- The KPMG TaxBot development consumed months of writing a 100 -page notice
- Fiscal advice written by members had dispersed in innumerable laptops
- KPMG Workbench Multiple LLM models of competitors
When large language models began to attract global attention at the end of 2022, KPMG digital leaders immediately recognized potential benefits but also the main risks.
The digital director John Munnelly admitted that the first experiments with ChatgPT produced “really terrifying” results, including the discovery of confidential financial data without guarantee on internal servers.
That incident caused the company to suspend the experiments, restrict access to the public tools of AI and reassess the dangers that unofficial implementation could introduce.
Building a private AI platform
Subsequently, KPMG began building a closed environment for AI work, backed by software licenses that allowed access to OpenAi and Microsoft Systems.
This movement gave the consulting the opportunity to design applications within the safest limits, which eventually led to a platform called KPMG Workbench.
The system combined the recovery recovery generation, multiple LLM options and agents housing capabilities.
Instead of depending on a single supplier, the company deliberately spread the use in OpenAi, Google, Microsoft, Anthrope and Meta.
Throughout 2023, an extensive effort was dedicated to training employees on how to write indications effectively and interact with AI writing systems.
By 2024, KPMG’s Australian arm initiated projects to build specialized agents. Among them was the so -called Taxbot, a tool designed to prepare fiscal advice.
Munnelly explained that development began by gathering advice written by partners who had been “stored throughout the place”, often scattered on laptops.
That information, combined with the Australian Tax Code, was placed in a RAG model to produce automated drafts. Taxbot, however, was not trivial to build.
According to Munnelly, its creation required a 100 -page notice, written for months by a dedicated team and finally fed to Workbench.
The result is a system that requests several entries, seeks orientation of human experts and then generates a 25 -page document for customer review.
Munnelly said the agent now performs tasks that once took two weeks in one day, a change that described as “very efficient.”
He suggested that the rapid change is particularly important for clients who participate in agreements sensitive to time, such as mergers.
However, he also emphasized that only licensed tax agents can operate the tool, recognizing that production without professional supervision is not adequate for general users.
Beyond efficiency, KPMG argues that the introduction of agents has increased staff satisfaction, since repetitive tasks can be avoided.
In addition, some clients have expressed interest in acquiring similar agents, generating income flows that KPMG did not originally anticipate. However, the company recognizes that measuring precise benefits is still difficult.
Through the registration
