Reinventing the Auditing Process in Insurance with Generative AI

Unlocking Financial Information and Insights from Reports for Faster, More Efficient Auditing Process

Johnson Lambert LLP, founded in 1986 and headquartered in Vienna, Virginia, is a certified public accounting (CPA) and consulting firm specializing in services for the insurance industry. The firm serves a diverse client base, offering comprehensive services in auditing, taxation, consulting, business advisory, and regulatory compliance.

Auditing for insurance entities is a core area of expertise for Johnson Lambert. The firm ensures clients’ financial statements are accurate and comply with regulatory standards, while effectively managing risks. Accurate audits are essential for maintaining stakeholder trust, supporting informed decision-making, and enhancing operational efficiency.

Recognizing the critical importance of the service, Johnson Lambert’s leaders sought ways to streamline and scale the audit process. Manually-performed audits posed a challenge because they significantly increased the time, effort, and cost required to achieve a necessary level of accuracy and quality. Johnson Lambert’s auditors were bogged down with routine manual tasks, having to prioritize them over more impactful, customer-facing work. It was a bottleneck that the firm needed to resolve to grow its client base.

One area that needed improvement was report processing. This crucial process provides essential information that auditors require to evaluate a client’s financial statements, risk management practices, regulatory compliance, and overall financial health. Johnson Lambert’s auditors had to manually extract and validate this information from unstructured reports in PDF format. This involved converting the files into CSV format, extracting the necessary tables, and manually validating the extracted values by tracing and referencing them throughout various reports. This labor-intensive process was inefficient, time-consuming, and prone to errors, hindering the audit workflow.

Manual report processing created a significant bottleneck that slowed down the audit workflow at Johnson Lambert. It took auditors between 60 to 80 hours to complete a single financial risk audit, diverting valuable time and resources from more impactful, customer-facing tasks.

It was clear that improving audit speed, accuracy, and efficiency was crucial for Johnson Lambert. Generative AI technology presented the ideal solution to address these challenges.

In 2021, Johnson Lambert conducted a workshop with Provectus. When the firm’s leaders began exploring the possibilities of using generative AI in their business, they were reintroduced to Provectus through their AWS account team. Impressed by Provectus’ expertise in generative AI, Johnson Lambert decided to collaborate with us on developing a GenAI AI-powered solution for report processing.

Leveraging the AWS Generative AI Stack and Cohere LLMs to Streamline and Scale Report Processing

At Provectus, we understand the importance of starting an AI adoption journey with use cases that can quickly make a significant, real-world impact, while not forcing businesses to reinvent their operations from day one.

With this in mind, our collaboration with Johnson Lambert began with a series of comprehensive discovery sessions to help Provectus AI professionals understand their processes, challenges, and desired outcomes. We conducted a deep dive session to observe and understand the steps taken in an audit process, ensuring that the developed GenAI solution would enable Johnson Lambert’s auditors to perform financial risk audits faster and more efficiently.

We discovered that, in order to perform their duties, auditors required extensive manual labor to extract, move, and normalize tables from unstructured reports in PDF format. Our suggestion was to reduce the time and effort involved by utilizing Amazon Textract and Cohere’s Command Large Language Model, hosted on Amazon Bedrock. These services could accurately extract and classify information (tables, table values, and table names), and as a post-process, define validation rules for table names, streamlining the audit process.

One of our key considerations was Johnson Lambert’s requirement for the solution’s flexibility and low maintenance. The solution needed to be designed so their engineers could easily add new components to modify and improve it in future iterations. To meet this need, Provectus developed the GenAI solution for report processing using our Managed AI Services methodology. This approach ensured that the solution would be easily adaptable and sustainable, providing long-term value and efficiency.

Amazon Bedrock, a fully managed service for high-performing foundation models, was used for interacting with the chosen LLM, Cohere’s Command. The flexible nature of Bedrock allowed us to quickly iterate over other options, but considerations for cost, performance, and output quality led us to select the Cohere model. The LLM was used to normalize and refine the OCR process and to improve custom keyword extraction and reference resolution. During the extraction process, tables, including their values and names, along with references were stored in a vector database, allowing for efficient semantic search and retrieval of specific table data. In the future, this integration will allow Johnson Lambert to easily test and substitute models to achieve better performance results if needed.

The solution architecture follows best development practices to support Johnson Lambert’s initiative to further optimize their audit process and use case. It is designed and built to be executed in a headless format once the information extraction quality is satisfactory.

