The Impact of AI on Fraud Investigations: What Forensic Accountants Need to Know

Generative AI systems like ChatGPT are not immune to mistakes that can lead to issues when users rely on AI. AI must be properly managed. 

Experts say forensic accountants should use their skills, knowledge, and expertise to conduct most of the investigation in financial fraud cases. According to Kniepmann, using AI to complement human forensic accounting, compared to a start-to-finish automated report, minimizes the risk of incorrect solutions. Human verification will also ensure that the AI is not plagiarizing or producing false information.

‍

Sang notes that even the most advanced AI models can exhibit bias and discrimination. AI can flag false positives, particularly among certain racial groups and geographic areas. Sang says humans can over-rely on these false positives without proper training and rush to incorrect conclusions.

However, there are ways to limit AI bias and discrimination. According to Sang, it begins with robust testing and validation before AI implementation. For larger firms, Sang proposes an oversight committee that adopts a governance framework within which AI can operate. At the same time, a smaller company should establish a policy on how to handle AI. With guidelines, the risk of AI bias and discrimination is mitigated.

Data privacy and security represent another concern with using AI, as many AI systems could require access to large amounts of sensitive financial data or personally identifiable information. Building guidelines and limits for an AI is essential in maintaining security, and it’s important to set those limits before implementing AI. 

ChatGPT, for example, has access to nearly the entire internet. An AI in an online server, such as Microsoft OneDrive or Google Drive, could hypothetically access all the cloud-based files shared with the user. For a firm intent on using AI to help with a financial fraud investigation, the company should carefully select which files it wants the AI to have access to, such as transaction data and non-identifiable information, while limiting information that is unnecessary for the AI to have. 

“If you’re not feeding it in a closed environment, then every other AI system out there is using your learnings and using your data,” Chee says. “It needs to be within a closed environment, which means it’s within your organization and not being leaked anywhere. The internal organization can have an AI hub where it constantly learns and grows, but that hub needs to be locked down.”

Kniepmann agrees, emphasizing the importance of knowing exactly how strong the company’s cybersecurity is and whether it can handle AI. “When you implement AI within your internal system, it has the ability to search all of the file folders in locations that you have available to you that you may not have realized you had available,” Kniepmann says. “You don’t want to implement it too fast until you understand your security infrastructure.” 

AI-based fraud detection limitations 

For now, AI has limited contextual understanding and struggles with nuance and context easily grasped by humans. Systems may focus on individual pieces well but often cannot recognize the relation to a bigger picture, which is crucial in a financial fraud investigation.

It’s another reason AI should still be used alongside human expertise.

Given that it is difficult for AI to use context in its transaction processing, it’s important for humans to use their outside knowledge when working with AI to verify all results that the program finds.

Sang offers a solution to this: retrieval augmented generation (RAG). Using RAG, a user can provide private, local knowledge as the context to the AI, which can help it solve for the solution. 

Ethical considerations in using AI 

AI users should spend significant time understanding the risks around privacy, data protection, and discrimination concerns and how to mitigate them before implementing AI.

To attempt to remove ingrained bias from AI, programs can be trained using data from people of different regions, backgrounds, experiences, and ethnicities. Studies such as the Gender Shades Project, which took place from 2017 to 2020, have shown massive racial discrepancies, likely due to a lack of training on diverse races. Chee says that training software on people from one race is likely to make the software less reliable for people of other races.

Likewise, AI must learn using a diverse data set to give its users the most accurate results. “If the data that we’re feeding it has already been skewed, it could skew even further as the AI is learning,” Chee says.

Another ethical consideration is the importance of transparency in financial fraud cases. AI can operate as a black box, keeping how it processes and reaches solutions a mystery for most users. Since forensic accountants need to explain to fraud attorneys, or in a courtroom, how they reached conclusions, it may be more difficult if AI is doing the work.

Plagiarism also remains a concern, as users could inadvertently use generative AI to regenerate someone else’s intellectual property when writing professional opinions, for example. 

Experts agree that updated AI regulations are necessary. While governing accounting bodies, including the American Institute of Certified Public Accountants (AICPA), have begun providing these frameworks, federal legislation lags behind. 

