- KPMG taxbot development has consumed months of writing a 100 -page prompt
- Tax advice written by partners had been dispersed on countless laptops
- KPMG Workbench welcomes several LLM models of competing suppliers
When large language models began to attract global attention at the end of 2022, KPMG digital leaders immediately recognized potential advantages but also major risks.
Digital Director John Munnelly admitted that the first experiences with Chatgpt had produced “really frightening” results, including the discovery of significant financial data guaranteed on internal servers.
This incident led the company to suspend experiences, restrict access to public AI tools and reassess the dangers that uncontrolled deployment could introduce.
Build a private AI platform
KPMG then began to build a closed environment for AI work, supported by software licenses that made it possible to access the OPENAI and Microsoft systems.
This decision gave the Council a chance to design applications within safer limits, which finally led to a platform called KPMG Workbench.
The system has combined the generation of recovery, multiple LLM options and agent accommodation capacities.
Rather than depending on a single supplier, the company has deliberately spread the use through Openai, Google, Microsoft, Anthropic and Meta.
Throughout 2023, in -depth efforts were devoted to the training of employees on how to write effectively prompts and interact with AI writer systems.
By 2024, the Australian KPMG branch launched projects to build specialized agents. Among them was the so-called Taxbot, a tool designed to prepare tax advice.
Munnelly explained that development began with the collection of advice with partners who had been “stored everywhere”, often dispersed on laptops.
This information, combined with the Australian Tax Code, was placed in a cloth model to produce automated drafts. Taxbot, however, was not trivial to build.
According to Munnelly, its creation required a 100 -page prompt, written over the months by a dedicated team and finally fueled workbench.
The result is a system that requires several inputs, requests human expert advice, then generates a 25 -page document for customer exams.
Munnelly said that the agent was now doing tasks that took two weeks in a single day, a change he described as “very effective”.
He suggested that rapid turnaround is particularly important for customers engaged in agreements sensitive to time such as mergers.
However, he also stressed that only approved tax agents are authorized to operate the tool, recognizing that production without professional supervision is not suitable for general users.
Beyond efficiency, KPMG maintains that the introduction of agents has stimulated staff satisfaction, because repetitive tasks can be avoided.
In addition, some customers have expressed their interest in acquiring similar agents, generating KPMG income sources were not originally providing. Nevertheless, the company concedes that the measure of precise advantages remains difficult.
Via the register
