KPMG Crafted a 100-Page Prompt to Build an Agentic TaxBot

KPMG crafted a 100-page prompt to build an agentic TaxBot.

Developed by the Australian arm of KPMG to streamline tax advice preparation. The creation of TaxBot involved crafting a detailed 100-page prompt, leveraging a Retrieval-Augmented Generation (RAG) model combined with Australia's tax code and internal tax advice documents gathered from various sources, including tax partners' laptops.

This initiative was driven by KPMG's chief digital officer, John Munnelly, who highlighted its efficiency after the firm's initial cautious approach to AI following security concerns with early ChatGPT experiments in 2022.

The TaxBot reportedly reduces the time to produce tax advice from two weeks to a single day, generating a 25-page first draft for clients. It utilizes models from multiple vendors like OpenAI, Microsoft, Google, Anthropic, and Meta, reflecting KPMG's strategy to avoid reliance on a single LLM provider.

The development process, which began in 2024 after staff training in 2023, involved significant effort, including months of work by dozens of employees collecting global tax best practices.

The system is designed to handle complex tasks such as analyzing regulations and producing detailed reports, though it requires human oversight (e.g., 4-5 experts) to ensure accuracy.

KPMG claims this innovation has not led to job losses, with staff surveys suggesting increased satisfaction as AI handles repetitive tasks, allowing employees to focus on more challenging work. Additionally, the firm has seen unexpected revenue from clients purchasing these agents.

Some skepticism exists regarding the reliability of LLM outputs for precision-critical fields like tax, with concerns about potential errors requiring expert validation.