Artificial Intelligence (AI) is making us question the way we have been performing work in almost all sectors, and accounting roles are no exception. While it is generally accepted that AI can improve efficiency by automating recurring tasks, the use case differs when we examine automation in a core accounting domain. What makes accounting a uniquely challenging opportunity is the rigid application of US GAAP guidelines, and the cognitive thinking required to interpret business transactions overlaying GAAP interpretations, to post auditable accounting entries. With the advent of sophisticated AI assistants, we may soon be able to merge cognitive AI into US GAAP interpretations, breaking down complex business transactions into accounting recommendations for accountants to consider while maintaining books and records.

Application, challenges, and opportunities of AI in core accounting

Within the core accounting domain, certain use cases are recognized as having “unresolved” solutioning opportunities. One common challenge concern processing recurring payables. The problem is that organizations—particularly those in consumer lending, mortgages, manufacturing, logistics, and entertainment businesses—deal with voluminous invoices, in most cases due to recurring expenses such as Rent and Leases, Utility, Travel Reimbursements, Valuation and Appraisal Fees, etc. Such expenses are usually coded to a specific general ledger account with worktags or dimensions like cost center, expense approver, and others.  While the job of AI models could become easier if every vendor transitioned into an electronic billing solution that sends digitized invoices to customers, we might be at least a decade away from that technologically matured state. Until then, AI models built to integrate optical character recognition (OCR) to automate data conversion, and a training platform to map Vendor-Expense definition – billing date to US GAAP acceptable codification, can empower accountants to embrace the power of next generation models.

In seeking a solution for this problem, platforms such as Bill.com® and Expensify® have already achieved a certain degree of automation, but these are still considered as Narrow or Weak AI. Such models operate in a limited context and cannot perform tasks beyond its predefined capabilities.  

Consider a business case where an Accounts Payables (AP) team receives an invoice for a purchase of office furniture or the installation of a server. A trained accountant will perform the following steps before concluding accounting treatment:

• Review the nature of each expense and reach out the department that placed the order
• Confirm if the invoice amount is close to the billing estimate, if there is one
• Receive approval from the respective department for payment of invoice
• Once the invoice is approved, the accountant will post a Journal Entry to record Expense/Asset and Liability; this phase involves cognitive judgement which also aligns closely with application of GAAP principals such as:
  • ASC (News - Alert) 606 – Revenue Recognition:– when to recognize expense; should expense be allocated to multiple periods. Recognition of expenses occurs at the same time as corresponding revenue, so it becomes important to factor ASC 606 into expense recognition decisions.
  • ASC 360 – Fixed Asset Accounting: is the expense within a policy threshold that allows expensing vs capitalizing? If capitalizing, setup a depreciation schedule.
• Output: Posting an auditable Journal Entry is an expected outcome from this process.

Translating the above steps into an algorithm, to perform tasks that can be trusted in the financial world, will require a step up from Narrow AI. Perhaps General AI (also known as Strong AI) might be what accountants need. General AI’s possess human-level intelligence, capable of performing intellectual tasks that a human can do.

While that is part of technology solutioning, the parallel challenge to achieve this milestone is the need to standardize processes across the organization.

For example, a mortgage lending firm introduces AI-powered accounting software in their accounting department and initially expects to achieve a run rate productivity of 1.6x per employee. The group soon realizes that there were more exceptions to the process which needed manual intervention. Their cost center budgets were not systemically documented, so at month end there was a need to review actual costs for each cost center and process a journal to reclassify expenses from one cost center to another. The savings achieved initially through automation were diluted, due to the manual reviews at the month end and the effort needed to process journals.

This example illustrates the imperative for organizations to relook at their processes and revamp their policies to make their systems more AI-friendly.

The 2024 Thomson Reuters survey on AI revealed that nearly 51% of businesses are using or considering transitioning into AI tools.  This presents an opportunity for accountants to elevate themselves from doers to strategic design agents. Accountants possess valuable insight on the start to finish of a transaction lifecycle. A smart organization will empower its accountants to participate in process standardization and infrastructure-build initiatives which are key to implementing AI.

AI use in determining the correct accounting treatment for a trade or a transaction

While there are scenarios that describe how exactly an AI assistant can help output the accounting treatment for a trade in general, an accountant can derive larger benefit if regenerative AI models can be integrated in a real-life business scenario that could recommend Debits and Credits. In the Payables example mentioned earlier, the current Enterprise Resource Planning (ERP) platforms need accountants to key in one leg of the account that should be debited, while the other end is automatically configured. In this instance, the accountant will go through the invoice and identify the type of expense and associated general ledger account (i.e. professional fees, legal fees, custody fees). The accountant will then navigate to a payable module in an ERP platform, and assuming header level information is already captured by AI, will select the expense account to which the entry should be debited to, with a credit automatically hitting the AP ledger.

This is a classic case that demonstrates where we still need a human accountant’s intervention in determining the type of general ledger (GL) account to book an expense to, thus interjecting a break in the full cycle of automation. In some cases, instead of selecting a GL account, the ERP is configured to select the type of expense, which is then mapped to a ledger account using a mapping library. In both these cases, there is a need for accountants to divert their valuable time to make a determination.

