Accountants can choose from a plethora of new ways to apply artificial intelligence (AI), as capabilities for language processing, decision-making, and automation are more accessible than ever.

“There are so many new tools,” said Tricia Katebini, CPA, a partner at GRF CPAs & Advisors in Bethesda, Md. She has helped the firm implement numerous new AI-based tools. “It’s nice to see, but it’s overwhelming.”

To help cut through the hype and the array of new options, the JofA asked six CPAs to share the most interesting use cases they had found for AI, from customized solutions to off-the-shelf software products (see the sidebar “How to Determine AI Technology Needs”).

Here are their stories.

CASE 1: SNAIL MAIL AUTOMATION

Karl Spanbauer, CPA, is controller of the not-for-profit Capital Area Food Bank, which provides about 60 million meals annually for people around Washington, D.C.

He had one major annoyance after taking the job: the mail. Not email, but actual physical mail. His small accounting team was inundated with vendor invoices, tax notices and statements, and more.

“I lost a couple documents. It was literally hundreds of pieces of mail,” Spanbauer said. “I knew we had to automate it somehow.”

Drawing on his experience with Microsoft Power Automate and other software, Spanbauer built his own solution to scan, summarize, and respond to mail.

Here’s how it works:

1. Staffers run all business-related mail through a scanner.
2. The scanned file is automatically saved to a shared drive.
3. An image analyzer extracts the text of the image.
4. The text is analyzed by a secure large language model (LLM), which generates a summary.
5. A ticket is opened in Jira, the team’s workflow system, which includes the summary and the original mail.
6. The team tackles the tickets, responding to the mail as needed.

Spanbauer has added further automations to the system, making it easier to look up the status of scanned invoices and carry out other follow-up actions.

Case 1 at a glance

• Degree of difficulty: Spanbauer relied on deep experience with Power Automate and other tools. The solution was quick to implement but was built on top of months of work to connect systems and data.
• Cost: The project didn’t bring any additional costs to implement; it relies on the team’s existing licenses.
• Payoff: The team now tackles the mail in 20-minute sprints for “Mail Mondays,” saving about four hours of total staff time per week.
• The takeaway: The project worked because Spanbauer already had a robust digital environment — he just needed a way to bring the postal mail into it. “AI can’t interact with a piece of paper sitting on my desk. I had to get that piece of mail into that ecosystem,” he said.

CASE 2: TECHNICAL ACCOUNTING MEMOS

Glenn Hopper is head of AI research development at Eventus Advisory Group in Memphis, Tenn. His firm serves a mix of private companies and small-to-midsize public companies with services including fractional CFO and controllership roles.

Among his recent projects is a specialized bot that can draft technical accounting memos.

“We get asked to do these accounting memos all the time,” Hopper said. For example, a memo might take on the topic of goodwill impairment in an acquisition. It’s a challenge because, he said, “You have to get very technical, and you want to make it clear and understandable.”

Using his expertise in generative AI, Hopper took the organization’s trove of accounting memos and prepared them for analysis by an LLM through a technique known as retrieval-augmented generation, or RAG.

“We’ve been writing these forever,” Hopper said, “and we have a style we use for them.”

This involved dividing the memos into smaller “chunks” that are more topically focused and easier for an LLM to process. Hopper then ran these chunks through a process known as vectorization and stored them in a vector database using technology from OpenAI.

The end result: a store of content that the bot can tap to create a “coherent and contextually relevant response,” Hopper said. The memo generator is powered by OpenAI’s Assistants API platform, which allows companies to create advanced, customizable bots.

Hopper added that he tried to “pound out any possibility for hallucination,” referring to generative AI’s tendency to make up facts when it can’t find the information it is seeking. “I think we’ve minimized it a lot by giving it very specific examples to refer to, very specific instructions,” he said. That process of “fine-tuning” involves exposing the customized bot to hundreds of examples, though Hopper said he was unable to divulge much more because the specifics are proprietary to Eventus.

Once the tool generates an accounting memo draft, Eventus keeps a “human in the loop,” reviewing and correcting the memos as needed.

“The people who are using it are kind of blown away — and for many of them, it was their first foray into even using ChatGPT,” he said.

