The Australian branch of KPMG, a consulting firm, developed a lengthy 100-page prompt to create an AI system that delivers tax advice much faster than human consultants can.
During a presentation at Forrester’s APAC Technology & Innovation Summit, John Munnelly, the chief digital officer of KPMG, shared how the introduction of ChatGPT in late 2022 significantly impacted the firm. He realized that this technology was crucial for KPMG to adopt.
Initially, KPMG’s experiments with AI led to alarming discoveries, including a document that contained thousands of employees’ credit card numbers. This revelation caused a lot of concern, prompting the company to halt its AI trials and restrict access to ChatGPT while they evaluated the potential risks associated with AI.
Munnelly noted that the AI could accomplish tasks that previously took their team two weeks in just one day. Shortly after, while returning from vacation, he found numerous missed calls on his phone.
A graduate employee had shared a screenshot on social media showing that KPMG had blocked ChatGPT, along with a mocking comment about the company’s view on innovation. This incident caught the attention of a major business publication.
Fortunately, KPMG was already in discussions with Microsoft to acquire new software licenses, which included access to OpenAI’s tools. The firm began to construct a private AI platform and applications utilizing this technology.
Different branches of KPMG worldwide also started similar projects, sometimes duplicating efforts made by colleagues in other countries.
Ultimately, KPMG decided on a unified approach and created “KPMG Workbench,” a tool that offers retrieval-augmented generation (RAG), large language models (LLMs), and agent hosting for all its global offices.
To avoid relying on a single vendor for LLMs, KPMG’s Workbench incorporates models from several companies, including OpenAI, Microsoft, Google, Anthropic, and Meta. Throughout 2023, the firm concentrated on training its employees to effectively use chatbots and craft suitable prompts.
By 2024, KPMG began developing various agents, including the TaxBot mentioned earlier. Munnelly explained that the creation of this bot involved gathering tax advice from partners, which was often scattered across different locations, including personal laptops.
The firm compiled as much of this information as possible and integrated it into a RAG model alongside Australia’s tax regulations to produce an agent capable of generating tax advice.
Munnelly emphasized the efficiency of this system, stating that it could create a 25-page initial draft for clients in just one day. He highlighted the importance of speed, especially for clients involved in mergers who need quick insights into tax implications, making timely advice crucial.
He remarked that this innovation is fundamentally changing how KPMG operates. The agent was developed by utilizing the extensive 100-page prompt that was input into Workbench.
When asked about the details of this prompt, Munnelly mentioned that a dedicated team worked for months to refine it, resulting in an agent that requires four or five inputs before it generates tax advice and consults a human for guidance.
Only licensed tax agents have access to this tool, as its outputs are tailored for individuals with specialized tax knowledge.
Munnelly noted that he believes lengthy prompts like the 100-page document will likely become unnecessary in the future, as KPMG has created an agent runtime service.
This service enables various agents, including writers, editors, and managers, to collaborate on tasks like research and writing project summaries, which KPMG refers to as “credentials.”
According to Munnelly, deploying these agents has allowed KPMG to tackle tedious and time-consuming tasks that employees would rather avoid.
Surveys indicate that employee satisfaction has improved as AI frees them to focus on more engaging challenges, leading to a perception that the firm is more innovative.
He pointed out that employees prefer to work on stimulating problems rather than mundane tasks.
Additionally, the introduction of these agents has generated unexpected revenue, as clients have expressed interest in purchasing them.
Munnelly mentioned that KPMG is seeing advantages in terms of time, quality, and financial gain from its AI initiatives.
While he acknowledged the difficulty in quantifying the benefits of AI, he did not address potential job losses, instead suggesting that KPMG is eager to explore new AI applications.
Other Stories You May Like