For this reason, the solution is divided into several components:

• A thin user interface that enables auditors to perform report processing. It can include a feedback loop feature, allowing users to use a built-in, semi-automatic evaluation mechanism to validate processing results.
• A backend API service and an ML pipeline powered by AWS Step Functions that facilitate report processing. It enables engineers to integrate other technologies, while keeping the system observable and monitored.

In simple terms, the generative AI solution for report processing is an information retrieval system that extracts tables (and their components) from unstructured PDF reports and adjusts table names to meet industry standards. It can trace and validate the extracted (and modified) table names and values across various reports. The GenAI solution augments the existing audit workflow for report tables, to optimize and enhance the performance of auditors.

The GenAI solution prototype was delivered as a production-ready system, but due to time constraints and report variability, several further improvements were recommended. These include, but are not limited to: pipeline enhancements (edge case processing, table variability, post-process formula validation); searchability improvements; better authentication and authorization; improvements to orchestration, observability, and lineage; enhanced user feedback features; integration with existing systems in the audit workflow; advanced headless operation capabilities; development of UI for comparison of historical tables; and the introduction of AI Governance best practices.

Provectus continues to work with Johnson Lambert to enhance the GenAI solution. These improvements are planned for future iterations.

From Manual Routine to Highly Impactful Work: Faster, More Efficient Auditing Process as a Driver of Business Transformation

Johnson Lambert clearly saw the potential of generative AI. By utilizing its ability to quickly and accurately extract and classify information from unstructured PDF reports, the firm planned to reinvent their manual audit process. Faster, more accurate, and efficient audits, performed with the help of generative AI, would enable their auditors to prioritize highly impactful, customer-facing work instead of spending hours on routine manual document processing. This transformation would unlock the firm’s potential, helping them to expand their customer base and grow their business.

Provectus was an invaluable partner to Johnson Lambert on this transformation journey.

In less than two months, Provectus developed a state-of-the-art prototype of the GenAI solution for report processing. Johnson Lambert’s auditors quickly adopted the solution to extract and manage the report information and insights required for financial risk audits.

While using the solution in their work, auditors reported that GenAI-enabled report processing was significantly more efficient and accurate than the manual process. The GenAI solution increased audit efficiency by 20% and accelerated document processing by 50%, all while minimizing the heavy lifting associated with finding, moving, and checking report information for accuracy.

Value of Using Generative AI for Report Processing

• Efficiency: Automates the management of report data, cutting the time required for a financial risk audit from 60-80 hours, to only 12-16 hours.
• Accuracy: Enhances the precision of data extraction, classification, and validation, reducing the number of errors associated with manual processing.
• Cost Savings: Lowers operational costs by minimizing labor-intensive manual work, allowing auditors to focus on customer-facing activities.
• Scalability: Enables Johnson Lambert to handle more clients and larger volumes of data without a proportional increase in workload.
• Compliance: Ensures adherence to industry standards and regulatory requirements through consistent and accurate data processing.
• Innovation: Positions Johnson Lambert as a forward-thinking firm leveraging advanced technology for better service delivery.

The GenAI solution for report processing has transformed the way auditors work at Johnson Lambert. They can now upload a PDF report to the solution to immediately start processing, extracting table data, normalizing table values and table names, and providing specific table metadata as references and notes. The extracted tables can be validated by reviewing the original PDF in the solution’s UI, and then exported as a CSV file for further processing by the auditing tools. All processed reports can be easily searched for and filtered in the UI to facilitate the auditing workflow.

Compared to their previous report processing experience, which required converting PDF files into CSV format and extracting and organizing all necessary information manually, Johnson Lambert’s auditors now save dozens of hours of work, significantly multiplying efficiency across the firm. The time saved allows the firm to focus on delivering highly impactful solutions for their clients, leading to a growing client base and a stronger bottom line.

Provectus will continue to support the AI transformation at Johnson Lambert. The teams will collaborate to refine and scale the GenAI solution for report processing, improving it based on auditors’ real-world experience. The next steps of the project are already in progress, focusing on further enhancements to maximize efficiency and impact.

Moving Forward

1. Learn more about GenAI’s potential from business and technology perspectives in The CxO Guide to Generative AI: Threats and Opportunities
2. Assess the readiness of your organization to adopt GenAI solutions with Generative AI Readiness Assessment with Provectus
3. Develop high-value generative AI use cases with “Generative AI by Provectus” offering