Millions of Americans abroad are out of compliance, not because they refuse to pay, but because complying costs more than the tax they owe. Artificial intelligence can change that equation, provided it works under a method that keeps a professional accountable for every word.
An American retiree in Bordeaux owes no American tax: her French pension and her French taxes erase it. To prove it, she needs a federal return, foreign tax credit forms, an FBAR, perhaps a Form 8938 and several forms for her French funds. A competent preparer on both sides of the Atlantic will charge her more than a year of her pension's American tax would ever be. So she does nothing, and joins the millions of Americans abroad whose non-compliance is not a refusal to pay, but a rational answer to a price.
In our analysis of FATCA fifteen years on, we measured that compliance gap. In our ten structural mismatches, we explained why the two systems collide. This article argues that the gap is first an economic problem, that artificial intelligence can lower its price, and that the conditions under which it may do so are already being written by the courts of both countries.
Part I sets out the economic diagnosis. Part II describes what AI changes and what it does not. Part III examines the risks the courts have already sanctioned. Part IV presents our method and our position.
I. The economic diagnosis
A. The tax owed is often zero; the cost of compliance never is
The American system taxes its citizens wherever they live, and treaties and foreign tax credits mean that most Americans in France owe little or no federal tax. But the absence of tax does not dispense with the returns, and the returns of an American abroad are among the most complex the IRS receives: credits computed by category, information returns on foreign accounts, assets and companies, and regimes designed for multinational groups that apply to a retiree's French savings plan. The cost of producing them is fixed, the tax at stake is often zero. For a large part of this population, rational compliance is not compliance at all.
B. An administration that has withdrawn, and a rare skill
The IRS no longer has any office abroad: it closed its last tax attaché posts in 2015, despite the repeated objections of the National Taxpayer Advocate, the independent voice of taxpayers within the IRS. In her 2025 annual report to Congress, the National Taxpayer Advocate again listed among the most serious problems of the tax system the severe compliance burdens borne by American taxpayers living abroad. On the private side, the skill required is rare: knowing American federal tax law and the French system well enough to align them, form by form. That scarcity sets the price, and the price keeps the gap open.
C. What the gap costs
The gap costs taxpayers in anxiety, in blocked banking relationships, and sometimes in disproportionate penalties for accounts that generated no tax at all. It costs the administration in the form of a population it can neither serve nor audit, whose compliance would bring little revenue but much legitimacy. Our position, set out in our previous analyses, is that coercion alone has failed to close it in fifteen years. Only one lever remains untested at scale: the price.
II. What AI changes, and what it does not
A. The marginal cost of a file
A compliance file for an American in France is a sequence of tasks that are each simple and collectively long: collecting bank statements and French tax notices, classifying accounts and products, converting amounts, identifying the forms due, applying the treaty, preparing the returns. Most of this sequence is reading, sorting and structured drafting, precisely what language models now do quickly and at low cost. What took a professional days can be prepared in hours, and the professional's time can be concentrated on what requires judgment. The marginal cost of a well-documented file falls; the price can follow.
B. What remains human: judgment, signature, accountability
AI does not decide whether an assurance-vie contract is a PFIC or a grantor trust, whether a taxpayer was non-willful, whether a treaty position should be taken and disclosed. Those are judgments that engage a professional's responsibility, before the client and before the administration. The American professional standards are explicit: the American Bar Association's Formal Opinion 512 of 29 July 2024 states that lawyers may not abdicate their responsibilities by relying solely on a generative AI tool to perform tasks that require professional judgment, and applies to these tools the duties of competence, confidentiality, communication and reasonable fees. AI augments the professional; it does not replace the signature.
III. The risks, and the courts that have already sanctioned them
A. Invented references
The best-known risk is the fabricated citation. In Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), the lawyers filed a brief citing six decisions invented by ChatGPT, then stood by them when questioned; the court sanctioned them with a 5,000 dollar penalty and recalled that nothing is inherently improper in using an AI tool, but that the rules impose on lawyers a gatekeeping role over the accuracy of their filings. French courts have followed. At the end of 2025, the administrative court of Orléans publicly warned a lawyer whose submissions contained some fifteen nonexistent decisions (TA Orléans, 29 December 2025, n° 2506461), and the judicial court of Périgueux invited the parties to verify that references found through search engines or artificial intelligence are not hallucinations (TJ Périgueux, 18 December 2025, n° 23/00452). In tax matters, an invented reference is not only an embarrassment: it is a position taken before an administration that can sanction it.
