How to use AI to detect accounting fraud

Picture of Mark Jolley
Posted by Mark Jolley

Artificial intelligence (AI) is proving increasingly valuable in the task of detecting and preventing fraud, but it’s in the field of accounting fraud where AI really shines.

One in 10 public companies are thought to commit securities fraud each year and virtually all manipulate their financial statements to some degree. Most do it to give the appearance that a company is worth more than it is. Some do it to manipulate their stock price. 

Whatever the reason, if we extrapolate the estimated cost of accounting fraud in the US to the entire world, the annual cost of accounting fraud to the global economy is at least US$1 trillion. It’s a number that we have quoted often, elsewhere. Putting that number in perspective, the world consumes about US$3 trillion worth of crude oil per year. 

So, accounting fraud matters and we haven’t even considered the indirect costs due to the resultant misallocation of resources. Some researchers have even found that a build-up of misreporting can lead to a recession or lower economic growth through its real effects.

This article looks at how AI can be used to detect - and therefore mitigate - accounting manipulation and fraud.

Table of contents


Accounting fraud is invisible

Accounting fraud is insidious because, like an iceberg, it is mostly invisible. It involves non-cash transactions embedded deep in a company’s financial accounts. Other forms of criminal fraud, by way of contrast, are transactional. 

Transactional fraud is quickly discovered. If a chap on the streets of Paris sells you the Eiffel Tower, you will very quickly discern you are not the true owner, even if you drink pastis and wear a beret.

Relative to the known prevalence of accounting fraud, therefore, most accounting fraud goes undetected. Even when detected, admission of guilt is rare and criminal charges even rarer. 

In the whole of 2022, the US Department of Justice tried just 72 individuals for fraud and convicted 56 at trial. Only one case of accounting fraud led to incarceration, that being Frank Okanuk, the former CFO of the PR firm Weber Shandwick. Of course, FTX was the big accounting fraud case of 2022, but even here it appears the authorities will press charges relating to embezzlement rather than account manipulation because accounting fraud is so darned hard to prove.

Red flags in traditional accounting

Although accounting fraud is difficult to prove, it can be detected via an array of forensic accounting techniques.  These techniques are essentially anomalous patterns in a company’s reported financial numbers or in its governance. 

These anomalies are known as “red flags." The greater the number of red flags, the greater the likelihood a company is cooking the books.  

Examples of accounting anomalies include:

  1. Strong revenue growth without corresponding growth in cash flows;
  2. Abnormally high accruals or capitalization of expenses in relation the size of the balance sheet;
  3. Unnaturally smooth earnings;
  4. Inverse correlation between operating cash flow and operating profit;
  5. Solid sales growth while competitor sales are struggling;
  6. Depreciation rates that don't correspond to the standard rates in a company’s industry;
  7. Outsized frequency of complex third-party transactions, many of which add intangible rather than tangible value;
  8. Frequent asset write-downs;
  9. Unseasonal improvement in a company's performance within the final reporting period of a fiscal year;
  10. Frequent re-statement of earnings;
  11. Unusual discretionary expenses.

The greater the number of red flags, the greater the likelihood a company is cooking the books.  

These and hundreds of other accounting anomalies can be used to detect potential accounting fraud. Some, such as those shown here, are simple while others involve complex ratios or lengthy computations.

Other indicators of potential accounting fraud rely not on financial statements, but on markers of poor corporate governance or business manipulation. This might include:

  1. Employing a little-known auditor;
  2. Replacement of an auditor resulting in missing paperwork;
  3. Lack of audit or other relevant committees;
  4. Disproportionate management compensation derived from bonuses based on short-term targets;
  5. Disproportionate compensation of management via options.

Return to top

Problems with traditional accounting 

A seasoned forensic accountant uses red flags for accounting and governance concerns to build a detailed picture of the extent of anomalies in a company’s financial accounts. From this picture, she will be able to provide an assessment of the risk of fraud at a particular company.

This approach faces two problems:

  1. First, it takes time and good forensic accountants do not come cheap. This is fine if we are assessing one or two companies, but what if we want to examine hundreds?
  2. Second, how do we analyse the data? How do we put all these red flags together in a coherent way so that we can get an overall estimate of the risk that a company is manipulating its accounts? How can we compare one company with another? How can we compare companies in different industries, or across different countries? Accounting standards vary widely among industries and countries.  

This is a complex problem. It is a big-data problem that frankly lies beyond the scope of the human brain. It is also a significantly more complex problem that we encounter with regular transactional fraud. In this case, the same patterns endlessly repeat and it is a relatively simple matter to build algorithms to detect anomalies.

However, companies are unique. A company’s business model, its products, its markets, its work force and its assets will be unique. Its approach to discretionary accounting decisions will also be unique. Accounting relationships that might be highly anomalous in some industries might be perfectly natural in others.

How can we compare companies that are unique in almost every respect?  The answer, of course, is AI.

What is machine learning

AI technologies are tailor-made for the detection of accounting fraud because they are able to analyse large volumes of data to detect patterns and anomalies which can be used for prediction. 

AI, in this instance, refers to machine learning (ML) algorithms, which are a subset of AI. 

An algorithm is a coded formula written into software that is designed to solve specific problems. The programmer details the problem and the data needed to solve the problem. The system analyses the data to deliver a solution. 

In this case, the problem is to find patterns of data that have been associated with particular outcomes. The ML algorithm uses these associations or patterns to create models which can be used to forecast the probability of certain events such as the probability that a company is manipulating its accounts. 

Machine-learning algorithms automate the process of finding patterns in data. As new data becomes available, the algorithms learn and adapt continuously. 


The key with AI is that these predictive models are not fixed. Machine-learning algorithms automate the process of finding patterns in data. As new data becomes available, the algorithms learn and adapt continuously. 

They require less human involvement over time and are able to deliver increasingly effective analysis. The entire system is constantly learning as fresh data is introduced or as some data restrictions imposed by the programmer are relaxed. When applied to previously unused data, the model can be used to anticipate patterns in that data. This process allows the user to test the accuracy of the machine-learning system. 

Return to top

AI system to detect accounting fraud

The AI system underlying the Transparently.AI Manipulation Risk Analyzer (MRA) is trained to look for patterns associated with accounting manipulation and various types of fraud. It does this by identifying these patterns in known historic cases of serious accounting manipulation, fraud and resulting corporate failure. 

Although every past case of fraud is different, the relationships among accounting variables are constant. Thus, even though different strings are pulled when accounts are manipulated, similar anomalous patterns in the relationship among accounting variables will result. 

These patterns are not linear and they are not necessarily obvious to the naked human eye, but they are present in the complex models woven by the machine learning algorithms.

The MRA allows the user to analyse hundreds or even thousands of companies in a single session, and thus can greatly facilitate the avoidance of potentially fraudulent companies. 


Traditional methods of detecting fraud are often time-consuming and costly, and many instances of fraud go undetected or unprosecuted. 

However, AI-powered solutions like the MRA have the potential to revolutionize the detection and prevention of accounting fraud, offering a more efficient and effective approach compared to traditional methods. 

AI algorithms continuously learn and adapt as new data becomes available, improving their accuracy over time. With the ability to analyze vast amounts of data and detect complex patterns, AI is a powerful tool in safeguarding against accounting fraud and promoting financial integrity.

Return to top

Subscribe to our blog