Case 2 at a glance

• Degree of difficulty: This was a highly customized implementation of OpenAI’s technology, requiring about 60 hours of Hopper’s time and deep expertise.
• Cost: Beyond the cost of ChatGPT licenses, the memo-generating bot and other advanced bots can incur costs of several dollars for each run. The paid versions of ChatGPT start at $20 per month for individuals and $25 per user for team settings.
• Payoff: Hopper estimates that accounting memos have been taken from a four-hour task to a 30-minute task that includes careful human review. Hopper also has created other bots that can tackle tasks like analyzing Forms 10-K.
• The takeaway: The memo bot was a win, but it wouldn’t have been possible without Hopper’s expertise. For many other organizations, it may be more cost-effective to wait for a vendor to develop an off-the-shelf product. “It takes time to do, and you need to have a certain skill set to do it,” Hopper said. “It’s not cheap to get these things done.”

CASE 3: MARKETING THE BUSINESS

Barrett E. Young, CPA, is marketing partner at GWCPA in La Plata, Md., which is focused on helping family-held businesses prepare for a transition of leadership. Since joining the firm in 2017, Young has led the team of 15 to make better use of technology, including ChatGPT and related technologies during the past year.

The firm has even deployed its own client-facing tool: the GWCPA Generations Advisor, a customized ChatGPT deployment that gives advice on ownership transitions. This custom GPT requires an active ChatGPT license but does not incur a per-use cost, as it is not as complicated as the Eventus bot described in the previous section.

“I’m using ChatGPT every night when I have ideas in my brain at two in the morning,” Young said. “I’m asking it questions and making notes, ‘What if I gave that to my customers?'”

Developing a custom GPT of this type is relatively straightforward: The creator supplies written instructions to ChatGPT, including factors like the purpose of the bot, the tone and structure of responses, and other instructions.

The goal was to create “something I could give away to our target audience that would start conversations and add value in their lives,” Young said.

Young enhanced the custom GPT by uploading GWCPA’s marketing materials. Custom GPTs can be instructed to refer to information or to imitate the style of user-uploaded documents — another version of the RAG approach. To get access, users simply sign up for GWCPA’s free marketing email list, which Young also uses to send weekly prompt ideas for readers. The hope is that the custom GPT will answer basic questions about succession planning — raising awareness of GWCPA and allow for deeper conversations when they’re ready to meet with the team.

Case 3 at a glance

• Degree of difficulty: Generations Advisor was relatively quick and easy to implement for Young, who already had created a few custom GPTs for personal use. These simpler forms of custom GPTs can be created within the ChatGPT interface through plain-language instructions.
• Cost: Creating a custom GPT requires a paid ChatGPT subscription, which costs $20 or $200 per month, depending on the level of access needed.
• Payoff: Subscribership has steadily grown for Young’s mailing list, with the custom GPT responsible for about 50 new members the past three months. It has steadily attracted a handful of subscribers each month. While the audience is relatively small, Young says it’s reaching the right people. “One or two clients from that makes the entire thing worth it because we’re high-value, high-touch services,” he said.
• The takeaway: Generations Advisor is a light customization of the core ChatGPT project, tapping into the platform’s abilities to create a useful tool for clients that Young hopes will grow interest in the firm’s services.

CASE 4: DATA ANALYSIS WITH AI-ASSISTED CODING

As a consultant for executives and companies, Don Tomoff, CPA, a director at Invenio Advisors in Cleveland, has used ChatGPT for coding and analysis, saving time and money and allowing for more efficient project execution.

“I used to outsource coding for projects, and I haven’t had to do that since January of 2023,” Tomoff said. “That’s been just a huge win.”

He now asks platforms like ChatGPT to generate code for tasks in Excel and other software. For example, he recently needed code to seek out a particular piece of text and delete all the rows above it in an Excel spreadsheet, no matter where it appeared. Tomoff described the problem in natural language to ChatGPT, and it returned functional visual basic (VBA) code. Other models like Anthropic’s Claude or Google’s Gemini have similar capabilities. To see a prompt used with Claude, see the sidebar “Claude AI Prompt for VBA.”

Tomoff breaks his requests into multiple stages, especially with larger projects. He asks ChatGPT what steps are required for a particular task, then instructs it to write code for each step. If the code proves faulty, he can describe error messages and ask for revisions.

Outsourcing coding previously cost up to $150 per hour, or up to $2,000 for a typical project, Tomoff said. Moreover, the time to develop, test, and improve solutions has dropped from days to hours, he said.

“You want to explain as precisely as you can what you want,” he said. “You don’t want to eat the elephant in one effort, because it will stop on you if you give it too much to digest. It’s iterative.”

Brianne Smith, CPA/PFS, Ph.D., a financial adviser, managing member of her own accounting firm, and assistant professor of accounting at Auburn University at Montgomery, Ala., offers similar advice to her accounting students as they use ChatGPT for four phases of data analysis: “Ask the question, master the data, perform the analytics, share the story.”

