- KPMG’s TaxBot development consumed months of drafting a 100-page prompt
- Partner-written tax advice had been scattered across countless laptops
- KPMG Workbench hosts multiple LLM models from competing vendors
When large language models began attracting global attention in late 2022, KPMG’s digital leaders immediately recognized potential benefits but also major risks.
Chief digital officer John Munnelly admitted that first experiments with ChatGPT produced “really scary” results, including the discovery of sensitive financial data sitting unsecured on internal servers.
That incident caused the firm to suspend experiments, restrict access to public AI tools, and reassess the dangers that uncontrolled deployment might introduce.
Building a private AI platform
KPMG subsequently began constructing a closed environment for AI work, supported by software licenses that allowed access to OpenAI and Microsoft systems.
This move gave the consultancy a chance to design applications within safer boundaries, eventually leading to a platform called KPMG Workbench.
The system combined retrieval-augmented generation, multiple LLM options, and agent hosting capabilities.
Rather than depending on a single vendor, the firm deliberately spread usage across OpenAI, Google, Microsoft, Anthropic, and Meta.
Throughout 2023, extensive effort was devoted to training employees on how to write prompts effectively and interact with AI writer systems.
By 2024, the Australian arm of KPMG initiated projects to build specialized agents. Among them was the so-called TaxBot, a tool designed to prepare tax advice.
Munnelly explained that development began by gathering partner-written advice that had been “stored all over the place,” often scattered on laptops.
That information, combined with Australia’s tax code, was placed into a RAG model to produce automated drafts. TaxBot, however, was not trivial to construct.
According to Munnelly, its creation required a 100-page prompt, drafted over months by a dedicated team, and ultimately fed into Workbench.
The result is a system that requests several inputs, seeks guidance from human experts, and then generates a 25-page document for client review.
Munnelly claimed the agent now performs tasks that once took two weeks in a single day, a change he described as “very efficient.”
He suggested that quick turnaround is particularly important for clients engaged in time-sensitive deals such as mergers.
Yet he also emphasized that only licensed tax agents are permitted to operate the tool, acknowledging that output without professional oversight is not suitable for general users.
Beyond efficiency, KPMG argues that the introduction of agents has boosted staff satisfaction, since repetitive tasks can be avoided.
Additionally, some clients have expressed interest in acquiring similar agents, generating revenue streams KPMG did not originally anticipate. Nevertheless, the firm concedes that measuring precise benefits remains difficult.
Via The Register