Executive Summary

→The customer due diligence process, a manually intensive task, presented inefficiencies, customer friction, and inconsistent information handling, prompting ING to identify it as a prime candidate for GenAI-driven transformation.

→ING embarked on a GenAI strategy initiative, beginning with the technology's potential rather than a predefined problem. This led to identifying customer due diligence (CDD) data extraction as a high-impact use case, focusing on reusing its components elsewhere in the business and a "human in the loop" principle to assist, not replace, analysts.

→Implementation was driven by a top-down push for safe GenAI adoption, while also drawing on a partnership with Google Cloud Platform and deploying cross-functional squads embedded with business SMEs for operational alignment.

→ The bank is implementing a phased rollout combined with robust, pre-established governance, frequent risk assessments, and a quality assurance process including human analyst review and a dedicated QA team, ensuring its controlled deployment and alignment with bank policy.

→ The GenAI solution is demonstrating tangible success, achieving nearly 90% accuracy in data extraction tasks, and projecting at least a 50% reduction in average handling time. Positive user feedback and early success are driving enterprise-wide learnings in AI adoption, risk management, and collaboration across functions and lines of business.

Introduction

The customer due diligence (CDD) process is a deeply manual and operationally intensive task, especially across wholesale banking divisions. Analysts are tasked with reviewing a wide array of documents submitted by clients, including ownership records, financial statements, compliance filings, and corporate structures. These documents vary widely in format, language, and complexity, and often result in back-and-forth interactions with clients, which take a toll in both efficiency and customer satisfaction.

ING shared a common pain point with the industry: analysts had to sift through content manually, often requesting the same document multiple times from different entities within a corporate group. This led to duplicate effort, inconsistent information handling, and delays in onboarding and review timelines. For ING, this inefficiency made CDD a prime candidate for transformation. We spoke with Joop Vahl, Chief Product Officer at ING Wholesale Banking Advanced Analytics and Artificial Intelligence about how his team leveraged a generative AI solution to improve the CDD process.

The CDD problem

ING identified the traditional CDD process as a prime candidate for GenAI intervention, due to an intense manual process involving extensive reviews of documentation. This manual effort was not only time-consuming but also a source of customer friction. Customers, particularly those with multiple entities within the same company structure, were sometimes asked repeatedly for the same documents as analysts followed procedural evidence collection.

Meanwhile, the legacy backend storage and software systems limited the bank's ability to achieve its full potential capabilities. This combination of manual intensity, customer inconvenience, and system limitations made the CDD data extraction process the "biggest problem" ING aimed to solve with GenAI at that point.

Why use GenAI?

ING embarked on a GenAI initiative to streamline and enhance the data extraction component of its CDD process. In a departure from their usual methodology of starting with a problem, the bank decided the potential of GenAI warranted beginning with the technology and then identifying the most impactful application.

The core of the solution is the application of GenAI to extract relevant information from customer documentation. A crucial element of ING's AI strategy is the "human in the loop" principle. The GenAI tools are designed to assist and empower CDD analysts, not replace them, by handling the brunt of the work of sifting through documents. This allows analysts to focus their expertise on analytical tasks. The bank also emphasises the importance of explainable GenAI, ensuring that the processes and outputs are understandable and transparent to all stakeholders involved.

A key strategic consideration for this use case was the development of reusable GenAI components. Recognising that many individual GenAI business cases might be "relatively thin," ING focused on building scalable solutions where the reusability across different applications would constitute the actual business case.

How was it implemented?

The implementation of the solution involved a wide range of stakeholders within the organisation, with specific processes set up to ensure everyone was aligned:

Strategy and approach

ING benefitted from an initial top-down push to use genAI safely and compliantly. The bank engaged external consultancies for guidance on market trends and potential GenAI use cases. This exploration led to an initial list of use cases, prioritising short-term impact and reusability. The CDD data extraction use case was selected from within the KYC domain, one of five initial buckets for GenAI experimentation at ING. ING Analytics employs squads of roughly 9-12 people comprising product management, UX, analysts, engineers, and data scientists. SMEs from the business are embedded within these teams to ensure a direct link to operational needs and facilitate smoother deployment.

Tech stack

ING made a strategic decision to partner with a strong vendor, selecting Google Cloud Platform. The CDD solution leverages Gemini models, BigQuery, and Vertex AI as core components. These are augmented with open-source tools; for instance, a preprocessor was crucial for achieving desired accuracy. The team adopted a progressive approach, starting with the base LLM and incrementally adding components like RAG as the need became clear, validating the "best in market" setups through practical application.

Rollout and governance

At the time of writing, ING is adopting a phased rollout of the solution starting with specific sectors and countries to manage complexity and ensure an inclusive approach, focusing not just on impact but also on incorporating challenging data privacy and complexity scenarios. Robust governance was established before the project commenced. This involved frequent risk assessments during the development stages, ensuring all stakeholders, including risk management, understood the technology and the steps taken.

Meanwhile, risk controls are embedded into the deployment pipeline on the development platform (GCP). A human-in-the-loop design is also a central pillar of this control framework, while a through-the-loop quality assurance process is in place: the LLM's output is reviewed by an analyst, but also reviewed by a Quality Assurance team to confirm that the model’s answers align with bank policy.

Outcome and lessons learned

The GenAI initiative for CDD data extraction is demonstrating tangible results and promising future potential:

→ The system achieves nearly 90% accuracy rate through the entire process, which includes the output review made by human analysts and QA teams. This performance is reported to be on par with, or even better than, purely human performance in the existing process.

→ A/B testing is performed with one group of analysts using the AI tool and another performing the task manually. Features designed to build trust include providing rationales for AI-generated answers, cross-referencing information across multiple documents, and displaying how similar cases were handled previously.

→ From the pilot stage, ING expects to achieve at least a 50% reduction in average handling time. The ambitious target is to reach a 75-80% reduction as the system matures and trust-building features are fully implemented.

→ Feedback from analysts during the Proof of Concept (POC) phase was highly positive. Users found the tool effective and expressed a desire to continue using it even before further enhancements, citing it as an improvement to their daily work.

→ Beyond the direct benefits of the use case, the project has served as a significant learning experience for ING. It has fostered a collaborative journey with risk management and enhanced the organisation's overall understanding and readiness for future AI advancements.