AI Demystified for Detecting Errors and Omissions in the Monthly Financial Close Process

Every month organizations go through a stressful period during the financial close process, where they validate account balance and transactions for the month. The finance teams follow this process with limited resources and time available to perform an accurate analysis of the data. Hence, the finance teams during the monthly financial close period are overburdened as there is never enough time to validate everything, so rules of materiality are applied.

Research shows that 40% to 60% of the time accountants spend on the close process is looking for errors and omissions in transactions and account balances, which we also refer to as anomaly detection. Ideally, anomaly detection should be an everyday process, but usually many organizations do it after the month-end close starts.

Accounting transactions are very pattern-based, and the finance departments handle certain transactions every month or quarter. For example, vendor invoices are typically charged to the same account segment/account combination, and revenue and write-offs are typically charged to a similar account segment/account combination. An everyday process of looking at the patterns and identifying the outliers (anomalies) can enable them to get corrected before the month-end period starts, which will not only reduce the burden on finance teams during the close process but also improve the accuracy. Read our eBook, Best practices for improving month end close accuracy, and find out more about the challenges that finance teams face while reviewing large amounts of transactions every month.

There are several different AI solutions available for detecting fraudulent transactions, banks use them, retail companies use them on their cash register transactions, and many other companies use them as well. Identifying errors and omissions in accounting is just an extension of those methods. Identifying Errors and Omissions is a perfect use case for the Artificial Intelligence process because accounting is very pattern based. Here are some reasons why:

Algorithms interpret patterns: AI technology uses computer algorithms. The software programs aim to mimic the human ability to learn, interpret patterns and make predictions. The key to identifying errors and omissions is “interpret patterns”. For identifying errors and omissions, “Machine learning” is a form of AI that uses data to train algorithms to recognize patterns and identify outliers/Anomalies.

Flagging errors: AI technology looks at the sequences and patterns of invoice numbers. If an Invoice falls out of pattern or comes in at a different time it could be a manual invoice and a duplicate billing, which will be immediately identified and flagged as an error/anomaly.

One might wonder, what is unique about AI algorithms and how is it different from rules?

Algorithms apply different dimensionality, tolerances, and focus to identify outliers that translate into anomalies. It isn’t a magic algorithm that finds everything but a series of algorithms that focus on specific AI-derived dimensionality and tolerances. It learns as data sets roll forward and the data is enhanced with the actual corrections applied vs anomalies detected.

Rules tend to be discreet “If/than/else” logic to find errors. Algorithms go beyond “If/than/else” logic to offer much border scope that is needed by Anomalies, replacing 100’s of rules logic.

In comparison to AI-based algorithms, rules bring many challenges trying to identify anomalies:

Difficult to Scale: Rules are discrete for dimensional combinations, you could easily have +1000 rules.  Management of those rules will become a problem and the system will stop reacting to newly added vendors/customers/accounts.

Scope:  The task of identifying unusual transactions follows two different threads:

• First identify all the usual transactions, everything left over is flagged as unusual
• Second identify unusual transactions based on what the system has never seen before (based on what looks unusual)

Rules don’t offer a broader identification of anomalies and a lot of individual rules need to be updated to cover the many different types of unusual transactions. While both rules and algorithms apply structured inputs and logic, rules tend to be discrete and challenging. 

With rules, finance teams need to manually intervene and maintain 1000+ rules for each and every tasks, which include: 

• Vendor tolerances: Reviewing each vendor and determining manually the tolerances of payment ranges on a monthly quarterly/annual basis and then creating if/then/else logic to check every vendor for those tolerances
• General Ledger Entries: Imagine reviewing each vendor and determining the standard ways it is applied to the general ledger (accounts, account segments, etc.) and then creating if/then/else logic to check every vendor for those tolerances.
• Balance sheet accounts: Imagine reviewing all the balance sheet accounts and the normal source of the transactions (A/P, A/R, Fixed Assets) and then converting that to if/than/else logic looking for transactions that are outside of the normal system/account match up.

Vendor payments and invoices: Imagine reviewing all vendor payments and the typical invoice number pattern to understand if an invoice is a manual invoice and then cross referencing that against other vendor invoices and payments to check for a duplicate payment.

Machine Learning can address this challenge with rule discovery 

What if there was a Machine Learning process that does all the rule discovery based only on those different sets of dimensionality and then takes those patterns and lists and processes them through an algorithm that:

• Analyzes 12-18 months of transactions history from General Ledger transactions, A/P system transactions, etc.
• Classifies the normal (usual) and abnormal (unusual) transactions based on the transaction history analysis  
• Creates patterns of abnormal transactions and flags them as well
• Updates the rules based of the transaction history analysis

Classification algorithms can be  broad, one algorithm focused on specific dimensionality such as time, amount, invoice number, vendor, etc., can be used to replace hundreds of discrete rules.  One algorithm can be used to identify missing payments, duplicate payments, missing journal entries/duplicate journal entries, etc. The concept of Rule Discover is a very simplified example of machine learning.  Think of learning algorithms as creating a list of usual transaction dimensional combinations.  Those algorithms provide the input of a related algorithm that looks at the current day’s transactions and identifies the outliers. 

Hence, AI and ML can be the perfect financial close management technologies for detecting errors and omissions; and overall improving the General Ledger entries and data accuracy. Watch this webinar by Ron Baden, VP, CFOTech, HighRadius, and learn more about how to apply AI in accounting with some real use cases.