From days to hours: how ADC helped a leading Dutch bank scale AI-powered lending

The challenge

To lend money to clients, lenders at banks have to assemble a credit proposal for each client. This process consists of looking through unstructured files, such as pdfs, Word documents, call transcripts or emails, and spreadsheets, to perform the same regulated six-step analysis for each proposal using a lot of copy-and-paste. Apart from being difficult to standardise and easy to introduce mistakes into, this way of working is too slow to match the ambitions of ADC’s client, a large Dutch bank, for growth, risk management, and customer experience. Therefore, the bank together with ADC came up with a solution to speed up the proposal writing, while complying with credit standards and upholding human accountability.

The approach

To tailor the solution exactly to the needs of the lenders, ADC worked together with them on mapping the end-to-end proposal process and to encode the lenders’ knowledge into a so-called ‘prompt-vault’: a structured prompt library that contains prompts on standard sections for credit proposals, bank-specific risk frameworks an policies, and requirements for evidence, tone, and argument structure.

Around this prompt vault, a RAG (Retrieval Augmented Generation)-based system was designed that is capable of taking unstructured documents as input and selecting the right evidence for each evidence section from this data. From this, a fully structured draft proposal is then generated, which is then presented to the lender for final review and judgment.

The solution

• The resulting solution is a full workflow lending assistant. This lending assistant is a full-stack application that:

• Takes unstructured inputs and parses and indexes them so the system can identify information in messy and overlapping data

• Uses a domain-specific prompt bank that breaks the proposal into predefined sections, and uses expert-designed prompts to specify objective, tone, evidence, and output structure for each section, drawing from a RAG-layer for the right supporting documents

• Generates a fully structured draft proposal in professional credit format, with explicit traceability back to information sources

• Supports lender review and iteration by providing a clean UI in which lenders can navigate and edit the report. The UI also includes an agentic assistant to help lenders with refinement

The lending assistant is fully integrated with the bank’s governance and IT constraints and is built on a robust AI foundation.

Impact for the client

With the lending assistant, the bank has seen increases in speed and productivity, as well as quality and consistency of credit proposals. Taking the pilot to production took a few months and was quickly scaled up across the lending team of over a hundred users. Active usage and strong retention show that the lending assistant has real day-to-day value. The bank’s locations in other countries have also started requesting the tool be made available to them.

Furthermore, the lending assistant has demonstrated that GenAI can safely be embedded into core banking processes. The bank’s CEO has cited it as a flagship AI initiative that can pave the way for AI adoption in other lending products.