The Australian arm of KPMG has developed a highly efficient agentic system using a 100-page prompt that generates tax advice documents in one day instead of two weeks. The consultancy firm's digital transformation accelerated significantly after ChatGPT's debut in late 2022, although initial experiments revealed security risks—including the discovery of a document containing thousands of employees' credit card numbers—which prompted the firm to temporarily restrict the technology's use. According to John Munnelly, KPMG's chief digital officer, the company subsequently began building its own private AI platform incorporating models from multiple major technology providers, including OpenAI, Microsoft, Google, Anthropic, and Meta.
Developing the TaxBot required months of work, which involved gathering tax advice documents previously written by partners and stored in various locations—often on laptops. This information, combined with the Australian tax code, was fed into a Retrieval-Augmented Generation (RAG) model, for which they wrote a detailed 100-page prompt. The resulting system begins work based on four or five input parameters, then seeks human guidance before generating a 25-page document. Munnelly emphasized that TaxBot usage is restricted to tax experts, as the system's output is not suitable for non-specialists. KPMG has since advanced the technology, creating an agent runtime service that enables different agents—writers, editors, and managers—to collaborate on tasks such as research or writing project summaries.
According to KPMG, implementing AI agents has not only increased efficiency but also improved employee satisfaction by freeing them from boring, time-consuming tasks. Employee surveys indicate that staff rate the firm as more innovative since AI allows them to spend more time on challenging assignments. Munnelly noted that the agents have also created unexpected revenue streams, as some clients have requested to purchase them. However, a Carnegie Mellon University study shows that AI agents generally only successfully complete assigned tasks in 30.3% of cases, indicating the technology is still evolving. Glenn Hopper, AI head at VAI Consulting, claimed just a few months ago that the time of AI agents are not here yet
Sources:
Simon Sharwood
Efosa Udinmwen
