Australian businesses and governments have become familiar with generative AI over the past year.

However, through experimentation, it’s become clear that many public-facing generative AI tools are unable to meet the needs of corporate use. A lot of this comes down to the output that public-facing tools produce when prompted for a response. It’s not always clear what data these public-facing tools were trained on or where the responses come from. This has a direct impact on the quality and trustworthiness of the answers being generated.

What businesses and governments alike have learned is that they need to be more in control of the AI platform and the data used by the AI model to generate responses.

By hosting a private instance of generative AI themselves, organisations have been able to train it only on data held in their own knowledge bases or other internal data repositories. As these private generative AI setups produce responses based on privately held data, staff and other users can have greater assurance and confidence that the outputs are at least grounded in their own data.

While that sounds a fairly straightforward proposition, in reality, there is a lot of complexity involved in being able to do this successfully at scale.

It’s one reason that corporate generative AI programs still tend to be quite limited in the number of use cases they support - because it remains challenging to structure and ingest multiple sources of data from across an organisation to power more sophisticated generative AI capabilities.

But if generative AI is truly going to be as prevalent in business and government environments as predicted, two key questions must be prioritised: How is the data that generative AI produces verified, and how can this activity be performed at scale?

For an increasing number of Australian businesses and government agencies, knowledge graph technology represents the ‘missing link’ that will allow them to use generative AI accurately and at scale. A knowledge graph is an organised representation of real-world entities and their relationships, typically stored in a graph database.

Specifically, a knowledge graph combined with GraphRAG (Retrieval-Augmented Generation) can provide the technical means to retrieve and organise data for ingestion by the AI model, and transparent reasoning around how and why the AI model generated the response that it did.

Addressing the need for more data

Traditionally, generative AI pulls content from relational databases and neural networks that, if given enough data, produce natural language outputs. But, the technology's shortcoming is that it has no means to organise, understand, or analyse the relationships and connections between data from various sources stored in rigid tables.

This is problematic because generative AI’s power lies in its ability to look across multiple datasets and to synthesise the contextually relevant pieces of data into a single source of truth, in a simple, easy-to-digest format.

That is not to say generative AI can’t produce value when it queries only one source of data; we have seen this work in complex and jargonistic domains such as processing insurance claims, where the circumstances of each claim needs to be matched to the specific coverage clauses of a policy to determine the extent of liability and coverage.

However, for generative AI to fulfil its potential as a digital assistant for staff, it needs to be capable of sitting across a whole organisation’s worth of data. The answer to a single question is likely to require data from more than one source, which requires an enabling mechanism.

Knowledge graph technology can act as a solution by providing a unified framework that connects data sources, allowing AI systems to understand the relationships and associations between organisational entities and datasets. With a graph database, businesses can utilise the predictive qualities of AI across their entire value chain, unlocking contextual awareness and enabling AI systems to provide more insightful and relevant responses based on interconnected data points.

Grounding and validating outcomes

While generative AI is usually able to synthesize answers, the people prompting the AI for a response need to know that the answer they receive is grounded in fact, that there are references back to the source data, and that these can be easily checked out.

Grounding is essential to maintaining a strong foundation of accuracy and trust when using AI systems. It uncovers the deterministic facts in datasets, ensuring that generative responses are based on factual truth determined through explicit relationships.

Businesses should aim to train the large language models (LLMs) that power their generative AI capability with high-quality, structured data in a knowledge graph. Doing so allows them to visually track connections between data and why an answer is accurate, as opposed to an answer appearing with no ability to trace the reasoning behind it.

Taking the next steps with generative AI

Businesses and government agencies have been able to demonstrate the potential of generative AI in different operational contexts, though often not on a whole-of-organisation basis.   The Australian Department of Infrastructure, for example, has selected Neo4j as its database for analysing how Australian commuters and road freight moves around the country. The Commonwealth Bank of Australia has also deployed Neo4j to use knowledge graphs for network observability supporting the company’s focus on cybersecurity and its transition of digital assets from on-premise to the cloud.

While booming in popularity, with new levels of sophistication being added all the time, we are still at the ‘early adopter’ stage of generative AI. Until large language models have the ability to create and predict answers with greater accuracy by uncovering the complex relationships hidden within datasets, this may remain the case.

Businesses are turning to knowledge graphs to capture the semantic relationships between words and phrases to understand their environments and the opportunities that may exist within them. This approach enhances language understanding and enables AI systems to deliver accurate and coherent responses, effectively engaging with users in natural conversations.