analysts to investigate, understand, and report on findings with the clarity and explainability that Agentic AI promises, turning a complex process into a manageable workflow.

Conclusion: The Future is Composite AI

The challenge of UBO identification has exposed the critical flaw in treating LLMs as a universal solution for every problem. Effective, scalable, and auditable automation in financial compliance demands a Composite AI approach-one that integrates graph-based data synthesis, powerful graph inference engines, and intelligent Agentic AI interfaces. As industry leaders and regulators demand greater certainty, the greatest risk is not choosing the wrong AI tool, but failing to adopt a purpose-built, automated solution that can meet the challenge.