Beyond oil: Rewiring financial services

“Data is the new oil”¹ has become a boardroom cliché. It paints a picture of data as a precious fuel to be extracted and consumed. However, this comparison can be misleading. Oil is a limited resource - once used, it's gone. Data, on the other hand, particularly in financial services, is renewable and becomes more valuable with each reuse and connection. Unlike a barrel of oil that is spent once burned, a piece of data can be used repeatedly. It can inform one's decision, then be repurposed in countless other contexts, often becoming more useful the more it is shared and combined. However, treating data like a scarce commodity and locking it away in silos reduces its value. The real value of data comes from its ability to flow freely, connecting teams, systems and ideas to unlock deeper insights and smarter decisions.

Financial service firms are realising that data is no longer just a passive by-product of operations, but a strategic asset in its own right. Crucially, it is not a static asset like inventory or infrastructure; it is a dynamic asset and a relational network. When managed effectively, data is not merely stored or reported on; it moves, connects, enriches and evolves. Transactions, accounts, contracts and customer interactions; are not isolated drops of oil, they are interwoven links in a complex web. Each piece contributes to a larger, living system where insight flows freely, decisions are empowered, and value is continuously created and harnessed.

Before AI transformed financial services, data was typically processed in isolated systems, relying on traditional statistical methods that often overlooked the interconnected nature of financial activities. Today, modern AI-powered by tools like Graph Neural Networks (GNNs) and Large Language Models (LLMs) have changed the game. By leveraging complex algorithms to map and analyse the rich tapestry of relationships within data, these tools enable a deeper understanding of how entities interact and influence each other. They capture the dynamic interplay of factors such as financial risk contagion, fraud and customer behaviour².

The next question to ask is: how can data be harnessed as an asset? This brings us back to its value as an asset, setting the stage for how it can be strategically applied in a networked financial services environment. 

GNNs and LLMs are cutting-edge AI technologies that, fundamentally, deal with different types of data. GNNs focus on analysing networked data by identifying patterns and relationships, akin to mapping a social network to see how individuals are interconnected. Meanwhile, LLMs are specialised in understanding and generating human-like text, enabling them to interpret written language in documents and conversations.

At first glance, GNNs and LLMs may seem like an odd pair. But together they tackle one of the biggest challenges in financial services: making sense of both the structured and unstructured information that financial organisations generate and consume. GNNs excel in relational inference by finding patterns in networks of nodes (such as people, accounts or institutions) and edges (relationships like transactions, ownership or communications). In finance, this is crucial; many datasets naturally form graphs and GNNs can capture the connectedness that traditional models overlook³. Meanwhile, state-of-the-art LLMs (e.g. GPT-4 class models) possess a different strength: they understand natural language and domain context and can reason or converse about information. 

Financial institutions handle vast amounts of unstructured text including research reports, news articles, regulations, customer communications, emails and more. LLMs can analyse and reason with this data, extracting insights, answering questions or generating narrative explanations. However, pure LLMs have limitations⁴. 

This is where GNNs add value. A knowledge graph (powered by GNNs) serves as a single source of truth for an LLM, grounding it in factual relationships⁵.

Conceptually, GNNs and LLMs complement each other. GNNs thrive on structure; LLMs thrive on context. Together, they enable AI systems that can both "think" in graphs and "speak" in language. For instance, a GNN might detect an anomaly in a transaction graph (perhaps a suspicious payment loop) and an LLM could generate a natural-language explanation.

For the purposes of this blog, we’ll refer to the powerful integration of GNNs and LLMs as 'GraphLLM'. It is a hybrid approach that combines the structural reasoning of GNNs with the contextual understanding of LLMs.

Hyper-personalised financial products and advice

GraphLLM hybrids are paving the way for hyper-personalisation in financial services⁶. By treating each customer as a node in a dynamic context network, GNNs can infer micro-segments, while LLMs generate tailored recommendations or advice in natural language. This hybrid approach is already taking shape: banks are already employing GPT-4 to assist their advisors⁷⁸ and incorporating graph context could enhance this capability further.

Predictive compliance and regulatory foresight

LLMs can interpret draft regulations, while knowledge graphs map operational dependencies. When combined, these technologies enable proactive compliance⁹ by identifying which internal processes or systems will be affected by regulatory changes before those rules come into effect. This proactive approach helps institutions identify and address issues before they escalate. That in turn shifts compliance from a reactive checklist to a strategic alignment function.

Towards ethical, explainable AI decisions

Graph structures provide traceability, while LLMs supply human-readable rationale. A graph-enhanced LLM can explain why a loan was declined or why a transaction was flagged. This not only improves fairness and accountability but also aligns with evolving global standards for AI explainability in financial services¹⁰.

Anatomy of an AI-driven fraud detection system (data to insights, insights to action).