B. Professional secrecy and the security of tax data
A compliance file contains everything: identity numbers, accounts, income, family. Submitting it to a public AI tool may breach professional secrecy and expose the client. The obligations are the same on both sides: confidentiality of client information under the American professional rules recalled by Opinion 512, and professional secrecy of the French lawyer, which the Conseil national des barreaux has addressed in its practical guide on generative AI of September 2024 and its guide on ethics and artificial intelligence of 17 March 2026. The choice of tools, their contractual terms and the place where data is processed are therefore not technical details but ethical decisions.
C. Ethics on both sides
A professional who prepares American returns and represents taxpayers before the IRS is subject to the rules of practice of Circular 230, which require diligence and accuracy whatever the tools used. A French lawyer remains bound by the duties of competence and prudence of the profession. Neither framework forbids AI; both make the professional answerable for what it produces.
IV. Our method and our position
A. Systematic verification at the source, on both sides of the Atlantic
Our rule is simple: no legal reference is cited from memory, whether produced by a human or a machine. Each French reference is checked against the official databases, each American reference against the statute, the regulations or the decision itself, and each article of this publication states, before publication, how its sources were verified. When a decision cannot be found in the official databases, we say so, and we say how it was verified. This discipline, more than any tool, is what distinguishes an augmented practice from a content generator.
B. Fixed fees rather than hours
If AI lowers the cost of a file, billing by the hour would capture the gain for the professional, not for the taxpayer, and Opinion 512 itself warns that fees must remain reasonable in light of the time actually spent. We apply fixed fees, set by the complexity of the situation and not by the time spent. That is the only way for the productivity gain to reach the price, and for the price to reach the level at which compliance becomes rational again for the retiree in Bordeaux.
C. What the administration could draw from it
An administration that no longer has offices abroad could see in augmented practices an ally rather than a risk: more complete files, better documented, filed through the existing regularization procedures, at a cost compatible with the tax at stake. The streamlined procedure, described in our analysis of its 2026 eligibility, already offers the legal framework. What has been missing is a price. For accidental Americans who hesitate between complying and renouncing, the stakes we described in our article on accidental Americans and in our analysis of renunciation depend on the same variable.
Conclusion
The compliance gap of Americans abroad will not be closed by coercion alone, nor by goodwill. It will be closed when complying costs less than worrying. Artificial intelligence makes that price reachable, because it lowers the marginal cost of a file without lowering the standard of the work, provided every reference is verified at the source and an identified professional signs and answers for every file.
The courts of both countries have already drawn the line: AI is not improper, unverified AI is. Our position is to stand on the right side of that line, publicly, and to pass the productivity gain on to the taxpayers for whom the price of compliance was the only real obstacle.
Frequently asked questions
Why are so many Americans abroad out of compliance?
Mostly because of cost, not refusal. Thanks to treaties and foreign tax credits, many owe little or no American tax, but they must still file complex returns and information forms. The cost of competent preparation on both sides of the Atlantic often exceeds the tax at stake, and the IRS no longer has offices abroad to help them. The National Taxpayer Advocate again identified these severe compliance burdens in her 2025 annual report to Congress.
Can artificial intelligence prepare my American returns?
It can prepare much of the work: collecting and classifying documents, converting amounts, identifying the forms due, drafting. It cannot replace the judgments that engage responsibility, such as the classification of a French product, the existence of non-willful conduct or a treaty position. American professional standards, notably ABA Formal Opinion 512 of 2024, require a professional to retain that judgment and to answer for the result.
What are the risks of using AI in tax matters?
Two above all: invented references and data confidentiality. American and French courts have already sanctioned or publicly warned lawyers whose submissions contained decisions fabricated by AI, from Mata v. Avianca in 2023 to the administrative court of Orléans in December 2025. And a tax file contains highly sensitive data that must not be exposed through public tools. Systematic verification at the source and a controlled choice of tools are the answers.
Does AI make tax advice cheaper?
It should, if the gain reaches the price. Billing by the hour would keep the productivity gain for the professional; fixed fees set by the complexity of the situation allow it to benefit the client. ABA Formal Opinion 512 also recalls that fees must remain reasonable in light of the time actually spent when AI tools are used.
