Years ago, when someone filed an IRS return, the assumption was that the recipient reading it and making decisions based on what it said would be a human. However as technology advanced, these humans began playing less of a role until, today, the majority of frontline review is being done by automated systems that read the data and make their own decisions based on it. Nowadays, the default audience for a tax return is not a person but a machine, which has created a slightly different set of factors for preparers to consider.
Darren Guillot, national director of consulting firm Alliantgroup, who previously worked in the IRS for 36 years (most recently as commissioner of the IRS' Small Business/Self-Employed Division), noted that automated enforcement mechanisms are certainly not new. As early as the 1970s, the IRS has had systems for detecting compliance shortfalls as well as analytics for determining whether something is a good case for an audit.
"The IRS has been using predictive analytics for a long time for determining what's the best case to audit, what's the best return for return to look at, for an audit or in a collection case, based on all these characteristics of this business or this individual and what they owe, what kind of notice should they get? How many notices should they get? At what cadence over what period of time? At what point should it go to enforcement or go to the field?" he said.
The big difference between then and now is not so much the introduction of analytics but, rather, their scale and scope. Before the focus was on the actual return because that was the information being presented to the IRS. Now, through connected data systems, examination is not just about a single return but its context within the broader array of data available to the government about the entity or individual attached to the return.
"When AI machine learning sits on top of the analytics, it can look at, theoretically, all of the data, and I mean all of it, millions of tax returns or cases where by the IRS in the last year, five years, 10 years, 20 years, and they can make all sorts of connections and analyze it, which would have taken a team of five analysts, hundreds of analysts, years to do before. It'll do it in seconds," said Guillot.
Jenny Groberg, founder and CEO of BookSmarts Accounting in Utah, said this means that while the process before was very much rule-based, where the overriding concern was making sure the specific filing followed all the instructions, things now are based more on patterns, where the main challenge is fitting within the expected norms of not only taxpayers themselves but other similar taxpayers as well.
"I would say with tax returns, the AI is looking for inconsistencies. It used to just be that you had a human reviewing a tax return that could look and say, 'Well, this looks off to me.' Perhaps now you have AI that can score risks for the return. You can detect anomalies. You can match data across systems, and it's industry specific as well," she said.
Filers need to be aware the systems will look for discrepancies in W2s and 1099s, inappropriate exemptions and inconsistent reporting. For example, IRS analytics might notice a restaurant is reporting an unusually high (or low) gross revenue compared to similar restaurants, or the systems might notice that charitable deductions suddenly went from $10,000 a year to $50,000.
"Rather than just saying, 'Oh, you've entered in all the boxes [correctly], it looks pretty good,' it's going to then compare you across industries. … If you have multiple data points, you can see trends. So if the IRS has multiple data points to compare you against other people, even to compare yourself against prior returns, they'll be able to flag indicators that there's inconsistencies," said Groberg.
Another part of this pattern matching is items that typically have produced audits in the past. Timothy Wijtenburg, a sole proprietor from Florida, noted that the IRS tends to pay extra attention to anything on the Schedule C because of previous abuse there. It therefore behooves a preparer to pay extra attention when working with the form.
"I think they continue to hone in the core areas where the most activity is, like the Schedule C. You're just going to have to prepare it even more efficiently and effectively," he said.
Because IRS automated enforcement relies so much on data, though, the system might have trouble with more ambiguous things that can't be strictly quantified. Machines are great at quantifiable systemization, but not so much with nuance.
"Machines are very good at spotting mismatches and patterns," said Tom O'Saben, director of tax content and government relations for the National Association of Tax Professionals, in an email. "They're less about nuance and more about consistency, which means we're preparing returns with the assumption that the first reviewer may be a system, not a person."
For instance, Wijtenburg brought up areas where the law is a little less settled and an algorithm might struggle, as its decisions will inherently rest on precedents that might theoretically be shown to be wrong in later court cases.
"What comes to mind is court cases and how they end up conflicting with the IRS. The IRS doesn't win every court case, and so it is [not] always going to be the case where an AI gets it right," he said.
Taveion McCutcheon, founder and CEO of Adept: Accounting Principles and Solutions, said nuanced situations might confuse a computer and are more difficult to communicate in terms of pure data, as opposed to when a preparer explains the situation to a human who would be able to understand. For instance, Form 8379 for injured spouse allocations requires detailed information and lots of documentation to explain complicated situations that a machine might struggle to comprehend fully.
"It takes you to be able to see a tax return with an individual saying 'Hey, I have debt. I don't want my spouse to take on this debt. How can we still file married jointly and still get the standard deduction, but so this individual won't take on my responsibility?' That is so detailed that you have to go through a lot before AI can even know that scenario is coming up in front of them," he said.
However, there was widespread consensus that while there might be some differences in how computers versus humans handle things at the IRS, trying to game the computers is not a viable tax avoidance strategy. Guillot noted that, first, the IRS is well aware that people have been trying to do so for years, and so devote considerable effort in heading off such attempts. But even more so, he said that while theoretically someone might try to manipulate the automated systems, such tricks can only get someone so far, as they likely won't stand up to even cursory scrutiny.
"Clearly you're thinking of how to game the system. There are people at the IRS who get paid to do nothing but think of how they are going to try to steal from us next," said Guillot.
He said some people in the past have tried to game IRS job platforms by inserting keywords into their resumes. While this might have gotten them an interview, once they actually sat in front of a person, it was clear they didn't know what they were talking about. He recommended against such tricks in the same way he recommended against trying to game taxes.
"You just guessed your way into this interview. That didn't help, because the first question I asked in the interview is: 'Tell me how much you know about the Notice of Federal Tax Lien in Section 6320 of the Internal Revenue Code.' Silence. You could get crickets, right? I think that would be a very dangerous game to play with taxes," said Guillot. The best way to avoid triggering IRS automated systems, he added, is to be honest.
O'Saben said that if something seems off, it will likely get more attention regardless of whether it is a machine or a human making that call.
"If something doesn't reconcile cleanly or would look inconsistent in the data, it's more likely to get attention, regardless of whether a human might ultimately understand it. So the focus isn't on trying to 'outsmart' automation. It's on making sure the return is accurate, consistent and well-supported from the start," he said.
Groberg agreed, saying that contending with automated systems does not mean finding some magic combination of words and numbers to trick them, but being mindful of their pattern-matching priorities, which usually means the boring but practical habit of maintaining good records.
"My recommendation is: spend your time focusing on consistency and record keeping so that if for some reason, something is different, then you can defend that position," she said.
This is a lot better than trying to outsmart the AI, she added: "If your goal in your tax return is just to try to make sure that you're not getting flagged by a bot, then there are bigger problems here."
