AI and Accounting Investigations: Incorporating Machine Learning and Natural Language Processing

Introduction

Advances in artificial intelligence (AI) promise to improve how work is done in nearly every field.? Two subsets of AI, machine learning (ML) and natural language processing (NLP), offer tools that can expedite and improve accounting investigations.? This article will discuss when and how to incorporate these tools, the limitations that the confidentiality of investigations imposes, and the indispensable role of human expertise and judgment.

ML and NLP

AI refers broadly to the replication of human thinking and decision-making with computers. Within AI, machine learning (ML) enables computers to learn relationships from data without explicit instructions through a process called ‘training’.? Natural language processing (NLP) enables computers to understand language as a human would.

The ability of these models to learn relationships without guidance is what separates them from the traditional analytics used in an accounting investigation, where investigators look to established financial ratios, accounting relationships, and behavioral tendencies in the data.

There are many algorithms to train models specific to the attributes of a particular data set and the goals of the analysis.? ‘Supervised’ models learn from labelled data (e.g., known fraudulent transactions) to predict specific outcomes or attributes, while ‘unsupervised’ models analyze data without any labels to discover patterns, structures, and clusters in the data.

Implementing these models requires use of a programming language such as Python, while understanding their mechanics requires a background in linear algebra, probability, and statistics.? As such, a cross-functional team is advisable for accounting investigations incorporating these tools.

Pre-trained models are available for general tasks via open-source or third-party applications (e.g., ChatGPT).? These models can give users access to complex and powerful models trained on much larger datasets than they could do themselves.? Confidentiality considerations, however, can limit the use of ML and NLP tools provided by 3rd-parties, as these often involve transferring data to an outside entity.

When and How to Use

AI tools work best when applied to very large amounts data and when relationships in the data are not well known.? As such, they may not be useful in all investigations. ?Several common areas of investigative evidence and their amenability to ML/NLP use are discussed below.

Financial Models / Spreadsheets – Low Potential

Accounting investigators often gain access to a repository of financial models and analyses in spreadsheet form.? Given the idiosyncrasies of these files, they generally do not lend themselves to ML tools presently available.? NLP tools can be utilized to parse any text in these files and identify clusters of related files, expediting their review by the team.? A trained forensic accountant remains indispensable in understanding the content of these financial analyses, their relevance to the investigation, and any further information requests they support.

Text Documents (e.g., contracts, invoices, memos) – Medium Potential

ML and NLP tools can assist the review of text documents in several ways.? Files received as PDF scans can become searchable text using optical character recognition (OCR).? Unsupervised models can cluster similar documents, identify the common topics discussed using co-occurring words and phrases, and provide summaries.? If some documents of interest have been identified, supervised models can use these labels to identify their distinguishing features and suggest other related documents.

General Ledger, Subledgers, Transaction Data – High Potential

The most granular data sources in accounting investigations are the general ledger, supporting sub-ledgers, or transaction source files. Supervised models can use labels, such as whether a journal entry or transaction was known to be fraudulent, to learn the predictive attributes of this activity.? The resulting model can then be applied to other periods for which the nature or occurrence of this activity is not known.? Unsupervised models can separate journal entries or transactions into clusters that share similar attributes for a targeted review.

Models Cannot Replace Accountants

Ultimately, ML and NLP tools can expedite and enhance an accounting investigation, but trained and experienced forensic accountants remain central.

The confidentiality of accounting investigations limits the extent to which teams can develop more sophisticated ML models.? As the data from one investigation cannot be used for any other purpose, teams cannot train an ML model on general tendencies across matters.? In this way, accounting investigators retain an advantage over models as they can learn from investigations over time while the models cannot.

The outputs of any ML or NLP analyses also require human assessment in the context of the business, regulations, and the unique circumstances of the investigation.? Models also cannot assess the status of an investigation and determine the information needed to move forward.? Lastly, a forensic accountant, not a model, must judge the totality of evidence, think critically, and communicate the results.? While further advances in AI may erode some of these distinctions, it remains unlikely to ever remove the forensic accountant entirely.

This article was written by William Floyd, PhD, CPA , an Expert Advisor to Floyd Advisory LLC .

About Floyd Advisory

Floyd Advisory LLC is a consulting firm providing financial and accounting expertise in areas of SEC reporting, transaction advisory, investigations & compliance, strategy & valuation, litigation services, and data analytics.