Patrick Ulrich is the Chief Technology Officer at TCG Process, leading the development of TCG’s DocProStar Intelligent Document Processing and Orchestration platform. With over 24 years of experience within the technology industry, Patrick brings deep expertise in AI, machine learning, and process automation. He has been working with AI and machine learning technologies since 2006, advancing the capabilities of Intelligent Document Processing and process automation solutions.

IDM: You've been working with AI and machine learning since 2006, long before the current AI boom. How has your perspective on AI's risks and rewards evolved over these years, and what lessons can businesses learn from your experience about implementing AI safely?

PU: I really think AI is exciting for everybody.  The things you can produce with Generative AI, such as pictures, texts, movies, whatever, it's really exciting. However, leveraging these same technologies for extraction and reasoning, as covered by Intelligent Document Processing (IDP), can be costly. And you need to be aware, that if your LLM is delivering incorrect results on some of your documents, you cannot just retrain the LLM.  It will continue to deliver incorrect results.

While the initial setup time for AI is shorter than with older technologies, such as rule-based extraction, the security and hallucinations challenges can be problematic. You need a platform approach to handle hallucinations from AI services. There is always the possibility, that the result from AI is a hallucination, and because there are limitations with confidence measures, it is difficult to identify the false positive results.

We had an example where an AI model was used to extract info about contracts and the prompts were built to collect typical information from a contract. But as it happens in real life, a wrong document, like a claims document, came into the process. The AI returned random, inaccurate values, even though it was a completely different document class. If you can audit this activity with human in the loop (HITL) or with other rules that run afterwards, I think you're fine. If your process cannot manage or afford such instances, you should reconsider AI.

IDM: With major platforms like Microsoft, Google, and AWS developing their own OCR capabilities, what unique value does TCG's DocProStar bring to enterprises, and how do you envision the IDP landscape evolving?

PU: We have not built our own OCR or AI technologies, and we will not do so in the future.  These technologies are going to evolve quickly from horizontal to verticalized industry specific services over the next few years and picking the winners will be difficult.  We recommend a platform approach to allow customers to test and integrate value add AI services as technology improves without having to redo built core processes.

We have always focused on easily integrating the best of breed technologies onto our platform and never has that been more important than now.  While challenging Microsoft and Amazon, it’s quite tough for any company. We offer our customers the best OCR or AI service for their process requirements and specific needs. Whether it is Amazon, Microsoft or Google, for us it is simply a case of adding another engine to suit the customer use case, security and cost.

We've always been in a position to integrate capabilities from different engines into DocProStar. It is our best of breed approach towards integration that has ensured we manage our clients’ needs respectfully, either with OCR or other AI services.

IDM: Can you walk us through how TCG approaches AI hallucination challenges in document processing? What specific strategies have you found effective in maintaining accuracy?

PU: It has been our recommended standard practice to measure accuracy confidence and to verify what is coming out of the engine whether it’s OCR or extraction through a myriad of methods. And we still believe having a human in the loop plays a key role in correcting information and training models.

It is important for humans to validate extracted data because business processes and customers are best served with the provision of 100% accurate information. We need to be really sure that the values that we send to the next system are correct. And for that, we need human in the loop capability.  That said, we add lots of automatic checks using master data to compare values coming from multiple AI engines to reduce and absolutely minimise manual effort.  Our aim is to configure fast and accurate processes whereby humans in the loop only manage exceptions and IDP does the heavy lifting.

IDM: As CTO, how do you balance customer feedback with emerging technology trends when developing your product roadmap? Could you share an example of when these two factors aligned or conflicted?

PU: At TCG Process, Sales, Consulting and Development are always closely aligned. We receive direct feedback from all those in our ecosystem, and we try to meet with customers on a regular basis to gain direct user knowledge and understand any problems or challenges. It is our mandate to quickly resolve issues and advance functionality, and when these result in useful features, we implement advancements such that all customers will benefit. Once a new function is implemented into the product it is available to all customers, and they configure it for their benefit or not.

IDM: The term 'orchestration' is increasingly used in the IDP space. Could you explain what orchestration means in practical terms for businesses, and how it's shaping the future of document processing?

PU: We think that orchestration will be very important in the future especially with AI Services. Orchestration, for us, means the integration and management of specific services for extraction, classification, verification, validation, analysis and output of information with 100% accuracy to a customer within seconds.  This is what our customer’s customer and employees have come to expect of their key business processes. 

We run processes that call upon multiple AI services such as OCR, Speech to Text or summarization of a document. Because we integrate with, and engage our customer’s ERP, CRM or DMS, we can orchestrate an end-to-end process resulting in information processed into one system, wait for feedback then move further along in the process to the next system. 

Our platform leverages a common database model to combine ingestion, information with 3rd party service or database content to enhance decision making.  Architecturally this capability sets TCG apart in the orchestration space.

IDM: Where does TCG Process see IDP evolving in the future?

PU: We are looking at our IP in the IDP area for expansion into the areas of BPM and BPA. Our customers want assistance with the ingestion and ability to act on information quickly and accurately. This is an area where we already have some considerable knowledge. I believe this will grow as we will see what makes the most sense for our customer needs and for us. We will select areas that will bring value to our customers. We are exploring this now— how we can expand deeper into the business processes for our customers.

With the help of AI and our platform, we will be able to provide more intuitive process designs for our customers. There are AI services available, which can solve problems like handwriting or summarizing a document, which were not available in the past.  We foresee the development of new secure and verticalized services within orchestration expanding.

We also see the user desktop as a growing area for us. Today many still rely heavily on their desktop for document repository, either through mail, Teams or other channels. We will enrich DocProStar to ingest such documents easier, with greater speed and accuracy into the correct process.

For more information visit: www.tcgprocess.com