Smith explained: “You wouldn’t load ChatGPT and ask it to do all four things at the same time — but we would focus in on various areas and chunk them out into intermediary goals at the end.”

ChatGPT and other tools are also capable of analyzing and transforming files directly, not just generating code for users to execute.

Case 4 at a glance

• Degree of difficulty: Low, as long as the user has some familiarity with generative AI bots. The key is describing the problem in specific terms and having the patience to experiment and troubleshoot.
• Cost: Code generation is possible with the free versions of the major generative AI models but may be more convenient and powerful with paid versions.
• Payoff: Tomoff now can tackle more complex projects without outside help. “Projects that I either would not undertake or would take too long that I would undertake, I’m doing in minutes,” he said.
• The takeaway: Embrace the bot as an extension of your abilities, Tomoff advised. “If you’re an accountant, you no longer can say, ‘I don’t know how to do this,'” he said.

CASE 5: BRAINSTORMING AND RESEARCH

LLMs can be powerful research assistants — when they’re used correctly.

When Tomoff is researching a topic, he might feed an outline into ChatGPT or Perplexity, the LLM-based search engine. He’ll ask: “What other areas would you add, and why would you add them?”

The answers can highlight areas for research that he may have overlooked, and follow-up questions can provide some basic education in those areas. The power, Tomoff suggested, is in the volume of ideas the bots can produce.

“People will say it’s not any more creative than you and me,” he said. “It can sit and come up with ideas 100 times faster.” He also pointed to ChatGPT’s recent implementation of a “deep research” option, a more powerful research tool for paid users.

Still, Tomoff cautioned that chatbots will get users only 90% of the way to an answer. Because of the risk of hallucinations, the user still must confirm any facts taken from the software’s responses.

Meanwhile, a new class of research tools is tailoring LLMs to accounting needs. Young’s firm recently started subscribing to Ask Blue J, whose bot is focused specifically on federal and state tax laws. It can return answers and citations for specific, detailed tax questions. The software also can generate memos and emails to clients. But it comes with some kinks. The answers are “95%” factual, Young said, but the model often struggles to return accurate calculations when oft-changing, inflation-adjusted figures are involved.

“I’m always telling my staff, ‘Double-check that,'” he said.

Case 5 at a glance

• Degree of difficulty: Chatbots are intuitive tools for research, but safeguarding against hallucinations requires both domain knowledge and experience with the bots.
• Cost: The free versions of chatbots can help with brainstorming and research. Costs for more advanced or profession-specific implementations will vary.
• Payoff: At best, chatbots can help accountants quickly answer questions and identify new areas for research. “This is the tax manager I always wish I had, where I could ask any question and not get judged,” Young said.
• The takeaway: Chatbots are not genies that can magically answer any question, but they can be tools for extending an individual’s knowledge. “It just augments what you know,” Tomoff said.

CASE 6: AUDIT AUTOMATION AND ASSISTANCE

Katebini is an adviser to not-for-profits at GRF CPAs & Advisors, where she’s helped navigate the flood in recent years of automation and AI-powered tools, especially for audit uses.

The firm recently started using Trullion’s Audit Suite, which brings a range of useful features. The platform can:

• Search documentation for evidence related to a sample for an audit. “You’re not scouring through 80 pages of a bulk PDF. It’s very easy in that respect,” Katebini said. The technology can help staff quickly identify if there is a lack of evidence for a particular point in the audit.
• Extract summaries from documents such as board minutes and lease agreements.
• Automate elements of financial statements, including footing statements and tying back disclosures to the balances on those statements. “Our production team doesn’t have to go through the process of manually footing everything,” Katebini said.

Trullion is just one of many audit-focused products on the market. Another product, the AICPA’s Dynamic Audit Solution, analyzes large datasets to identify patterns and potential risks. Other products can automatically search for discrepancies and missing data or easily reformat data, if it’s sent in the wrong format.

“It’s the press of a button for us now,” Katebini said.

Taken together, the firm’s new tools are improving the staff’s ability to complete high-level analysis and build relationships with clients.

“Our staff is so inundated, there’s so much work to be done, it’s hard to take a step back and talk with your client sometimes, and technology advancements in assurance engagements is creating efficiencies to allow us more time to advise,” Katebini said.

Case 6 at a glance

• Degree of difficulty: Companies increasingly offer off-the-shelf products, but identifying and implementing solutions require familiarity with the technology and its applications, as well as project management expertise.
• Cost: Varies by product.
• Payoff: Substantially increased efficiency in managing and generating documentation.
• The takeaway: There’s a new generation of technological solutions and an emerging role for finance technologists to implement them.

