Whether or not you're on the artificial intelligence (AI) bandwagon, you'll want to pay attention to the rapid advances in the generative pretrained transformer, better known as GPT, the large learning model used by ChatGPT.
About half of all tasks U.S. workers perform could be completed faster by generative AI without losing quality, according to a 2023 study by University of Pennsylvania researchers and OpenAI, the company that developed ChatGPT. Many accounting firms are open to more automation. Already, 24% of top-performing client advisory services (CAS) practices use AI, according to the 2022 CAS Benchmark Survey by CPA.com and the AICPA's Private Companies Practice Section (PCPS).
That percentage will surely rise because AI technologies such as ChatGPT can help CPAs do more work more accurately with fewer people. This, in turn, can help the profession tackle pressing challenges related to production capacity, staffing shortages, and accountant burnout.
Based on inputs or prompts, GPT guesses what should come next. These guesses are trained on massive data. The most recent version, GPT-4, reportedly was trained on more than 1 trillion parameters.
That said, GPT can't be original; it cannot develop new ideas. It is, by its nature, derivative. It can write, but it isn't a writer. It can create, but it isn't creative. It can tell you a joke, but it will probably be a joke you've heard before, or just a bad one. GPT is pretty much a sophisticated version of autocomplete.
But if a large part of your firm's work is inserting numbers into boxes, I advise you to stay tuned.
How GPT is being used today
GPT has become very adept at certain repetitive business tasks. Here are the 10 most common:
How many of these tasks do you and your team spend time on regularly? Caleb Newquist, a former CPA and the editor-at-large at Gusto, a tech company that offers HR software, told me recently that GPT "quickly figures things out with basic prompts. It starts to do them. It starts to follow your instructions. That's not very different from a first-year accounting associate."
Kashif Ali, CEO of TaxGPT, which automates tax filing, said he's seen his productivity increase tenfold using GPT. Ali likens it to having "a second brain or copilot" that you can talk to and that can troubleshoot a problem with you.
Chris Vanover, CPA, president and senior chief auditor at AuditClub, said audit seniors and associates can use GPT as a de facto senior manager or audit partner when they're going through an audit. Which would you rather have, Vanover said, the traditional approach: "Hey, can I talk to you at the end of the day? I have a bunch of questions for you," or asking an AI tool: "Hey, I'm going through these particular steps. Can you help coach me or navigate me through things that I should be considering when I'm executing these procedures?" Most people would choose the latter, Vanover said.
Aaron Harris, global chief technology officer at accounting and business software maker Sage, told me that ChatGPT is really good at writing application programming interface (API) calls against the Sage Intacct API. Harris said he's quickly learning the Python programming language just by asking ChatGPT to help him code.
From where I sit, if we can use AI to write API calls, which allow different applications to communicate with one another, think about how that could empower us to move data around without relying on developers to build every single connection manually.
Alleviating staffing shortages and burnout
Citing alarming U.S. Bureau of Labor statistics, The Wall Street Journal reported last December that 17% of accountants and auditors — one out of six — quit the profession in 2020 and 2021. That's 300,000 accountants and auditors exiting the profession with not nearly enough new entrants to replace them.
AI technology can help. When we think of labor, we think of workers toiling away at their desks. We rarely look at productivity, that is, getting work done faster and more efficiently. If we can increase productivity substantially through AI-powered automation and workflow improvements, then we don't need nearly as many workers.
Ask anyone who quit our profession why they left and most often they say it was because they felt burned out. Quite simply, there was way too much work to do and not enough staff to help them. Unfortunately, this mindset becomes a doom loop because as more people leave a firm, the ones who remain get slammed with even more work until they too get burned out and leave.
Even though accountants were among the first professionals to welcome computers, we didn't use the technology to its potential.
Consider how a tax practice operates in most firms. Workflow is often managed with spreadsheets. Instead of writing down numbers, they are being manipulated in Excel. In audit and tax, we have routing sheets. Those sheets used to be on the outside of a folder, and the folder had all the paperwork in it.
