Julia Chirkova
The recent release of Alert 2026-19, "Introductory Guidelines for Responsible AI Use in Federal Tax Practice," by the IRS Office of Professional Responsibility (OPR) on June 24, 2026, has ignited significant discussion within the tax and accounting profession. While much of the immediate commentary has focused on the implications for billing practices, particularly concerning Section 10.27(a) of Circular 230, which addresses unconscionable fees, a deeper examination reveals that the alert’s most transformative impact lies in its redefinition of practitioner responsibility under Section 10.37, governing written advice. This shift compels tax professionals to prioritize the "traceability" and "verifiability" of AI-generated outputs over purely cost or billing considerations, fundamentally altering how AI tools should be evaluated.
The core of the OPR’s guidance, as outlined in Alert 2026-19, largely reiterates existing duties under Circular 230, but crucially applies them to the burgeoning use of Artificial Intelligence in tax practice. These established principles include the practitioner’s ultimate responsibility for all work product, the stringent prohibition against submitting inaccurate or fraudulent documents, and the imperative to safeguard client data from unauthorized disclosure. The alert also directly addresses the phenomenon of "hallucinated citations" – instances where AI systems generate citations to non-existent legal authorities – warning that such errors can lead to disciplinary sanctions. The novelty, therefore, does not lie in the introduction of new ethical obligations but in the application of these long-standing duties to the unique challenges presented by AI technologies, with the issue of AI opacity emerging as the most critical compliance hurdle.
The Opacity Problem: A New Compliance Frontier
The OPR’s guidance places a significant emphasis on Section 10.37, which governs the provision of written advice by tax practitioners. The alert clarifies that relying on an AI system’s output without the ability to trace its logic to verifiable sources can be deemed unreasonable. This directive challenges the prevalent practice of employing general-purpose AI systems, such as large language models, which are designed for fluent and confident output generation but often lack the inherent architecture to meticulously document their reasoning processes. These systems may produce sophisticated answers that appear authoritative, yet they frequently fail to provide clear, direct links to the specific statutes, regulations, or judicial decisions upon which their conclusions are based. Furthermore, when these systems do generate citations, there is no guarantee of their accuracy or existence.
This concern is underscored by a widely reported incident involving Deloitte Australia. In July 2025, a 237-page report commissioned by the Australian government, and prepared by Deloitte Australia, was found to contain fabricated quotes attributed to a judge and citations to non-existent academic works. These errors, reportedly generated by AI, were allegedly not identified by the firm before the report’s delivery. The inaccuracies were ultimately uncovered by an external academic who cross-referenced the citations with their purported sources and discovered their absence. In response to this oversight, Deloitte Australia subsequently refunded a portion of its fee to the client. While this case involved a substantial report and a government client, the underlying risk is directly transferable to tax practitioners. A tax memorandum, a client return, or a position paper developed using an opaque AI system carries a comparable risk of containing factual or legal inaccuracies that may go undetected until external scrutiny or a client dispute arises. The potential for such "invisible" errors to permeate a firm’s work product, regardless of the scale of the engagement, is a primary concern highlighted by the OPR.
Elevating Due Diligence and Competence in the AI Era
Alert 2026-19 further reinforces the OPR’s expectations by extending the due diligence requirements of Section 10.22 and the competence requirements of Section 10.35 to AI-assisted work. Practitioners are now explicitly mandated to verify the accuracy of all facts, citations, and calculations within AI-generated outputs before they are presented to clients or submitted to the IRS. This necessitates a profound understanding of how AI systems generate their content. The alert clarifies that mere perusal of vendor marketing materials or reliance on benchmark scores is insufficient. Instead, tax professionals must possess the capability to articulate the provenance of every factual assertion and legal claim within an AI-assisted work product. This includes verifying that the cited source is not only extant but also remains current and authoritative.
The convergence of Section 10.22’s verification duty and Section 10.35’s competence requirement leads to a practical imperative: firms must implement a rigorous independent verification process. This process mirrors the traditional manual research methods that AI tools were intended to streamline. When an AI system cannot inherently provide a clear audit trail of its research process, the firm is compelled to manually re-derive citations, re-calculate figures, and cross-reference the underlying regulations and statutes. This effectively recreates a significant portion of the research workload that AI was acquired to mitigate. In essence, a practitioner utilizing an opaque AI tool faces a dual cost for research: the initial cost of AI generation and the subsequent cost of manual verification to ensure accuracy and compliance.