How to determine AI technology needs

Navigating the growing variety of artificial intelligence (AI) technology options can be a challenge. Tricia Katebini, CPA, a partner at GRF CPAs & Advisors in Bethesda, Md., said that she spends so much time on technology, “I’m a person in the firm that has been designated to head up and lead the technology implementations, the vetting.”

Here’s how Katebini suggested identifying and implementing new solutions:

• Identifying pain points: Instead of looking for what “looks cool,” Katebini said, the firm began by asking staff about the tasks and work that were the most time-consuming and unexciting — such as creating note disclosures or getting data in the right format. “How can we solve that problem?”
• Seeking solutions: The team looks for solutions, including by asking vendors directly. “We’re bringing our pain points and problems to our vendors, too, and they’re very willing to listen.”
• Making a choice: The right tool is a matter of cost, utility, and compatibility with the rest of an organization’s tech stack.

Consulting on technology implementation can even become a line of business for a firm, said Brianne Smith, CPA/PFS, Ph.D., who runs her own accounting practice and a financial advisory firm in addition to being an assistant professor of accounting at the Auburn University campus in Montgomery, Ala.

“My firm and other firms are starting to consult on this type of process,” Smith said. “We make sure that they have proper security and privacy. We also recommend various tools for pain points.”

She and others suggested starting small with AI — even just experimenting for personal uses — and asking others for help and ideas along the way.

“There’s a whole community. We are all very willing to collaborate and talk about what tools we use,” Katebini said. “It’s kind of exciting to be at the crossroads.”

Claude AI prompt for VBA

Below is an Invenio Advisors LLC prompt instructing the Claude AI model to create a visual basic (VBA) bot to move items from one list to another list in Excel.

Prompt Structure

ClaudeAI PROMPT to move items from a list to another list (VBA)

You are an Excel VBA expert. I would like your help in creating a VBA macro that will MOVE (not copy) rows from one table to another when they’ve been checked. Specifically:

1. Source worksheet: {INSERT WORKSHEET NAME} with table named {INSERT EXCEL TABLE NAME}
2. Destination worksheet: {INSERT WORKSHEET NAME} with table named {INSERT EXCEL TABLE NAME} (create these if they doesn’t exist).
3. The destination table should start at cell A10 if it needs to be created
4. Move only rows where a checkbox in the last column is checked
5. Transfer all columns EXCEPT the last 3 columns from the source table
6. Prevent duplicates based on the video title (column 2) in the destination table
7. After moving a row, DELETE it from the source table
8. Include error handling for cases where tables don’t exist
9. Alert the user with a message upon completion

The macro should first check if items are already in the saved list before moving them.

Once you have created the code, please advise if you have any recommendations for improvement. I will then let you know if I want to make any changes to the code you provide.

After you complete this task, please provide instructions on how to implement the process in my workbook. Take a deep breath and let’s do this!

{INSERT WORKSHEET NAME}: ExcelTips_VideoList

Source: Invenio Advisors LLC

About the author

Andrew Kenney is a freelance writer based in Colorado. To comment on this article or to suggest an idea for another article, contact Jeff Drew at Jeff.Drew@aicpa-cima.com.

LEARNING RESOURCES

Navigating the AI Landscape: An Introduction to Generative AI and ChatGPT for Accounting Professionals

Learn how to use ChatGPT to streamline your workload, save time, and boost your overall efficiency.

CPE SELF-STUDY

Ethics in the World of AI: An Accountant’s Guide to Managing the Risks

This webcast discusses the current uses of AI in business, reviews nine risk areas, and provides practical suggestions to address these risks effectively.

June 4 and Aug. 5

WEBCAST

AI Strategy & Governance

Learn what organizations should consider when implementing AI solutions.

June 16, Gaylord National, National Harbor, Md.

CONFERENCE

ENGAGE 25

With nine tracks of expert content, you’ll gain exclusive insights, develop practical skills, and walk away with tangible guidance to evolve at your own pace. AICPA & CIMA ENGAGE 25 will be held at the ARIA in Las Vegas and is also available online.

June 9—12

CONFERENCE

Not-for-Profit Industry Conference

The conference, which will be held at the Gaylord National in National Harbor, Md., includes AI-related sessions on AI strategy and governance, advanced automation techniques using RPA and AI, and the C-suite’s road map to integrating AI in financial operations.

June 16-18

CONFERENCE

For more information or to make a purchase, go to aicpa-cima.com/cpe-learning or call 888-777-7077.

AICPA & CIMA MEMBER RESOURCES

Articles

“What AI Can Do for Auditors,” JofA, Feb. 1, 2024

JofA Artificial Intelligence (AI) Coverage