Auditing and Account practices and the use of manual operations with experience certified professionals is known throughout the business world. There have been some notable cases where some less than stellar practices were adhered to so more stanards, practices, and laws were creating to prevent the negatively impacting practices from happening again. And so it goes.
Where we’ve been able to make a real material difference in accounting, auditing, and fraud detection is within Machine Learning. It’s been about understanding patterns in the information (such as emails, transactions, journal entries, reports) and linking connection between the data (again transactions, movements of funds, hiring dates) and other soft sentiment analysis even at investment banking trade levels and compliance communications.
This is where organization of data is critical. Identifying all of these relationships mean that the data must be stored in a way where the data is accessible. The format of the data matters and so does the purpose and context. Are we looking at JE transactions stored in the ERP Financials operational database and we can use SQL to retrieve information? If we want to combine financials with emails, where do we create a taxonomy for the emails? Labels? After our sentiment and sentence structure analysis where do we store the scoring, ranking, and other feature attributes? And when and where do we combine it with other data? What part of the workflow do each of these transformation transpire? And what is the frequency requirement: hourly, daily, near-real time as the transactions flow in? Where do the exceptions get routed? When do we require a human to inspect results otherwise automated to ensure quality control/assurance our process isn’t creating false positives, and who is the person and how is their role defined? Is the data stored in a data lake? a a data warehouse? data lake house? Or, retrieved directly from the source like a virtual data warehouse or data lake? if virtual storage then how can a snapshot of testing evaluation data be retrieved in order to conduct hyperparameter tuning and create comparisons to the model going forward?
Obviously from the geeky questions above there’s a lot for your organization to consider in order to get into forensic accounting with machine learning and doing it correctly and successfully the first time. It’s so new that many times we are telling the business what we think they need in practice as a great starting point. Our services also use some prebuilt solutions we’ve put together to accelerate getting organizations through a forensic accounting process as a one-off or to build out their own internal ML enabled forensic accounting infrastructure.
Either way ML for accounting and auditing really enable key areas where there is a large volume of data at play. Think Wire transfer fraud, tax fraud, and GL fraud and GL reconciliation, just to name a few.
At the end of the day your organization needs to know where it most likely needs help. Where there are very manual processes and human insight from CPA, CFO, Treasury, Chief Compliance Officers, etc., who understand that there must be a better way to deal with the minutia or potential for too many false positives of an operation, they are the first lead into a requirement which can be turned into an ML model that can ultimately add value. It’s the expertise to combine the need and the technology that is the secret sauce.
