Recapitulation of Generative AI Use in Accounting

Recapitulation of Generative AI Use in Accounting

Over the past seven-series articles, which began on January 5, 2024, we considered generative AI use in accounting: more precisely the promise and the practice of generative AI capabilities, the risks of generative AI use in the accounting practice, accountancy standards and regulatory bodies, the policies of those bodies regarding generative AI, the implementation of a generative AI (LLM system) in the accounting practice, custom API scripts and plugins for generative AI in accounting, for increasing the scope and scale of use, and private database implementation of generative AI in accounting, particularly for obviating the risk of confidential data leakage.

In the initial article “Generative-AI Capabilities, Promise, and Practice” we saw that the range of generative AI capabilities, from routine to sophisticated analyses and decision-making, lend themselves to use in professional services firms like an accounting practice. Consequently, the emergence of these generative AI capabilities is timely to mitigate the current shortages of skilled professionals in professional services industries, including the accounting industry.

In the next article, “The Risks of Generative AI Use in Accounting Practice” we noted that the risks applicable to the use of generative AI systems are the same ones applicable to the accounting practice. But there are specific compliance risks, associated with the practice in the USA, as there are in other jurisdictions, relating to the accounting profession as dictated by the Securities and Exchange Commission (SEC,) the Internal Revenue Service (IRS,) Financial Accounting Standards Board (FASB,) and Public Company Accounting Oversight Board (PCAOB)—a scion of the Sarbanes-Oxley Act of 2002. A derivative implication is that only CPAs or a registered public accounting firm can perform audits and prepare financial statements for publicly traded companies:

Then, we noted a substantial number of US standards and regulatory bodies, in the article “Accountancy Standards And Regulatory Bodies,” that can have a say in using generative AI in accounting. Their varied authority are derivatives of the securities acts of the 1930s. But, in the cases of the Internal Revenue Service, an agency of the Department of the Treasury, and the Federal Reserve System, the former manages government revenue, gained its authority from the Act of Congress in 1789, and the latter manages the US money supply, make loans and provide oversight to banks, and serving as a lender of last resort; has its authority from the Federal Reserve Act of 1913 and the Banking Act of 1933 (Glass–Steagall Act).

Following, in “Accountancy Standards and Regulatory Bodies Policies on Generative AI,” we found that “wait and see” can best describe the current policies of accounting regulatory and standards bodies regarding the use of generative AI. This is despite its use by some of these bodies in their operations. It is not required to say, that their delays are due to the limits imposed by the law and want of a full understanding of the technology and its ramifications.

In the article “Implementation of Generative AI (LLM System) in The Accounting Practice” we showed that with manual operations on the LLM, an accountant can complete a large part of his/her routine work, such as data preparation, tax preparation, inference, report composition, &c. Even so, manual operations are cumbersome for large-scale activities such as automatically retrieving journal entries from a database for producing financial statements and other reports. Notwithstanding, improved manual operations in ChatGPT, in general, and in accounting particularly, can be facilitated through the use of WebChatGPT or ChatGPT File Uploader (places a file upload button on the interface, 15,000-character limit/upload, especially useful in ICL or data prep,) or ChatGPT Sidebar & File Uploader extensions for your browsers.

Thereupon, in “Custom API Script and Plugins for Generative AI in Accounting” we noted that custom API scripts, for interacting with the LLM, are generally coded in Python and PHP programming languages; that plugins are off-the-shelf or ready-made API scripts available from various vendors and are accessible from the Plugin Store in the ChatGPT UI. Both API Scripts and plugins advance the use cases of ChatGPT in scope and scale. For example, instead of manual cut-and-paste operations, the accountant can upload complete files of various formats for “processing” via prompting. So, the accountant can build his/her API script that focuses on accounting or enable specialized accounting plugins from ChatGPT UI such as QuickBooks amongst others.

In the penultimate article “Private Database Implementation of Generative AI in Accounting” we introduced practices/solutions, i.e., anonymization, and encryption of data, as well as, vectorization, for use of private databases that protect confidential information with the additional benefit of functioning at scale.

Anonymizing data on the database means removing or replacing personally identifiable information (PII) from the data, such as names, phone numbers, email addresses, &c., during in-context learning with the LLM. Encrypting data from the database means transforming data into an unreadable form using a secret key or algorithm, also, during in-context-learning with the LLM, thus preventing unauthorized access to private data. Only those with the key or algorithm can decrypt the data and restore its original form.

The use of a vectorized database means converting data into numerical vectors that can be used for mathematical operations and machine learning. A vectorized database can also help leverage the general knowledge and information in the data without compromising the data subjects’ identities; and it is used as an adjunct to an LLM, which does not and cannot memorize the vectorized data.

Vectorization is not necessarily superior to anonymization or encryption, but rather a different approach that has its advantages and disadvantages. All accomplish the task of protecting personal identifying information (PII) and, particularly, avoiding data leakage, and misuse. Depending on the use case and the requirements, one may use one or all methods to protect the data.

Undoubtedly, the accountant can benefit from the use of generative AI. Although he/she must be aware of the risks and ensure that their operation complies with established regulations regarding such things as CPA licensing, maintenance of client confidentiality, accuracy, &c. Furthermore, instead of manual engagement with the LLM, which can be tedious, the accountant can choose automation using scripts and plugins whilst engaging with the LLM; and can improve the confidentiality of client data by utilizing anonymization, encryption, and vectorized databases.