A lot of workflow software is still based on those old paper routing sheets. The difference is that now it's digitized. In some ways, it was more efficient when we just had all the paper in a file folder; at least it was easy to pick through and retrieve. Now we have to search through digital file folders of PDFs, and it's often harder to find things. A Future of Work survey conducted in 2022 by the Virginia Society of CPAs among its members found that about 50% used workflow software.
That's not automation or efficiency. It's what Cornell University professor Louis Hyman calls the "productivity paradox," which results from digitizing paper-based processes without substantially improving them.
AI technology represents a breakthrough from the productivity paradox.
Ashley Francis, CPA, a trust and estate expert, told me she's found GPT helpful to generate client handbooks, create tax workpapers, and pull IRS publications. She can also envision using AI products to write tax memos. "That technology is coming," she said, referring to CoPilot in Microsoft products in partnership with BNA's plugin.
"Imagine going into the BNA portfolios and instead of having to read all of the portfolios, you start outlining your client situation," she said. "It's in its infancy but will be advancing [fast] and probably be mature in six months."
Research for a 12-page tax memo can take 40 to 80 hours and Francis believes AI will substantially reduce that time for her, once it is integrated into existing tax research tools. "It's like having a very well-educated intern who you would never trust to actually send [the memo] to the client," Francis said.
I have found ChatGPT to be an incredibly powerful and speedy tool. Just make sure you verify the sources and conclusions it comes up with before sharing its findings with clients, strategic partners, or staff. ChatGPT is still a work in process and the margin of error can be quite high.
Over the past 10 to 20 years, the big benefit of cloud-based accounting has been keeping all our data in a secure place, where we can access it 24/7 and easily move it between systems. Sure, there are rules to automate that, but the rules are limiting because you can only set up so many. But now, with GPT, we can have "fuzzy rules."
That means you can state in plain English what you want GPT to do. It can follow your instructions with a reasonable degree of accuracy. As a result, I don't have to make a rule such as, "Only approve expense reports below $10 automatically." Instead, I can provide the AI with a list of historical expenses flagged as material or immaterial. I can then prompt it to "Only approve immaterial expenses automatically." Based on the historical dataset, the AI technology can infer the meaning of "immaterial" in context and automatically approve or deny future expenses without a rule. It can also be prompted to provide a confidence score for each decision, which can be used to flag questionable expenses to a human reviewer for a second opinion.
The feedback from the human reviewer is then used to train and improve the model, increasing its confidence for each subsequent transaction.
Sage's Harris has even bigger plans for AI technology. "Why do you need approval levels when AI can be smart enough to identify when something should have a human review?" he asked. He believes there will be AI agents that learn to take input from the business leaders the same way a staff accountant takes input from the manager of the accounting team, then follows those very high-level job instructions.
"The new generation is going to be AI agents taking ownership of a specific workflow or part of the business," said Harris. "You'll interact with them in much the same way you would interact with a junior accountant."
Vanover said his firm had been in discussions with a member CPA firm going through a Public Company Accounting Oversight Board (PCAOB) inspection. Manual journal entries came into focus because they are a highly challenging area for firms to address. They are always connected to a significant risk related to the potential for management override of controls. This particular CPA firm pulled an extract from the general ledger and 22,000 entries were flagged as manual journal entries. No human could go through every one of those and say whether they were indicative of fraud. But AI can do that.
"And that's just the power of being able to synthesize that data and tell you exactly where to focus," Vanover said. "That's where we're headed."
AI technology is not going to fade. It will be as impactful as the internet. And the best thing you can do right now is to understand it and learn how best to use it. Sign up for a free account with ChatGPT or next-generation AI assistant Claude.ai and start asking questions.
— Blake Oliver, CPA, is the founder of Earmark, which produces accounting and tax podcasts, and co-host of The Accounting Podcast. To comment on this article or to suggest an idea for another article, contact Jeff Drew at jeff.drew@aicpa-cima.com.
Resource
CPA.com Generative AI Initiative: This multi-part project features a toolkit, security and risk considerations, a glossary of key terminology, and video walk-throughs showing how to build prompts.