KPMG delivers huge productivity gains and drives new business with generative AI

Many companies are still grappling with the prospect of introducing large language models to their workstreams. KPMG Australia, on the other hand, has already developed and deployed KymChat, a generative AI agent that could revolutionize how the firm’s 10,000 employees communicate and collaborate. KymChat epitomizes what has been at least a decade-long focus on embracing technology and innovation at KPMG Australia. This focus has helped the firm develop new tax, audit, and advisory offerings, and has been a powerful driver of the company’s growth as well. 

Because KPMG, which is made up of member firms across 150 countries, embraces both technology and the change that new technologies so often create, the company is in a prime position to help its clients navigate changes in the digital landscape for themselves. KymChat helps expand these client benefits by enhancing the discoverability of the firm’s large stores of unstructured data, automating mundane tasks, and offering rapid responses to common queries. It does all this with the aid of GPT-3.5 and GPT-4 in a highly secure, private environment, meaning company data never has to leave the firm’s control.

A powerful and scalable generative AI solution

Launched in March of 2023, KymChat quickly began helping KPMG Australia employees surface insights from the company’s external and internal websites, knowledge repositories, and Microsoft 365 productivity files. Two months later, KPMG Australia began to scale the solution up, adding new users, new data ingestion sources, and new features. As the solution grew in scope, search times grew longer. Also of note, the KPMG Prompt Confidence Index, another automated in-house tool, began registering lower quality responses from KymChat. 

The same month, Microsoft announced that it would be adding vector search capabilities to Azure Cosmos DB for MongoDB vCore, a fully managed native Azure service capable of deep integration with multiple other Azure services. Vector search uses machine learning to determine the meaning and context of unstructured data, expanding search beyond text to include images, audio, and video files. Within two weeks of the announcement, KPMG Australia had adopted Azure Cosmos DB for MongoDB vCore for vector searching and retrieval in KymChat. The solution pairs this functionality with GPT-4 through Azure OpenAI Service. “The results of our adoption of Azure Cosmos DB for MongoDB vCore were startling,” says Robert Finlayson, Senior Product Manager at KPMG Australia. “Search quality jumped from 50 to 91 percent, and our results were being delivered in under a second.” 

KymChat also makes use of Azure App Service and Azure Container Registry, which helps the solution deliver faster results at scale. Backend code responsible for data preparation, prompt construction, and AI integration with Azure Cosmos DB for MongoDB vCore is stored in an Azure Container Registry docker image, and KymChat uses Azure App Service to deploy a containerized web app based on that docker image. “A major part of KymChat’s scalability comes from Azure App Services, which allows it to automatically scale based on live metrics and events within the business.” says Finlayson. “We can add new use cases and ingest new data sets without any degradation of service.”

After deployment, the KPMG Australia team reached out to the Azure Cosmos DB team at Microsoft, who offered support through problem resolution, architectural advice, and the sharing of best practices. “The Microsoft team behind vCore were passionate about their product, deeply interested in how we were using it, and open to feedback,” says Finlayson. “We had experience with Cosmos DB, which gave us confidence in its potential capabilities, but Microsoft's support for vCore was instrumental in ensuring smooth operations as we increased the scope and capabilities of KymChat.”

Putting the user at the center of design

New users start their KymChat experience with a quick tutorial that introduces the solution’s enterprise capabilities and conversational experience. Users are already reporting that the solution drives deeper subject matter understanding through its easily digestible document summaries. They also enjoy the speed at which KymChat helps them discover not only data but client insights as well. 

Finlayson describes how easy it is to incorporate generative AI into client-facing workloads. “Let’s say an employee is about to interview a candidate for an open senior tax advisor role, and they want to know what questions to ask,” he says. “In the past, this employee might have had to liaise with HR or conduct a time-consuming web search of multiple sources to obtain viable questions. Now they simply ask KymChat to create a list of 10 role-specific interview questions, and they have their interview prep finished in seconds.” 

To help employees make the most of KymChat, KPMG Australia has created a prompt engineering training module. This has been helpful for all employees but has been especially valued by new hires and recent university graduates. KPMG Australia also recently held a generative AI hack day, encouraging interested and talented people from across the firm to identify new opportunities for KymChat to integrate with employee workloads. The day was so successful that some of the ideas it generated are already in development. 

Meaningful change and much more to come

KymChat has already proven itself as a driver of productivity and creativity at KPMG Australia. The reduction in data surfacing times alone has created more time for client interaction and has begun to foster a culture of continuous learning within the firm. “After all,” notes Finlayson, “why not educate yourself on a wide variety of subjects when the time investment that was once a barrier between you and that information has been removed.” This ease of access to data can be especially valuable when users are unfamiliar with a specific subject and need a quick overview, interwoven with KPMG research.

So far, 70 percent of the firm’s employees have adopted KymChat, making over 120,000 requests since its release. Even more impressive, the solution has led to as much as a 50 percent jump in productivity across workloads like document drafting, communications, sales proposal development, and research. KymChat has also unlocked new avenues for growth at KPMG Australia. Launched as a client-side market offering, the solution is helping the firm’s customers progress along their own generative AI adoption journeys. 

Other KPMG member firms from all over the world have taken notice of these advances and have begun asking for the ability to implement KymChat locally, adding to the potential benefits of the solution’s scalability. “We were successful in developing KymChat on the Microsoft platform,” says Finlayson. “Our vision was realized by Azure Cloud Services and enabled by Cosmos DB. Our extensive product roadmap is built on the continued innovation we expect to see from Microsoft Azure services.”

Find out more about KPMG Australia on Twitter, YouTube and LinkedIn.