You can safeguard data with AI

The information security team at a financial services company faced challenges in managing confidential information, after a review revealed opportunities to strengthen their data handling processes.

In response, the company created a highly controlled environment to store sensitive applications and data files, but challenges remained in securely providing access to authorized employees, including financial advisors, marketing personnel, and human resources.

To access protected data, employees needed managerial approval, and files had to be manually classified into one of five categories based on the PII they contained. The process was time-consuming and prone to errors, leading to backlogs and potential oversights.

Existing artificial intelligence (AI) tools were inadequate for detecting unstructured data, like names and addresses, raising concerns about the accuracy of information classification.

The institution sought a more efficient solution for scanning and classifying documents to enhance their security posture and operational efficiency.

KPMG developed a generative AI tool to analyze and classify data within a controlled environment, helping to ensure that only authorized personnel could access sensitive customer information. The solution automated document classification based on the types of sensitive information they contained. By creating an algorithm that sampled documents and identified various types of confidential data, the project aimed to reduce reliance on time-consuming and inaccurate manual reviews.

Designed to enhance the client’s overall security framework, the project focused on accurate detection and classification of confidential information, thereby minimizing the risk of data breaches.

To implement this system, KPMG undertook the following steps:

• Developed a proof of concept using a sample dataset of files similar to those managed by the company.
• Created an algorithm that enabled a generative AI large language model (LLM) to detect 50 types of PII—including credit cards, names, bank accounts, stock holdings, and passwords—and return accurate classification results.
• Allowed the AI tool to identify anomalies within documents, such as misclassified fields or hidden columns.

KPMG also evaluated multiple LLMs for their efficacy and cost, ultimately recommending the most effective option based on accuracy and cost considerations.

The implementation of generative AI has transformed the client's ability to analyze and classify sensitive data files. With an accuracy rate of 97%, the AI solution significantly outperformed the previous manual approach and allowed the client to process documents containing unstructured data—approximately 95% of the files they needed to scan—efficiently and accurately.

The automation of document classification drastically reduced the time required for scanning. While manual reviews typically took 10 to 15 minutes, the AI solution accomplished this in just 5 to 10 seconds—approximately 100 times faster than human review. As a result, the client was able to reassign five full-time equivalents (FTEs) from manual document review to higher-value activities, enhancing workforce efficiency and allowing for real-time handling of sensitive information.

By extracting personally identifiable information (PII) from various sections of documents, in a process known as sampling, the client could achieve accurate classifications without sending entire documents to the AI model. This innovation reduced the cost of analyzing a page to just 5 cents.

The project not only addressed immediate operational challenges but also strengthened the client’s security posture. By enabling enhanced classification and handling of PII, the client mitigated risks associated with data breaches and enhanced overall data governance.

Step forward into a future enabled by AI.
Our data and AI consultants can show you how.