Initial IRS Guidance on AI

In June, the IRS released thoughts on proper AI usage by tax professionals.The ground covered in the IRS’s news release says many of the same things observers have noted in recent times on AI cautions for practitioners. The guidelines do contain solid reminders for practitioners — especially since some parts of improper AI usage could lead to prison time!

Need for Guidelines

As noted above, the release from the IRS is not exactly “guidance;” it’s “guidelines.”

Many of us are using AI in our practices, and we probably think we know exactly what AI we are using, but keep in mind that there might be AI embedded in things, meaning you could be using AI even if you don’t consciously think about it being AI.

Circular 230 is the governing document for professionals who practice before the IRS, and the IRS’s guidelines tie ethical AI usage to Circular 230.

Circular 230 was last updated in 2014, but it actually handles AI quite nicely, because the concepts relating to ethical usage of AI are timeless and universal.

Hallucinations

An AI hallucination is the AI making something up that is inaccurate and without any basis. This doesn’t happen every time you use AI, of course. But it’s always a risk.

Even big players can get caught up in this. Deloitte Australia used AI to generate a report for the Australian government on the use of automated penalties in Australia’s welfare system. Deloitte charged the Australian government $290,000 U.S. Dollars for this report — which contained at least 20 errors, ranging from false citations to made up quotations. Deloitte refunded part of its fee because of the problems in the report.

In addition, numerous attorneys have been busted for submitting court briefs and other documents generated by AI where the AI has hallucinated the citations and other items in the documents.

Now, you might be thinking, “Nothing that dramatic has ever happened to me.” And I am sure you are right (I hope you are!) But any practitioner who uses AI needs to be aware of the Circular 230 expectations.

Due Diligence, AKA Human Review

After a setup similar to the setup we’ve just gone through, the IRS then dives into a regulatory framework of proper AI usage, starting with due diligence.

We can let Section 10.22 of Circular 230 do a lot of the talking here:

A practitioner must exercise due diligence in preparing or assisting in the preparation of, approving, and filing tax returns, documents, affidavits, and other papers relating to Internal Revenue Service matters; in determining the correctness of oral or written representations made by the practitioner to the Department of the Treasury; and in determining the correctness of oral or written representations made by the practitioner to clients with reference to any matter administered by the Internal Revenue Service.”

Whatever you’re asking the AI to do, you must review what it spits out. The problem with the Deloitte report mentioned above, or the attorneys who submit reports with hallucinated citations, is that quite clearly, no one reviewed what the AI produced.

This is, as the cliche goes, a recipe for disaster. It’s not just a violation of Section 10.22 of Circular 230, but it’s also simply an embarrassing thing in general. Imagine sending something to a client, and there are AI-generated inaccuracies that you didn’t catch … but the client does. Since we are using clichés in this paragraph, you can choose the cliché — egg on face, foot in mouth, raining on your parade, etc. etc. etc.

A related part of Circular 230 is Section 10.37, regarding standards for written advice. This is another spot where we can let Circular 230 do most of the talking. When giving written advice, you must:

“(i) Base the written advice on reasonable factual and legal assumptions (including assumptions as to future events);

(ii) Reasonably consider all relevant facts and circumstances that the practitioner knows or reasonably should know;

(iii) Use reasonable efforts to identify and ascertain the facts relevant to written advice on each Federal tax matter;

(iv) Not rely upon representations, statements, findings, or agreements (including projections, financial forecasts, or appraisals) of the taxpayer or any other person if reliance on them would be unreasonable;

(v) Relate applicable law and authorities to facts; and

(vi) Not, in evaluating a Federal tax matter, take into account the possibility that a tax return will not be audited or that a matter will not be raised on audit.”

This again gets back to reviewing what the AI is producing. It brings up one of your author’s favorite things to talk about with AI and tax pros. We are uniquely suited to using AI, because we develop a sixth-sense after awhile, where we are able to simply “tell” that something is off when looking at a tax matter. This is true when reviewing tax returns and it’s true when evaluating a piece of tax guidance or research. We can look at something and say “something ain’t right.” We might not know immediately “what” is wrong, but we know there’s something off.

An experienced professional can, in time, develop a similar thought process with AI. And, even if you don’t develop that thought process, at the very least, you need to review what you’re putting out there.

The 80/20 Concept

Some have said AI will get you 80% of the way. (Side note: “Who” has said that, dear author? Well, your author says, “people” have said it. He has read it and/or “heard it said.” Trust me, bro. But anyway, your author’s own anecdotal evidence when using AI lends a lot of credence to this thing that “some people say.”)

Your author has seen the 80% figure thrown out as a criticism of AI, but he prefers to look at it as a good thing. It can take a task and do most of the heavy lifting — leaving you with the task of reviewing and cleaning up. Still better than having to do everything yourself.

Think of AI as being junior staff — knowledgable in many ways, with a powerful ability to look things up. But the junior staff doesn’t always know how the pieces fit together. And of course, always verify any court cases, code sections, numbers given for IRS or Treasury guidance, etc.

This part — treating AI as junior staff — feeds into the next part of the IRS guidance.