Additionally, the U.S. Supreme Court's June 2024 overturning of the Chevron Doctrine made it more difficult for federal agencies to publish and enforce rules that govern the use of AI, according to Sang.

Chee believes that regulations could move toward an “obligation and self-reporting” model for AI creators and users. This means that the AI creator will be obligated to adhere to the rules established by legal entities while also being required to report details of their AI, such as when it’s being used, how it’s being used, which version of the AI is being used, and more. 

Forensic accounting work must also be admissible to courts, which is why AI is best used to increase efficiency rather than being relied upon to do the actual investigative work. If AI is used to produce a professional opinion, it may not be admissible, leaving the decision to determine the reliability of the AI to the judge's discretion. This could also result in the need for additional expert witnesses on the proper use and application of AI.

Establishing an AI framework allows companies to get ahead of the curve while also ensuring safer cybersecurity habits. Kniepmann’s firm, Anders CPAs & Advisors, implemented an AI policy and procedure guide, allowing them to very strictly define what they would use AI for and what their AI was allowed to do.

How forensic accountants can use generative AI effectively

AI is generally programmed to do what the user tells it to do, and the way to get the best results is to be as descriptive as possible. Prompt drafting has become a significant skill when using generative AI, and those who master it maximize their results. Not knowing how to make an effective prompt may result in a loss of efficiency as the user has to repeatedly change and correct their original prompt to get the desired result. 

Even with new technology, Kniepmann believes the most critical skills remain the ones that forensic accountants have always used. “In some ways, people who have been working in their particular area and have experience doing the old-fashioned hard way are at an advantage because they know what it takes to get to where you need to be,” he says. “Now, you have a tool that can make it more efficient.”

Furthermore, understanding how AI works can ensure that the user is not blindly trusting an unknown system. Most AI solutions are embedded in a platform and interact only with the user interface. Understanding the underlying processes is imperative because it enables more intelligent decision-making regarding implementing AI.

“If you don’t know what the code is saying and what the AI is doing and you go on the stand or in court and testify as an expert, you can’t say, ‘I just plugged this into a box, and it gave me this.’ That won’t hold up well in court,” says Chee.

Other emerging technologies impacting financial fraud investigations

Sang spends significant time working with the technologies that fuel AI in forensic accounting. He has his eye on several new programs that could change how financial fraud investigations are done. 

While not mature enough yet to be used viably, a blockchain-based technology, such as a distributed ledger, will ideally be able to provide immutable and transparent transaction records, thus making it a “perfect channel for evidence” and tool in fighting fraud, Sang says. Recent studies show that 45% of forensic accounting companies are working with and experimenting with blockchain-based technologies to enhance their fraud investigation abilities.³

Additionally, Sang predicts that quantum computing, which is currently only affordable for large tech companies, will eventually revolutionize cryptography and fraud detection. Robotic process automation, or RPA, is also a potential game-changer that can make forensic accounting investigations more scalable and efficient.

However, emerging technologies can also make it easier for criminals to commit fraud and for fraud accountants to trace threat actors. One of the most prominent examples is the use of AI to generate fake bank statements and financial documents that closely mimic authentic records, making it increasingly challenging for forensic accountants to detect discrepancies and confirm document validity.

AI best practices

The idea of AI making forensic accounting faster, easier, and more accurate is exciting. However, it’s important that financial fraud investigations do not become reliant on this new technology and instead use it to complement their existing workflows and methodologies. Artificial Intelligence is still a work in progress, but it continues to improve as we discover better and more efficient ways to use it.

Just as refusing to adopt AI altogether can perpetuate human errors, adopting AI too quickly can result in errors fueled by improper or inadequate training. Therefore, Chee says it’s best to be in the middle — adopting AI to be used in specific use cases while being careful to safeguard against its shortcomings.

‍This white paper is brought to you by Valid8 Financial, the global leader in Verified Financial Intelligence (VFI), leveraging the latest in AI and automation technology. Accounting, Legal and Government professionals follow the money faster using Valid8’s VFI Platform to rapidly parse, reconcile and categorize financial data from bank transactions, hand-written checks, and deposit slips. 

For more information about our VFI platform: