How Australian firms are using graph databases

Neo4j, the graph database from the US-Swedish company of the same name, is used by 76% of the Fortune 100, and its Australian customers include organisations in the healthcare, policing and banking sectors.

Examples include the Australian Federal Police using it to assist investigations, while ANZ and Commonwealth Bank apply it to tasks such as creating a digital twin of an IT network or managing data lineage.

A graph database does not replace transactional databases or data warehouses, said Peter Philipp, Neo4j’s general manager in Australia and New Zealand. Rather, it is used to find relationships within data at a large scale. This makes it well-suited to tasks such as investigating complex data or making recommendations.

Because a graph creates deterministic relationships – for example, between a person and the products they have purchased – rather than relying on aggregated behaviour, it can create more explainable and accurate product recommendations, said Philipp. Unlike vectors, graphs show the relationships between data points and can therefore be used to explain how a particular conclusion was reached.

Another advantage of graph databases is that they allow very fine-grained control over which data any particular user or group can access. For instance, users in a specific role could be denied access to dates of birth. Such controls can also show the existence of a connection between two entities, such as people or organisations, without revealing the sensitive nature of that connection.

Applying graphs to retrieval-augmented generation (graph RAG) provides better answers from large language models (LLMs), in part by grounding a query with business data that the public model otherwise lacks.

For example, Swedish fintech operator Klarna uses Neo4j to store and connect data that was previously held in multiple cloud-based applications. This data is then made available to the company’s Kiki chatbot, which is based on LLMs from OpenAI. In mid-2024, Klarna reported that 85% of its employees were using Kiki.

Australian small business lender Prospa is another Neo4j customer. It initially used a graph database to quickly understand complex ownership structures and relationships, a process that is time-consuming with relational databases. Neo4j is now the company’s system of record, connecting multiple applications and supporting self-service analytics via Microsoft Copilot.

Two recent additions to Neo4j’s local customer roster are both ASX 20 organisations: one is a bank and the other a miner, said Philipp.

Graphs are an intuitive way of understanding data and providing context to AI agents. The graph provides context that results in more accurate answers and better communication between agents

The bank is using a graph database to represent the links between its institutional policies, which provides explainability. The project created a knowledge graph from the relevant documents using Neo4j’s Knowledge Graph Builder, with an LLM used to query its contents. “That can be done in a matter of minutes,” he said.

Similarly, the miner has created a graph of its internal policies and documentation from a large amount of unstructured data, along with chat agents to make it accessible. The goal with such applications is not to surface popular data, but the correct data. Philipp noted that graph databases can provide the necessary context – including the identity of the person asking the question – to ensure the right information is delivered.

According to Philipp, the past six months have seen an uptick in the use of graphs to improve the accuracy and explainability of generative artificial intelligence (GenAI), leading to “a very noticeable trend” of projects moving from experiment to production.

Attention is now turning from using GenAI to assist people, towards agentic AI capable of performing tasks without direct human involvement. Agentic AI typically breaks a task into smaller pieces and allocates each part to a suitable agent.

“Graphs are an intuitive way of understanding data and providing context to agents,” said Philipp. “The graph provides context that results in more accurate answers and better communication between agents.”

Digital twins are another area where graph databases are being put to work. They are relevant “anywhere there is a network in the real world”, he said, citing logistics and IT networks as examples, and they make it easier to automate tasks such as identifying events that are out of policy. 

While there is potential for agentic AI to take action without involving humans, “we’re not there yet”, Philipp cautioned, because “the accuracy hasn’t been there up until now”.

While graph RAG helps improve accuracy, some of the industries most interested in AI – such as finance and healthcare – are also among the most regulated.

“Customers are being responsible, and nothing is being done in a hurry,” he said. “Since banks are generally cautious by nature, if they are happy, then something is going right.”

An organisation’s data can be its differentiator, so there is an increasing focus on ensuring corporate data is cleansed and linked to provide context, said Philipp.