Beyond Predictions: Using Machine Learning to Inform Tax Strategy

Future court decisions can now be predicted with speed, accuracy, and precision using machine learning, a subset of artificial intelligence. Given a sufficient dataset of past adjudicated outcomes on a question of law requiring the weighing of relevant facts and circumstances, the machine learning algorithm can achieve over 90% accuracy in its predictions across multiple areas of tax law.Take the question of whether or not a person is liable to pay trust fund taxes under § 6672 of the Internal Revenue Code (IRC). The algorithm is first provided with the inputs (the facts of over 400 U.S. court decisions) and the outputs (the adjudicated outcomes of said decisions). The algorithm then learns the relation between the inputs and outputs and is trained to make a prediction as to how a court would rule when faced with new situations and fact patterns in that same area of law.Blue J Legal's machine learning-powered platform, Tax Foresight, uses a questionnaire to collect the relevant information from the user and uses its learning from past decisions to provide a prediction. The ability to obtain an accurate prediction for a certain fact situation requires that the user knows the answer to all, or almost all, of the questions informing the various factors.But the platform's practical value goes beyond its accurate predictive functionality in a situation where your client is already aware of a tax delinquency and possibly already facing the trust fund recovery penalty (TFRP). The platform's algorithm can provide value much earlier in the game through its insights into the relative significance of the various factors vis-à-vis each other and how they interact in order to affect the outcome.

The § 6672 TFRP: A Brief Review

Any person can be personally liable for an amount equal to 100% of delinquent “trust fund taxes”—usually payroll tax withholdings[1]—of the business they are involved in, if that person is a “responsible person”[2] in the business and “willful”[3] in failing to remit. Numerous factors relating to a person’s influence and control of the business, their actual or constructive knowledge of the tax delinquency, their actions and efforts upon learning of same and the circumstances preventing payment go into a two-step analysis, first to determine “responsibility” and then, if the person is responsible, to determine “willfulness.” In over 90% of cases where a person was found (or conceded to being) “responsible,” they failed to show that they were not “willful” and the TFRP applied.

Insights from TFRP Factor Testing and Comparison

Imagine a person who is the president and director of a corporation, who signs checks and tax returns and also reviews and approves payroll. For quite some time he doesn’t know that there is a tax problem because someone else is hiding information from him. When he does find out, he tries to do something about it right away, but he doesn’t actually cause any payments to be made. An experienced professional might predict that this person will probably, although not certainly, be found both “responsible” and “willful” and be liable for the TFRP. The algorithm does as well. So far, as a predictive tool only, the algorithm may not appear to show anything interesting.Let’s try testing and comparing factors. What does the algorithm reveal about the impact of changing this person’s review and approval of payroll? It turns out this factor is quite strong. Everything else remaining the same, if this person never reviews and approves payroll, the classifier’s confidence in the probability of him being penalized goes down significantly—in fact, the classifier now indicates it is slightly more probable that he will not be penalized. Reviewing and approving payroll has a stronger impact on the probable result than status as president and director, or check-signing, which is a traditional indicator of control over disbursement of funds. But changing other factors to increase the person's other powers—such as the power to hire and fire and sign financial contracts—can move the predicted result back to a probable penalty, despite the person not having anything to do with payroll.It may be fairly obvious that the titles of “President” and “Director,” while not by themselves causative of significant financial control of a business or knowledge of its tax status, are traditionally associated with duties and responsibilities that evidence such control and knowledge. It is no surprise that having such titles does make it substantially more likely that the person will attract the TFRP. Data analysis confirms that the vast majority of decisions applying the TFRP involved a person who was either an officer or a director or both. In contrast, only about 15% involved a person who was neither.It may not be as obvious, however, that if a person does have such titles and the power associated with them, then simply avoiding signing checks and tax returns will probably not be enough to insulate the person from being found, at the least, to be a “responsible person” if they authorize payroll at any time. Or, to put it another way, if a person has such titles, and does sign checks and tax returns, they are not necessarily doomed to be found to be a "responsible person" if they truly have no power over payroll.These insights stimulate further thinking as to why the impact of reviewing and approving payroll is so strong. This task requires authority and control with respect to expenditure of company funds. It also requires authority and control with respect to the priority of creditors, in the situation where there is not enough money to pay both the employees in full and the trust fund taxes in full. Authority and control with respect to expenditure of funds and priority of creditors are the essence of “responsibility” in the TFRP context.[4]  It follows that a person who has a role traditionally associated with authority and control and who does fulfill some financial duties, like signing checks and tax returns, might still not be responsible if the business is structured and run in such a manner that the person nevertheless has no authority or control over disbursements and creditor priority. Further, reviewing and approving payroll is also related to knowledge and willfulness in that a person who has such a duty would usually be expected to know the status of payroll withholding taxes.In the scenario, changing factors around hiring and firing and signing financial contracts counteracted the strong impact of not having authority over payroll. Why? Is it simply a matter of having more “bad facts” than “good facts”? A coherent and commonsensical way to look at it could simply be that it stretches credulity that a businessperson could have the power to hire and fire, and enter into financial contracts on behalf of the business, and sign checks, and sign tax returns, and yet not truly have power over payroll. In such a situation, this lack may well be artificial and contrived.

