- KPMG’s Taxbot Development Consumed Months Preparing a 100-Side Prompt
- Partner-written tax advice was spread over countless laptops
- KPMG WorkBench is hosting multiple LLM models from competing suppliers
When large language models began to attract global attention by the end of 2022, KPMG’s digital leaders immediately recognized potential benefits, but also greater risks.
Chief Digital Officer John Munnelly admitted that first experiments with chatgpt produced “truly scary” results, including the discovery of sensitive financial data that is unsecured on internal servers.
This incident caused the company to suspend experiments, limit access to public AI tools and reassess the dangers that uncontrolled implementation can introduce.
Building a private AI platform
KPMG then began constructing a closed environment for AI work, supported by software licenses that provided access to Openai and Microsoft Systems.
This step gave the consulting firm a chance to design applications within safer boundaries, which eventually led to a platform called KPMG Workbench.
The system combined recycling-augmented generation, multiple LLM settings and agent hosting functions.
Instead of depending on a single supplier, the company deliberately spread its use over Openai, Google, Microsoft, Anthropic and Meta.
Through 2023, extensive efforts were used to train employees on how to write prompt effectively and interact with AI authoring systems.
By 2024, the Australian arm of KPMG launched projects to build specialized agents. Among them was the so -called Taxbot, a tool designed to prepare tax advice.
Munnelly explained that the development began by gathering partner-written advice that had been “stored everywhere” often spread on laptops.
This information combined with Australia’s tax code was placed in a RAG model to produce automated draft. Taxbot, however, was not trivial to construct.
According to Munnelly, its creation demanded a 100-page prompt, prepared over months by a dedicated team and ultimately fed work bench.
The result is a system that requests multiple inputs, seeks guidance from human experts and then generates a 25-page document for client notification.
Munnelly claimed that the agent is now performing tasks that once took two weeks in a single day, a change he described as “very effective.”
He suggested that quick turns are especially important for clients dealing with time -sensitive offers such as mergers.
Nevertheless, he also emphasized that only licensed tax agents are allowed to operate the tool, acknowledging that output without professional supervision is not suitable for general users.
In addition to efficiency, KPMG claims that the introduction of agents has increased staff satisfaction as repeated tasks can be avoided.
In addition, some clients have expressed interest in acquiring similar agents, which generates revenue streams that KPMG did not initially expect. Nevertheless, the company admits that measuring accurate benefits remains difficult.
Via the register
