AI Advantages in Tax and Accounting Tilt Toward Big Firms

The jostling among tax and accounting firms for early-adopter advantage in generative artificial intelligence is shaping up as an unequal contest between the giants and their smaller rivals.

The Big Four—Ernst & Young, PwC, Deloitte and KPMG—and other top firms are going all in with multibillion-dollar investments to optimize their AI systems on industry-specific data. Their smaller competitors—less able to afford such investments and other costs—may have to make do with off-the-shelf models trained only on public data, like ChatGPT.

“There is a lot of vocabulary information that’s very unique to the specific business,” said Junta Nakai, global head of financial services at software company Databricks, which recently acquired MosaicML, a software developer that trains models. “E&Y or PwC or Deloitte are probably going to have company-specific lingo and information they have to train on.”

Fine-tuned models are better for specific tax and accounting-related tasks. Off-the-shelf foundation models are fostering a revolution in business for their ability to sort and synthesize information, draft letters and process documents. But without additional training, they lack in-depth knowledge of the tax and accounting worlds.

“Adding large language models to datasets like the tax code, governing documents, sourcing, compliance—you’re giving the opportunity for someone to ask specific questions to information,” said Jason Juliano, director of digital transformation at EisnerAmper.

There’s also a greater likelihood that models trained only on general data will “hallucinate,” or return fictitious data—a potentially catastrophic outcome were it to make its way into official filings.

“When they don’t know something, they hallucinate,” Kimberly Church, an accounting professor at Missouri State University, of said of the public models. Custom models are efficient because firms “define the input parameters” of their data, restricting their ability to fabricate.

Different Options

Most big firms have access to vast datasets on which to train models as they build upon their existing AI practices, Church said, but the reality for smaller firms looks different, “because of the lack of in-house IT specialization—they’re unwilling to accept the risk.”

And that is before concerns about data privacy. Firms must consider the privacy issues that come with using a public model, according to Mfon Akpan, an accounting professor at Methodist University in North Carolina. Any confidential information inputted into a public model, even for the most basic tasks, is at risk.

The model choice is a trade-off. Firms that are heavily investing in proprietary systems retain the most control over their data. Other options—which leverage more public systems—are more accessible but may have limited knowledge bases.

“Fine-tuning” an off-the-shelf model involves training it on a proprietary dataset, sometimes with human input, to tailor its knowledge base. Smaller firms may also have access to less data overall—another disadvantage.

“Certainly, there’s great, reliable public sources that we can all leverage, but that doesn’t give you any kind of advantage,” said Jeff Schmidt, chief technology officer at Crowe.

The generative AI revolution is in its early stages, and while it is impossible to say who the ultimate winners will be, firms across the industry are determined not to be left behind.

“They’re going to face pressures to try to catch up,” said Akpan, “which means they’re going to have to pay money to train up.”