Procedures to Ensure Compliance

Circular 230 10.35 relating to competence, and 10.36 relating to procedures to ensure compliance, both come into play as well.

Under 10.35, the practitioner needs to understand how the AI works. Whatever tool you use, you need to know how it works. As the IRS news release says, “Practitioners must understand both the law and the technology used in their representation of clients before the IRS, including AI systems’ operational mechanics, limitations, and risks.”

Under 10.36, you need to establish procedures in your firm for proper usage of AI. This includes procedures for how you plan to use AI, but also adequate training for staff (or yourself). The IRS also adds that all AI tools should be vetted. Know and understand the AI solution you are using.

All of these things might seem like somewhat-basic ethical things. But now we get into things where we go, “Wow, that’s heavy, man” — and that is, things where you could go to jail.

Sections 6713 and 7216

When you use AI, you could be violating privacy policies that are supposed to be in place between your firm and your clients regarding how you use your clients’ data. You cannot disclose client data to a third party without the client’s consent.

The AI tool is, in fact, a third party. What is it doing with the data you feed it? Is it using that data to train its system? How might the data be retained? As discussed in the prior section, you need to have answers to those questions at a basic level, but there’s also a very scary level to this as well.

There are two code sections — 6713 and 7216 — that govern usage of client data and disclosure to third parties. Section 7216 is often thought of as something you need to worry about when you offshore work to India, but it can come into play in many more ways. Usage of AI is one of those ways.

If you use AI, and you’re either feeding client-specific data into the AI, or the AI simply has access to client files by virtue of the AI existing (such as a solution that can view all your firm files or emails), you need to disclose that to your clients under section 7216.

Violations of section 7216 come with monetary penalties and up to one year in prison per violation. When your author teaches on these things, and section 7216 comes up (regardless of the setting — be it AI, offshoring, sending a tax return to a bank or investment advisor) — your author always says “do you really want to go to jail?”

Oh and the monetary penalties — section 7216 violations start with a basic fine of $1,000 per violation. Section 6713 and section 7216 then combine to say the penalty under section 6713 is $250 per violation, up to a maximum of $10,000 a year.

Oh, and this little nugget — if there is a crime involving identity theft because of your actions, the penalty under section 6713 increases to $1,000 per violation and $50,000 maximum penalty per year. And section 7216 can impose a whopping $100,000 penalty per violation in this situation.

Be careful with what you’re feeding our robot overlords. If you’re giving them client-specific information, you better have the client’s consent!

Fees

We leave you with something that your author is still pondering. It’s something we all need to ponder. The IRS news release mentions a piece of Circular 230 involving the charging of unconscionable fees. The IRS opines that AI can reduce the time involved in researching and drafting advice, and, quote:

Cost savings should be passed on openly, with billing practices that reflect the efficiencies gained from the use of GAI. Practitioners should not only disclose, in general or specific terms as needed, the AI activities performed, but also fairly credit to the client’s account any cost reductions.

This statement has caused quite a stir online among practitioners. Your author would say we should think about it before making any hasty changes. As some online have pointed out, the IRS’s words here would be true for sure if you’re billing based on time. A research task that took you two hours but now takes one hour because of AI can’t very well (or ethically!) be billed at two hours now.

But, as many of our readers know, we’ve been told for nearly 40 years that the billable hour is dying. Perhaps you have joined the revolution of billing with flat fees or value pricing. One could argue, quite compellingly, that the time savings helps you deliver even better service to your clients. Therefore, the client fee shouldn’t go down, despite what the IRS bizarrely says about “crediting people’s accounts.”

Your author can’t quite say with 100% certainty that the IRS is talking about hourly billing, but that seems to be the logical conclusion. The IRS says this, in that section: “AI can reduce research and drafting time, such that billing clients for manual labor or time that was not actually spent or double billing for AI-assisted tasks may violate § 10.27, depending on the facts (e.g., a noticeable pattern across clients or the size of the billing differentials).”

While this doesn’t explicitly say it’s referring to hourly billing, it seems like that is a logical conclusion. So, if you bill by the hour, think about how AI affects this practice. And if you don’t bill by the hour, it’s still a good idea to reflect on how your fee structure fits into an AI world.

Conclusion and Best Practices

We’ve used a lot of cliches in this article, and your author will provide another one: AI is here to stay. Whatever AI bubbles or such might happen, it is not going to disappear.

How do we navigate this? Your author recommends:

• Write down all of the AI you know you use — this might just be ChatGPT or Claude.

• Then, think about and write down all the AI that might exist in places such as your tax prep software, or your email.

• Boom, now you have your list, so you can add it to your client disclosures.

• Think about, and start documenting, AI procedures.

• Think about your billing, especially if you bill by the hour.

What if you miss an AI tool when you’re making your list? Will the IRS smite you? They could, I suppose. But I think the true “smite from the face of the earth” sort of thing would come into play if you’re feeding the sensitive data of your clients to an AI. An AI tool that is merely embedded in something else, and for which you never use and it might not even be doing anything, is much less likely to be worthy of “OMG IT’S ALL OVER.”

Don’t fear AI; use it as a tool. But as the cliche goes (we have to have one more cliche) — use it wisely.