The process of preparing, reviewing, and correcting detailed financial reports is an operational burden both to asset managers and their service providers. In addition to the mechanics of summarizing data, preparing tables, and drafting summaries, the report process presents an opportunity to validate performance, ratios, and positions. Many reports run in the hundreds of pages, making it difficult to complete comprehensive checks of the report against source records.

Our work with major, global asset managers presented several major challenges:

• Accurately and comprehensively parsing data from PDFs. This presents a number of challenges given changing formats, different headers for the same information, and the ability to read from tables, footnotes, and summaries.
• Creating and running specific checks and reconciliations. It was important to develop a library of checks and algorithms that could be tailored to the types of checks teams normally ran.
• Comparing PDF extracted data with electronic source data. Teams must manage differently formatted worksheets and tables, map them to the PDF line items, and reconcile the two sources.
• Tracking and resolving the issues. The process often requires up to three drafts, each consuming time on both the service provider and asset manager. For each draft, exceptions must be diligently tracked through resolution.

We developed a comprehensive end-to-end system that can be configured to read the various sections of a financial report from the PDF. The system then retrieves the corresponding electronic source documents and runs pre-configured checks matching the PDF extracted tables to the source tables.

We recently upgraded our document parsing with a series of AI tools that visually identify sections and tables of a report - irrespective of format and location changes. In addition, our AI applies natural language processing and named entity recognition algorithms to understand differences in headers and descriptions across different reports. For one major asset manager, we have developed a library of nearly 200 checks.

‍

There are two phases to our financial reporting AI: 1) configuration and testing, and 2) production runs. In the first phase, we setup the system to parse specific fund reporting formats. We then deploy the checks specified by the customer and conduct a full cycle check on a past report. Once we have setup a report and fully tested it, we work with the teams to schedule the production runs for the report. While we are increasingly developing self-service features to support production run checks, we provide technical support to ensure the scheduled rounds are supported. As the diagram below shows, our system reads the draft full report, then ingests the electronic supporting documents. The system assigns exception tasks to key personnel and provides a management oversight interface. Dashboards show all team members the status of the checks and reports.

As illustrated below, dashboards in our system highlight the progress on resolving exceptions.

There are three major outcomes we have achieved in this financial reporting case.

• A 400+ page report was completely checked in under 40 minutes. We were able to show customers that a lengthy report with thousands of positions could be thoroughly reviewed and reconciled in a short period of time.
• 92 percent time savings. We have baselined the time it takes to manually check reports and have seen as much as a 92 percent savings in labor when running comparable checks. Often, our customers don't have the staffing nor the time to run comprehensive checks. In all cases, our AI is able to complete a 100 percent coverage rate.
• Audit found zero errors. We have had reports audited post check by external audit teams. In one of our first reports, the team found a clean report with no errors.