How a UK-based bank used AI to increase operational efficiency

This guidance is part of a wider collection about using artificial intelligence (AI) in the public sector.

AI technique used

  • machine learning

Objective

A large global bank wanted to automate its UK sales quality (SQ) process to improve 3 key areas:

  • compliance risk – faults and errors in unchecked cases going uncorrected
  • accuracy risk – inconsistency and potential oversights due to human involvement
  • slow processes – time-consuming reviews and delayed feedback loop

Situation

The SQ team is responsible for reviewing the sale of financial products for regulatory compliance. Currently the team is required to check a sample of 10% to 15% of completed sales. A team of 120 reviewers had to look at more than 10 different data sources and 180 data points to find and extract the information they needed to complete the audit. Each review took around 4 hours.

Action

The bank worked with an AI service provider to develop software to automate this internal compliance process.

Phase 1: Minimum viable product

The first phase was to develop a minimum viable product (MVP), showing the technical feasibility of the automation software.

The AI service provider worked closely with the bank’s SQ team to understand the current process including the:

  • sources for input data
  • data points to be extracted
  • checks conducted
  • human decision-making in the manual processes

The structured data, which made up 20% of the total data, was fed directly into the new system. The other 80% was unstructured. This included letters and memos in PDF, as well as payslips and bank statements saved as images. At least 70% of the checks involved in the SQ process required unstructured data.

The service provider then developed AI models to extract the required data, using a different model for each document category. This development process was complicated by the quality of the documents and the differences in the types of documents. Through ongoing analysis, the models were trained to address these inconsistencies with sample data.

The team tested and improved the models based on their real data sets and feedback from the bank.

Phase 2: refinement and deployment

After successfully completing the MVP, the focus shifted to:

  • further developing and refining the AI models to extract more data points from the documents
  • more closely replicating the manual process
  • deploying the solution into the bank’s live environment

Impact

The AI models can extract the required data from unstructured data sources much faster and more accurately than the previous manual process.

Using AI is helping the bank by providing:

  • greater compliance – the team can review 100% of cases, instead of just a sample set
  • improved accuracy – automated checks using AI models achieve close to 100% accuracy
  • faster process – elimination of the backlog of checks and moving checks closer to real-time
  • more time for team members to focus on making improvements

/Public Release. View in full here.