A General AI tool would come in handy here. A tool that encompasses capturing “critical to process” components from an invoice through OCR, then utilizes that data to run a lookup with organizational policy, including a search with GAAP or IFRS codifications, pushes out invoices to approvers, converts the data into debit and credit with limited human intervention can lead to real productivity gains in the field of accounting. There is a huge opportunity here, in that generative LLMs could build a data lake of how transactions have been treated historically within an organization and use that data as an anchor to process transactions of like nature.

Accountants often end up with questions such as, “How was this trade booked historically?” so a precedence could be used for auditability. A SOC Type I and II- certified LLM could promise a consistent financial result, making “apples-to-apples” comparison analogous in an investor reporting paradigm.

PWC’s GL.ai comes to mind when thinking about  an audit enabler bot. GL.ai algorithms are trained by experts from across the company’s global network to identify unusual transactions. The bot also has the ability to triangulate every line of journal, together with the behaviors of the people posting those journals.

We have a good use case of a similar tool in internal accounting reviews, which accountants can use to analyze entries posted to LLMs to get additional comfort around transactions that hit an income statement or balance sheet. This would be a clear-cut state of elevating accounting roles from mere debits and credits to strategic contributors.

AI in Treasury

Are accountants confident enough to delegate cash transactions to bots? Treasury management involves payment processing, collections, cash flow forecasting, liquidity, and debt management among its primary functions. One of the high-risk areas in treasury management is payment processing, in which an organization is initiating cash outflows impacting its liquidity. There is no question that  LLMs have the capability to handle huge volumes of data and  convert them to actionable business intelligence. In this case, a well-trained bot can look at historical cash needs relative to sales and purchase projections, and calibrate a working capital model for the accountants to integrate into their decision-making process. But has the LLM infrastructure matured enough to handle payments themselves, without a human eye verifying the outflow?

Almost all organizations have a treasury analyst who manually reviews daily cash outflows, such as invoices or funding wires, and compares the request to supporting documents. In their review, the treasury analyst or accountant will verify banking information for new beneficiaries by placing “call back verification” to the beneficiary contact provided in either invoice, funding documents, or any other supports. This is a risk mitigation step in light of increasing cyberthreat exposure.

While it does look like parts of treasury functions can be handed over to bots, it might take more proof of accuracy before handing bank accounts over to AI agents to initiate payments on their own. Perhaps this can still be achieved with a hybrid model that involves human + bot collaboration.

Within Treasury, AI bots can be useful in areas such as:  

• Guidance on selecting payment types. An AI tool can analyze information such as the size of a particular payment, how fast it must be delivered, and the payment types the supplier accepts, and then recommend the most appropriate payment type to use. For example, if AP needs to deliver a payment today and the supplier accepts payments via the RTP® network or some other instant payment method, that instant payment type would be the likely recommendation. If the treasurer has three days to make the payment, the system might suggest ACH or some other method.
• Intelligent email automation to support faster customer service. Intelligent bots in large businesses are reducing response time to vendor inquiries. Treasurer and AP teams generally receive high volumes of inquiries about the timing of payments. The bots review incoming customer emails, and assign inquiries to respective analysts. If the vendor inquiry pertains to the payment status of an invoice, the bot assigns the email to the payables team. But if the inquiry relates to issues like ACH or Wire bounced, then bot assigns such queries to the Treasurer.  This reduces the float time of an email, since the query lands directly in the right department for appropriate handling.
• Cash flow forecasting.  LLMs have successfully proven the ability to use historical data to project future outcomes, offering a starting point to execute a scheduled forecast process for planning and decision-making.  Forecasting LLMs use time-series AI algorithms for analyzing and modeling forecasting data over time, detecting seasonality and trends, and enhancing forecasts with automatic outlier detection.

There are sufficient use cases in a Treasury function for AI to improve efficiency. For instance, one critical activity in payment processing is “call back verification.” Call back verification is a step in payment processing wherein the accountant or treasurer calls the beneficiary to confirm banking instructions as provided in either the invoice or funding documents. This is a security measure used by almost all businesses to safeguard themselves against bad actor interventions in regard to payment instructions received. A publicly known case is that of the Toyota Boshoku Corporation, a subsidiary of Toyota, which in 2019 fell victim to a vendor scam. Cybercriminals impersonated a legitimate business partner and tricked the company into transferring approximately $37 million to fraudulent accounts.

Call back verification became increasingly  popular as an aftermath of cyber risks in payment processing. This area might continue to be a “human job” until at least such point that AI can guardrail itself against cyber risk exposure.

There is a vast opportunity for businesses to adopt AI enabled by a standardization of business functions. AI also opens the door for accountants to elevate themselves into subject matter expert roles, helping technology teams to design practical algorithms and play an oversight role in training models to adopt to changing business requirements. Nevertheless, AI will still have to mature in certain areas before bots can totally replace a human presence in handling critical functions.

Author:

Sumangali Krishnabhat is Senior Vice President, Controller of Redwood Trust, a publicly-traded real estate investment trust (REIT) specializing in housing credit. Ms. Krishnabhat has worked for more than two decades in accounting, finance, advisory, and private equity roles with some of the world’s largest consulting and fund management companies. As a senior executive she has developed special expertise in Fintech implementation, financial analysis and stewardship, debt facility management, taxes, audits, and technical accounting, liquidity planning and management, capital optimization, due diligence, reporting, and regulatory and credit compliance. In India, she earned a Bachelor of Business Management from Bangalore University and received an MBA in Finance from ICFAI University. She now resides in California and holds an active CPA license from the California Board of Accountancy and an active CMA credential from the Institute of Management Accountants.