The challenge: time-consuming manual work
In a world that is becoming increasingly automated, working with unstructured documents remains a challenge. Think of insurance files, medical documents, permits, and HR onboarding: documents play a crucial role almost everywhere. This also applies to the classic example of accounts payable and invoice processing. Invoices need to be matched with purchase orders and received items, but reference codes and product information are used differently by everyone. In some cases, information is simply missing or there is handwritten text within the document. The result? A lot of manual work is required, which is extremely labor-intensive. Processing time can take up to several days or even weeks, while technology is available to fully automate this process from start to finish.
Smart automation with Intelligent Document Automation (IDA)
KPMG's ‘Intelligent Document Automation’ (IDA) takes a holistic approach to process analysis, with the ESSAR method (Eliminate, Simplify, Standardize, Automate, and Robotize) at its core. The IDA team always looks at the entire process, the desired goal, and what the future process should look like. You could automate the process ‘as-is’, but is that the most sensible choice? Which process steps should be different, what is the biggest bottleneck and wherein lies the biggest improvement potential? These are questions we discuss together to come to the right answer, as automation and the implementation of Artificial Intelligence are complex issues with multiple possible solutions.
How does the technique under the hood work?
A key component in modern document automation is the use of ‘Large Language Models’ (LLMs). These are AI models trained on enormous amounts of text data, sometimes tens of millions of documents, and they are therefore able to understand natural text, recognize missing information, and make suggestions for completion. Unlike traditional, template-based systems, an LLM can interpret a document it has never seen before. For example, it understands that fields like ‘order number’, ‘order ID’, or ‘purchase reference’ mean the same thing in different contexts. This enables a much higher degree of automation.
Understanding context and reliability
Additionally, an LLM can classify, understand, or even complete documents with a much higher level of certainty when information is missing. AI can often explain the context well, even on the basis of incomplete data. A practical example that we like to illustrate this with is shown in the image below.
When we ask the same questions to a random AI model, we get a consistent answer: dog (with 95% certainty, based on its coat structure, paw pads, the shape of its paws and ears, and its behavior of rolling around on its back, which is common among puppies), and his breed is probably Pembroke Welsh Corgi (with 85% certainty, due to the color pattern, short legs, elongated body, and round ears, which together is characteristic of a Corgi). The difference between humans and AI lies in its level of certainty and explicit reasoning. Human reasoning sometimes relies on a gut feeling, while an AI model reasons on the basis of tens of millions of trained datasets. This is how an LLM also interprets language and information within a document.
Customer case: automating orders up to 95%
One of our customers, a large supplier with over 40,000 products in its catalog, submitted a request: "Help us process incoming orders more efficiently with smart technology." Customer satisfaction and efficiency were an important part of its strategy. In this case, there was a lot of manual work due to missing information, references that did not match, or other causes that made order processing take a long time. It took so much work that the organization was reaching the limits of its growth.
Given the situation, the AI tool 'Rossum' best met the customer's needs and formed the basis for automation. Rossum specializes in automatically reading and understanding large volumes of documents, such as the orders at issue, and uses AI to extract relevant data directly from various formats. Additionally, it can link the information to master data. KPMG supported the transition to the new way of working by integrating the tool into the customer's processes and work environment. The result? Almost 95% of all required data now were automatically recognized, leading to more work efficiency, less time lost to errors, and more time invested in strategic issues. Equally important: customer satisfaction increased because (urgent) orders were processed faster.
The next step in AI: agentic AI
Although most documents are now processed automatically, there are still cases in which the system needs extra context or situations in which you would want an additional verification layer, even if the information has been read correctly. These scenarios still need to be handled by employees. The expectation is that AI agents will be able to handle these exceptions in the future. The big difference between an AI agent and an LLM is that an agent can actually perform actions, such as automatically sending an email to a supplier when data is missing. As the use case with our client shows, the deployment of Digital Process Excellence (DPE) does not make people redundant; it ensures that they can focus on tasks at which they really make a difference, such as improving services and leveraging their expertise and talent.
