In the fast-paced world of financial services, financial institutions are constantly seeking innovative ways to enhance efficiency and ensure regulatory compliance. For many, KYB (Know Your Business) is one of the most challenging processes to manage due to the role it plays in preventing fraud while maintaining regulatory compliance and efficient corporate client onboarding. In an effort to enhance this process, Sinpex partnered with Gradient to develop an AI-powered data extraction automation within the KYB process - creating significant ROI and ease-of-use for Sinpex customers.
One of the main challenges in KYB efforts is extracting data and information from official documents and comparing it with the data and self-declarations provided by the customer. Moreover, performing this step manually, as most companies do, is highly prone to errors and leads to significant quality issues. While the use of AI seemed like a viable solution to reduce the noticeable friction, getting there proved to be more difficult than anticipated.
Unstructured Data: The team needed a way to handle different documents from official registries, all with a different structure. Some in a very poor quality and difficult-to-read and some totally unstructured like e.g. shareholder lists. Primarily in PDF format - making data extraction even more challenging.
Inaccuracy in LLMs: The existing Large Language Models (LLMs) are not well-suited for KYB/Compliance topics, resulting in inaccuracies. Making it difficult to rely on the responses generated for KYB-related queries where even the slightest error could lead to significant compliance issues.
These challenges are not only straining on resources, but it also poses a risk to the team’s demand for complete smart automation and operational efficiency.
The AI-Driven Solution
Working with the team at Sinpex, we focused our efforts on attacking the challenges head on and finding a way to integrate it seamlessly into their existing infrastructure. In just two weeks, our team successfully launched the initial implementation with minimal effort required from the Sinpex team. Time to value was important for the team since this meant that they could now address other priorities for their business.
Automated Data Extraction
The Sinpex Team's first step of the process involved deploying an AI-powered tool specifically designed to parse and extract data from the unstructured PDF documents. This tool utilized advanced machine learning algorithms to identify, extract, and organize data from these documents - turning a previously manual and error-prone process into a streamlined, automated workflow.
Custom Gradient LLM for KYB
To address the issue of accuracy in KYB responses, Sinpex leveraged Gradient’s AI Foundry to develop a custom LLM that could fulfill the KYB requirements that their customers frequently encounter. This model was trained on a vast dataset of regulatory compliance materials and commercial register excerpts and shareholder lists to ensure that it could provide precise and reliable answers to KYB queries without the risk of hallucination. Hallucinations generally occur when an LLM generates a response that is either factually incorrect, nonsensical, or disconnected from the input prompt.
Integration into Existing Infrastructure
To ensure a smooth transition and minimal disruption to the team at Sinpex, we integrated the solution seamlessly into their existing GCP stack. This way Sinpex could leverage its existing infrastructure, while benefiting from the advanced capabilities of Gradient's platform for their solution.
The Impact
Gradient helped the Sinpex team drive significant impact across the business including:
Improved Response Capabilities: Likelihood of the system not being able to respond (e.g. returning no answer to the user) decreased from 45% to 1%. This means users can expect a response to their question 99% of the time, enabling a better customer experience.
Increased Accuracy: Number of correct answers improved by 30 percentage points, drastically reducing the need to modify the answer provided in comparison to before. The precision of the answers provided also improved by 35 percentage points, drastically reducing the work required for the team.
Expanded Coverage on KYB Related Questions: Customers are now able to ask any question that they might have via the KYB Questionnaire. Previously, Q&A was limited to predefined questions that were established via KYB documents. With the expanded coverage, customers are now able to receive more clarity and resolutions at their own convenience.