Neil Walker, Head of Product for IDP vendor TCG Process, sat down with IDM editor Bill Dawes to explain the company strategy behind the utilisation of a different kind of Large Language Model (LLM) and the push for ‘explainable AI”.

IDM: Where does TCG Process see the future of AI in Intelligent Document Processing (IDP)

NW: When it comes to AI, there are many who see the hype but don’t understand the value. Everyone knows that they will at some point have to look at AI for their business, but they don't really know how to make it work and bring value to their operations and that's essentially where we are trying to bring some context to AI as a solution.

Generative AI is game changing in what it can deliver for IDP with its ability to really read, understand and comprehend content in an entirely new and more human like way. But how do we make that usable inside of an organization? That is what we are focused on at TCG, orchestrating AI services, combining them with other technologies, validations, and people inside of a process to deliver the best possible outcomes.

We believe that it is this combination that allows our customers to embrace the latest AI technologies, without the fear of hallucinations and other challenges people perceive when talking about AI and be able to demonstrate how they are using AI responsibly within their business processes.

IDM: You recently announced a partnership with an AI LLM partner, Lazarus, can you tell us more about that?

NW: The ultimate goal of most generative AI technologies is to provide an answer, whether it be right or wrong. This is partly why “hallucination” has been such a big topic. When we talk about mission critical business processes in highly regulated industries, that's the one thing we really can't afford.

With this in mind, at TCG Process we are carefully selecting outcome driven partners. One of which is Lazarus, who originated in the challenging medical records processing world.

Lazarus built their own OCR engine for reading medical handwriting document which is obviously quite a challenge in itself. From there, they progressed to build a Large Language Model LLM focused on answering questions, based only on the context of the document that's being processing. Which they define as Extractive AI.

IDM: How does their approach differ from other Generative AI?

NW: The Lazarus LLM is looking to generate an answer based on a statistical relationship between words and phrases to extract the information that we're interested in, but only in the context of the document being analysed. This ability and their capability to also provide context about the answers was one of the big attractions for us, plus they use what is called a zero-shot approach which means that all of the model training has already happened therefore reducing the time, complexity and effort required for each project.

Information is collected based on natural language prompting, enabling features such as summarization, sentiment analysis, urgency detection in addition to content extraction. Whilst we are able to put measures in place to try and ensure the quality of the prompts, it’s still important that we can validate the responses from the LLM in some way. With that in mind DocProStar takes the results, right or wrong, and inputs that information into our platform. To ensure we deliver a correct outcome, we run our own validations against the data and cross-check the result, avoiding the chance for false positives from the AI.

The Lazarus LLM is trained on a very large data set, with a significant number of data points. This data point volume provides the model itself with a vast amount of knowledge and comprehension of information, it essentially knows how to read. It's like teaching a child to read and then you put a new document in front of them and you don't need to train them specifically about that new document because they understand enough about the wider world in order to understand something that is new. Extractive AI adds the context to then turn a prompted question into information to extract from the document.

IDM: What is an example of a use case?

NW: An example of how we can orchestrate AI within a business process is in insurance claims processing. We'll use Lazarus to answer some questions about a document. We'll also use an AI image analysis tool that specializes in analysing photos and video streams of vehicle damage. So, it will look at the video stream and say “This is a damaged panel, it's new damage. It's definitely part of this claim or yes, it's damaged. But it happened previously.” The idea here is ultimately you want to reduce fraud within Claims scenarios but of course there's multiple other use case scenarios for that as well. And there we bring the results of both of those together. Being able to use that approach really allows us to leverage all the value and advantages that LLMs bring, but without the risk of data pollution.

We are typically handling very complex documents. Where the combination of DocProStar and the Lazarus model’s comprehension that validates the data by connecting it to line of business information. That makes it usable in a business case.

I think in the IDP space, technologies are being somewhat commoditized, new technologies emerge at record speed which means that the Lazarus of today, might be something else tomorrow, or we might choose from three, four, five partners in that space and select the right one for the specific task in hand. But the process piece is where we really see the true business value.

Having been in the intelligent process automation space for many years now, I see the need for Process Automation, RPA, and Information management to converge with multiple other technologies, so they gradually become a layered solution.

IDM: Do you see that a challenge remains to integrate these platforms with legacy systems?

NW: In Tier 1 organizations, we are really seeing a desire to consolidate their technology stack not just for financial reasons, but more so to streamline processes. So, they're bringing in what they call a COE, a Centre of Excellence. And a lot of their goal is to work collaboratively between IT and business. Collectively they look at what technologies they have and how can they move on, future proof and become more agile to keep up with their competitors.  

It is where these issues arise that we feel we have a great solution as we can layer our processes across that entire technology suite which gives them path to transition and allows technologies to seamlessly work together through effective orchestration.

A current topic in our industry that keeps coming back around is process mining. Having the ability to look at document processes, reflect on them and determine where should we focus our effort. Process mining is difficult because it's got to bring information in from all those systems. We see an opportunity to migrate your processes more or less ‘as is’ into a platform such as DocProStar, where the improved visibility can be a great platform to begin your continuous improvement process journey, which can be an alternative and potentially a better approach.

TCG Process, with headquarters in Switzerland, is an international organization that develops and integrates input management and intelligent process automation software. Its solutions are used in industries such as banking, insurance, healthcare, government and public administration to digitize and automate document-driven processes. TCG Process sells both directly and via partners globally.  

TCG Process Australia; Tel: +61 2 9060 3727; info.aus@tcgprocess.com or sales.aus@tcgprocess.com

For more information visit www.tcgprocess.com and follow TCG Process on LinkedIn. 

Recent TCG Process/Lazarus Announcement (here)