AI Governance Lighthouse Case Study Series: KPMG & KymChat
HTI’s Lighthouse Case Study Series shines a light on organisations that are actively exploring different approaches to human-centred AI governance. In the second case study, HTI presents insights from KPMG Australia.
KPMG’s AI Governance of KymChat
HTI’s second Lighthouse Case Study features KPMG Australia (KPMG) and its development and use of an internal Generative AI agent, KymChat. KymChat is being currently used across KPMG to answer staff questions about internal policies, compile thought leadership, and, in approved cases, preparing draft tax advice statements for human review and refinement.
KPMG recognised the opportunity and value of Generative AI to its organisation but was keenly aware of the risks, particularly around the potential exposure of sensitive client data. Such concerns initially lead KPMG to block the public version of ChatGPT soon after its launch.
In order to realise the benefits without exposing the firm to those risks, KPMG partnered with Microsoft to build an internal facing generative AI system on Microsoft Azure’s OpenAI Service, which was named KymChat. This ‘trusted cloud’ approach allows its employees to use this technology securely without client data leaving the KPMG environment.
“In this new era of GenAI we’ve been early innovators while at the same time building the structures and approaches that give us confidence that we’re avoiding potential risks. This is an on-going journey and we will continue to learn along the way,” said Jason Fogaty, Partner, Connected Technology Group, KPMG.
KPMG took a multi-layered and sequenced approach to AI governance for KymChat. It started with low-risk use cases, such as helping identify internal KPMG experts. As its functionality expanded, KPMG instituted appropriate guardrails to address the risks of Generative AI systems. These guardrails included the application of strict data governance processes to any data being used by KymChat, as well as appropriate staff training, policies and disclaimers.
“There’s a temptation for organisations to adopt AI systems without careful consideration of the risks and potential harms of these systems. In its development and deployment of KymChat, KPMG has shown the importance and value of engaging in thoughtful design and experimentation,” said HTI Co-Director and Industry Professor Nicholas Davis.
This case study explores how KPMG developed KymChat, how it is being used, and how KPMG is approaching the governance of this Generative AI tool. Important insights include:
The responsible AI guardrails and controls need to be specific to the system and context for deployment. One way to identify these is through careful experimentation with, and gradual deployment of, AI systems.
High quality data is critical for achieving the best results with Generative AI. Organisations will need to consider uplifting their existing data governance processes.
The development, deployment and governance of AI systems is more likely to succeed where an organisation takes both a bottom-up and top-down approach.
In April 2024, HTI launched its Lighthouse Case Study Series, as part of its AI Corporate Governance Program (AICGP), to highlight the insights generated and challenges faced by organisations on the frontier of human-centred AI development and deployment. KPMG is one of HTI’s advisory partners for the AICGP.
The first case study featured Telstra, highlighting their approach to AI governance, particularly the governance structure and complementary investments that support the development, use and oversight of AI systems within that organisation.
HTI will publish its third case study in the series next week featuring the University of Technology Sydney.