Dixon v. Commissioner: The Case of the Tycoon-In-Training

The above scenarios actually happened in the 2019 Tax Court case of Dixon v. Commissioner.[5] In 2005, while in his early 20s, Mr. Dixon was the sole director and president of a corporation and signed checks and tax returns. The IRS levied the TFRP on him for 2005 delinquencies. At trial, Mr. Dixon established that his parents, who owned the corporation, were the true directors and he knew very little about the business. He signed all documents as they directed and did not prepare, review or question them. He exercised no authority or control over disbursement of any funds, let alone payroll, or priority of any creditors. After the IRS contacted him about the tax problem, he called his parents for guidance and accepted that they would handle it. Held, Mr. Dixon was not a responsible person in 2005 and not liable for the TFRP then.The IRS also levied Mr. Dixon again for 2008 to 2012 delinquencies. He used the same argument as for 2005, but it did not succeed this time. By 2008, with his parents' approaching retirement (and incarceration for tax crimes), he took on management duties including hiring, firing and entering into leases. He no longer signed checks only when told to by his parents. He knew of the past and present payroll tax obligations, permitted them to keep accruing, failed to pay them and deliberately avoided contact with the IRS. Regardless of whether he did or did not personally review or approve payroll, this one potentially “good fact” for him could not be reconciled with the other “bad facts” in a holistic manner to arrive at a common sense conclusion other than he he was no longer a clueless trainee completely dominated by his parents, but had grown into the traditional roles and duties of Director and President for which he had been groomed. He was held liable for the TFRP for the 2008 to 2012 time period.

Better Learning with Machine Learning

With respect to Dixon, the insights into the TFRP analysis gained from machine learning confirm that it is worthwhile to ask and ponder questions like:

  • In 2005, could the government have succeeded if it had more evidence going to influence and control besides Mr. Dixon’s titles and signed checks and returns? Does being “in training” really make sense for why he had the titles of president and director without the associated powers? Was there something artificial or contrived about his having nothing to do with payroll? Could he have refused to sign a check or tax return? Ought he to have known there was a tax problem or risk earlier?
  • In 2005, would Mr. Dixon have been likely to succeed in avoiding the TFRP if the focus of his argument was that his titles were only nominal, or that he only signed a few checks, instead of his parents’ complete control? Would that argument have been persuasive if he were older, or had prior business experience or education, or if the owners were not his own parents? Would it be fair to say that this was a close case?
  • After 2005 and before 2008, would it have been advisable for Mr. Dixon to have changed procedures and personnel around tax compliance? If he could not do this because of his parents’ control of the corporation, could and should he have taken steps at this time to reduce their control?
  • In 2008 to 2013, was it Mr. Dixon’s best strategy to repeat the argument that his parents controlled the business in the face of evidence that he had gained more control and discretion over the corporation? Was there any theory incorporating the relevant factors that might still have avoided a “responsibility” finding?
  • In 2008 to 2013, ought Mr. Dixon to have focused more on factors going to knowledge of the delinquency and attempts to address them? Could he have avoided the willfulness finding? Was his knowledge possibly hindered by his parents’ dishonesty? Would it have been advisable for him to reach out to the IRS sooner instead of waiting for them to contact him?

Machine Learning and the Professional Workflow

As demonstrated, machine learning is valuable for more than just making predictions. Whether one is putting a hypothetical scenario to a client as part of giving advice, or thinking through the theory of the case as part of preparing for litigation, machine learning insights inform and inspire better and more complete thinking around analysis, risk management, and advocacy.


[1] Trust fund taxes include excise taxes but when the Internal Revenue Service seeks to apply the TFRP, it is usually with respect to payroll tax withholdings, such as federal income tax, Medicare and Social Security. These withholdings are often referred to as “trust fund taxes” and the penalty is called the “trust fund recovery penalty” because once collected, the amounts are considered held in trust for the United States until paid to the Treasury, pursuant to IRC § 7501(a).

[2] A “responsible person” is generally someone who has significant control over the financial affairs of the business, especially over disbursement of funds and priority of payments to creditors, such that they could be said to be under a duty to collect the trust fund taxes. Gephart v. United States, 818 F.2d 469 (6th Cir. 1987), at 473.

[3] “Willful” in the context of § 6672 is generally defined as a voluntary, conscious and intentional choice to prefer other creditors to the United States government. Denbo v. U.S., 988 F.2d 1029 (10th Cir. 1993), at 1033.

[4] See, for example, Gephart, supra.

[5] T.C. Memo 2019-79. Tax Foresight’s Trust Fund Recovery Penalty Classifier predicted both rulings correctly, at 58% confidence for the 2005 tax year and at 95% confidence for the 2008-2012 tax years.

Discover More

Related Blogs


Get blogs directly to your inbox

Sign up for the Blue J newsletter today.

Get Started

Start enjoying better, faster analysis with Blue J

Whether you have questions or are interested in booking a demo, we would love to hear from you.

By clicking “Accept All", you agree to the storing of cookies on your device to enhance site navigation, analyze site usage, and assist in our marketing efforts. See our Privacy Policy for more info.
Deny All