Defining "Traceable" AI in Practice
To navigate these evolving compliance standards, it is crucial to differentiate between compliant and non-compliant AI tools. A typical general-purpose AI system, when queried on a tax research question, may produce a coherent and confidently worded analytical paragraph, often concluding with a definitive statement. While it might include citations presented as text, these often lack direct, actionable links to the primary source documents. This format prevents practitioners from independently verifying the existence of the cited authority, its accuracy, or its current legal standing without leaving the AI interface and conducting separate research. The reliance on the AI’s confident assertion of the citation’s validity is precisely the vulnerability the OPR’s alert seeks to address.
Conversely, an AI system designed for traceability treats primary source material as the core output, rather than an ancillary attachment. In such systems, every factual or legal claim is directly linked to the specific statute, regulation, ruling, or case law it is derived from. These links allow practitioners to access and review the original source material directly, in its original language, without solely relying on the AI’s interpretation. This fundamentally alters the nature of the Section 10.22 verification step. Rather than embarking on an exhaustive re-research effort, the practitioner can efficiently confirm that the AI-surfaced source accurately supports the claims made, a task consistent with the diligence expected of a competent professional. This approach transforms the verification process from an open-ended research project into a targeted confirmation, significantly reducing the time and effort required while maintaining robust compliance.
The development of such traceable AI systems is an engineering challenge that has practical solutions. Some advanced systems are architected to separate the component responsible for retrieving and processing primary sources from the component that generates the written analysis. In this model, the drafting component is strictly confined to using a finalized research record, rather than drawing from the AI model’s general knowledge base. Within this framework, each piece of quoted text or factual assertion can be tagged with a unique identifier linked to its precise location within the retrieved document. This marker, rather than free-form text, becomes the sole mechanism for citation. Upon assembly of the final output, each marker is validated against the actual retrieved content. Any marker that does not correspond to genuine retrieved information is excluded from the output. This architectural distinction—between a generative model that can deviate from its instructions and a system with built-in constraints—is critical for enabling firms to effectively supervise AI use under Section 10.35.
The Prematurity of Billing Debates Without Traceability
The focus on Section 10.27(a) and the concept of unconscionable fees, while understandable, appears to be a premature debate in the current AI landscape. The OPR’s guidance clearly states that cost savings realized through AI should be transparently reflected in billing, encouraging practices that align with actual efficiency gains. While this may eventually lead to a broader shift towards value-based billing models, applying this principle directly to AI implementation without addressing the underlying traceability issue is problematic.
A firm cannot accurately price services based on value delivered, nor can it defensibly demonstrate efficiency gains to clients under Section 10.27(a), if the AI tool used to generate the work cannot substantiate its own reasoning. The inability to verify an AI’s output fundamentally undermines the ability to claim and justify cost savings or value delivered. Opacity, therefore, is not merely a Section 10.37 concern; it is the fundamental barrier that renders the billing question unanswerable until the issue of traceability is resolved. Without a clear audit trail, firms are left in a position of not being able to credibly account for how AI has reduced their internal costs or improved client outcomes.
Strategic Steps for Firms in the AI Transition
Alert 2026-19 unequivocally places a new, critical question at the forefront of every AI tool evaluation: can the system demonstrably trace its output back to a specific, verifiable primary source, or does it merely generate a confident narrative? Tax and accounting firms do not need to await further OPR pronouncements to begin implementing strategies to address this.
During any evaluation or renewal process for AI tools, firms must directly engage vendors with this crucial question: "For any given output, can your system display the specific primary source behind each material claim in a format that allows for direct verification, or does it present citations solely as unlinked text?" Furthermore, firms should proactively begin documenting, on an engagement-by-engagement basis, which claims within AI-assisted work products were independently verified against primary sources and by whom. This detailed record will serve as essential evidence of compliance with Section 10.22 due diligence and Section 10.35 competence requirements, particularly if the work product is ever subjected to IRS scrutiny.
It is also prudent for firms to acknowledge and account for verification time as a genuine cost associated with using opaque AI tools. This cost should be factored into engagements honestly, rather than assuming that advertised time savings are fully realized once the necessary human review layer is reinstated.
Ultimately, all other considerations raised by Alert 2026-19—including the burden of verification, the defensibility of billing practices, and potential exposure under Section 10.37—are contingent upon the answer to this core question of traceability. An AI tool meticulously designed for tax research, one that inherently demonstrates its work against primary sources, offers more than just enhanced usability. It provides the essential framework for firms to meet the rigorous standards now formally articulated by the IRS, ensuring both compliance and client trust in an increasingly AI-driven professional landscape.
About the Author: Kevin Boeckholt, CPA, spearheads Accordance’s innovation practice, where he collaborates with tax and accounting professionals to integrate AI into their daily workflows. His prior experience includes leading product development at a nascent tax AI company and serving in Tax Technology Consulting at Deloitte.