Let’s explore GraphLLM with a detailed example. In financial institutions, the ability to detect fraud is critical. Failure to put adequate measures in place can prove to be detrimental to the country and not just individual institutions¹¹. Traditional fraud detection systems often rely on static rules that fail to capture the complex and ever evolving nature of financial crime. GraphLLMs allow us to move beyond the static rules by understanding both the hidden relationships and the human context behind financial activities. The AI fraud detection systems¹² can be broken down into higher level steps:

• Step 1: Unified data architecture forms the asset to be leveraged to derive your network: Start by creating a centralised data framework that brings together both structured and unstructured data. This becomes the key asset for deriving insights from your network. Structured data: gather transactional data, account details and network interactions from multiple sources. Focus on ensuring data quality, consistency and integration into a unified data architecture. Unstructured data: collect related documents, communication records and external reports. This data should be pre-processed for text analysis by removing noise and extracting relevant entities and sentiments.
• Step 2: Knowledge graph/GNN maps the relationships between entities that are involved: Use a GNN to construct a comprehensive knowledge graph of entities (e.g. customers, accounts, transactions) and relationships (e.g. transfers, ownership). This graph represents the interconnected financial network.  
• Step 3: GNN identifies any anomalies: Now that the financial network has been mapped, the GNN can explore the nodes and edges. The GNN can perform relational inference, detecting subtle patterns across the entire financial network. The GNN is uniquely suited to surface any hidden relationships. For instance, the GNN could identify a sophisticated scheme where money is passed through a series of accounts to obscure its origin and destination. While this might be close to invisible when looking at transactions in isolation, it becomes a lot clearer when looking at the financial network.
• Step 4: LLM generates a human-readable summary for an investigator: What the GNN actually finds is a complex mathematical pattern that, in this context, is deemed a potential threat. This needs to be turned into something that is understandable by a human. The LLM takes the GNN’s output and generates a clear natural language summary that explains why any activity was flagged. A forensic investigator, for example, would receive a concise narrative output. This provides the rationale for fairness, accountability and swift action.

The AI fraud detection system leverages three key components in tandem: GNNs to map relationships and identify anomalies, LLMs to explain findings in human-readable format and unified data architecture to provide the foundational data asset that enables both structured and unstructured data analysis.

It is evident that the combined capabilities of GraphLLM transforms fraud detection from reactive and rule-based to proactive and intelligent.

While GraphLLM offers transformative potential, its deployment cannot be a technological free-for-all. The reality is many challenges in the AI space remain unsolved. For these systems to be trusted and adopted, they must operate within a robust framework of controls and ethical principles. These guardrails aren’t roadblocks, but the foundation of responsible innovation. They ensure that AI-driven decisions are fair, transparent and secure. There are some notable ones:

What does setting up data as an asset really mean? It entails rethinking data not just as an operational necessity, but as a core enabler of strategy, insight and innovation. It means architecting data ecosystems that allow data to be continuously connected, trusted, governed and ready for use in intelligent applications¹³. 

All of this presumes one critical and often overlooked truth: your AI is only as good as your data plumbing. Financial institutions must overcome fragmented systems, inconsistent ontologies, poor data lineage and patchy governance. Graph and LLM systems require robust data architecture built on data quality, interoperability and traceability.  This foundational shift is what underpins the true value extraction from advanced models like GraphLLMs.

Governance isn’t a barrier; it’s a catalyst for success. Without strong data ownership models, access controls, metadata and master data management and semantic coherence, even the most advanced GraphLLMs will underperform. Semantic coherence refers to the consistency and clarity of meaning across data definitions, labels and relationships. In data architecture, especially when building knowledge graphs or integrating AI systems, it means that terms and concepts are used in a unified and consistent way throughout the organisation. In simple terms, it means giving the AI meaningful context built on high-quality, well-governed data, enabling it to reason more accurately and responsibly. In the AI-driven financial future, data architecture and data governance are foundational infrastructure for success.

Adopting a GraphLLM approach is a strategic journey, not a flip of a switch. For financial institutions, the path to harnessing this technology can be broken down into three manageable phases beginning with the critical foundation: your data.

Financial institutions that treat data as static oil will be left behind by those who see it for what it truly is: a dynamic, relational asset. GraphLLMs are not just another tool; they represent a new architectural paradigm. Success won’t belong to those with the biggest models, but to those with the strongest foundations. The leaders will not be those with the largest models, but those with the most resilient foundations. Strengthen your digital plumbing, embrace the graph and prepare to lead in a financial landscape where intelligence is relational by design.

^References

1. ^{The Economist. (2017). The world’s most valuable resource is no longer oil, but data.}
2.  ^{IEEE TNNLS. Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Philip, S. Y. (2020). A Comprehensive Survey on Graph Neural Networks.}
3. ^{NVIDIA Developer Blog. Supercharging Fraud Detection in Financial Services with Graph Neural Networks (Updated)}
4. ^{OpenAI. (2023). GPT-4 Technical Report.}
5. ^{Microsoft. GraphRAG (Knowledge graphs + LLMs).}
6. ^{McKinsey & Company. (2021). How personalization is reinventing the consumer experience in banking.}
7. ^{OpenAI. Morgan Stanley customer story (GPT4 in wealth management).}
8. ^{BBVA. (2025). BBVA is now using ChatGPT to streamline legal queries and marketing processes.}
9. ^{The Bank of England and the Financial Conduct Authority (FCA). (2023). FS2/23: Artificial intelligence and machine learning.}
10. ^{European Commission. The AI Act (overview) + Explainability context.}
11. ^{International Monetary Fund. (2021). The Macroeconomic Effects of AML Intensification and AML-Related Gray Listings (WP/21/268).}
12. ^{AWS. Amazon Fraud Detector (AI/ML for fraud).}