Can AI Deliver Defensible Professional Tax Research?

Professional responsibility comes first

CiteMax is designed around a fundamental principle: AI-assisted research informs professional judgment; it does not replace it. The practitioner—not the AI system—is responsible for evaluating the facts, applying relevant law, checking authority, and deciding what advice can responsibly be provided to a client.

A legally accurate response and a usable professional research workpaper are different deliverables. A well-organized workpaper makes assumptions, citations, limitations and competing interpretations easier to identify and evaluate. Its presentation, however, cannot cure a legal error. Practitioners must independently inspect the source material rather than assume that citations, quotations or conclusions are correct merely because they appear in a polished report.

The practitioner review standard

Establish and reconcile material facts. Identify missing documents, inconsistent evidence, relevant tax years and applicable jurisdictions.

Open and independently verify controlling statutes, regulations, judicial decisions and published administrative guidance; confirm that each source actually supports the proposition for which it is cited.

Distinguish binding authority from nonbinding commentary, and identify adverse or limiting authority, material exceptions, effective dates and unresolved interpretations.

Recalculate material amounts, deadlines, election consequences and procedural requirements using independently confirmed inputs.

Apply professional judgment, document any corrections, and determine whether additional research, client inquiries or specialist review is needed before advice is issued.

Protect client information. Evaluate confidentiality and data-handling requirements before entering client records into an AI platform.

Circular 230: professional duties that remain with the practitioner

Circular 230 §10.22 requires due diligence concerning the accuracy of papers filed with the IRS and representations made to the Treasury and clients. Section 10.35 requires competence appropriate to the work undertaken. Section 10.36 addresses compliance procedures, including supervisory responsibilities. Section 10.37 requires reasonable factual and legal assumptions, consideration of relevant facts and the application of law and authority to those facts. These obligations do not disappear because an AI system drafted a response.

Selected regulatory excerpts (short quotations): §10.22: “A practitioner must exercise due diligence”; §10.37(a)(2)(v): “Relate applicable law and authorities to facts.” The complete provisions should be consulted in their current form; excerpts are not a substitute for the regulation.

IRS Office of Professional Responsibility: June 2026 AI guidance

On June 24, 2026, the IRS Office of Professional Responsibility published Issue 2026-19, “Introductory Guidelines for Responsible AI Use in Federal Tax Practice.” It explicitly applies Circular 230 obligations to generative AI. Its discussion emphasizes independent review of AI-created material, verification of factual statements, citations and calculations, competence, supervision, ethical billing and the protection of taxpayer information.

OPR also cautions practitioners to examine the underlying cases and citations rather than accept AI-generated legal analysis uncritically. This is the professional rationale for evaluating CiteMax as both an answer-producing system and a structured workpaper-producing system.

Separate IRS internal AI policies

The IRS Internal Revenue Manual 10.24.1 (revised August 10, 2026) addresses AI governance for IRS personnel. It reinforces output verification and human oversight, but is an internal IRS policy—not an additional Circular 230 rule imposed on private practitioners. The June 2026 OPR guidance is the more directly relevant discussion for this publication.

Disclosure of sponsorship and limits of this research

This publication was prepared by ARNCO Development LLC, developer of CiteMax. It is a CiteMax-sponsored comparative evaluation, not an independent third-party study. Historical benchmark scores are retained as recorded and must not be represented as newly audited accuracy-only or workpaper-quality grades. The October 18 five-case study and the October 31 single-issue wash-sale comparison are separate exercises and are not pooled into an overall ranking.

Readers should consult the accompanying technical evidence ledger for original-source availability and unresolved gaps, and independently review every legal proposition before professional reliance. Original third-party responses and full CiteMax workpapers are not reproduced where the verified source files or reproduction permissions are unavailable.

Primary references for practitioner responsibility

IRS Office of Professional Responsibility, Issue 2026-19 (June 24, 2026): https://content.govdelivery.com/accounts/USIRS/bulletins/41d6e70

Circular 230: https://www.irs.gov/tax-professionals/office-of-professional-responsibility-and-circular-230

31 CFR §10.22: https://www.ecfr.gov/current/title-31/subtitle-A/part-10/section-10.22

31 CFR §10.35: https://www.ecfr.gov/current/title-31/subtitle-A/part-10/section-10.35

31 CFR §10.36: https://www.ecfr.gov/current/title-31/subtitle-A/part-10/section-10.36

31 CFR §10.37: https://www.ecfr.gov/current/title-31/subtitle-A/part-10/section-10.37

IRS IRM 10.24.1 (August 10, 2026): https://www.irs.gov/irm/part10/irm_10-024-001r

The essential distinction: answer versus workpaper

The research question is not merely whether a model reaches a plausible tax conclusion. A practitioner must also determine whether the response identifies the governing authority, exposes material assumptions, acknowledges contrary authority and can be retained as an auditable workpaper. These are related but distinct achievements.

Substantive accuracy: Does the conclusion follow from current, controlling law applied to the stated facts? A concise answer can earn a high score.

Professional workpaper quality: Is the analysis organized, sourced, qualified and documented sufficiently for practitioner review and retention? A correct answer can still be incomplete on this dimension.

Time and operational reliability: How long did it take, and did the system complete the assigned task? Report separately rather than folding speed into legal correctness.

CiteMax is designed to produce a structured tax research workpaper. General-purpose assistants are capable of producing substantial memoranda when prompted, but an ordinary answer should not automatically be treated as equivalent to a standardized professional workpaper. Conversely, CiteMax does not deserve a high legal-accuracy score merely because its report is longer or more polished.

Two independent evaluation scales

The two 100-point scales above are the proposed publication-standard rubric, not scores already calculated. The historical numerical tables later in this edition reflect the original blended rubric or previously recorded review scores. They must remain labeled as such until source-level rescoring is completed.

What the preserved record can actually establish

Illustrative comparison: response versus research workpaper

This is a deliverable-format comparison, not a scored result for every general-purpose model. Specific products should receive a separate workpaper grade only after examining the original prompt and output and applying the same required deliverable specification.

Executive findings

The central distinction is between a defensible answer and a defensible workpaper. This edition reviews the October 18 five-case diagnostic protocol, shorter specialist-response records, and the October 31 wash-sale comparison. These datasets address different questions and use different scoring histories; they are therefore presented separately, not combined into a single product league table.

CiteMax correctly distinguishes S-election termination from an automatic partnership conversion; the LLC classification case is its strongest retained full-case example.

The Hawaii workpaper properly identifies an inability to verify state authority in its session. That is commendable source discipline, but it still falls short of answering the state-law question.

Anonymized specialist responses include an affirmative misstatement on automatic LLC reclassification and a failure to distinguish the electronic IP PIN rejection deadline from the paper fallback.

The October 31 wash-sale comparison includes prior recorded assessments of CiteMax, Grok, Copilot, Perplexity, Claude and Gemini. ChatGPT and Duck.ai were within the broader program, but neither has an attributable October 31 wash-sale score in the records recovered for this edition.

How the scores were derived

Each criterion receives an editorial 0–100 assessment; the weighted total is rounded to the nearest point. A response that fails to provide an answer is recorded 0 as a response-completion failure—not a statement about the vendor’s normal capability. Where a primary legal outcome is demonstrably incorrect, the row is marked CRITICAL and cannot be described as passing regardless of score.

These are first-pass editorial judgments, not blinded independent reviewer scores. Source statements were spot-checked against selected IRS and Hawaii primary materials; full citation-by-citation verification remains outstanding.

Comparability and anonymization

CiteMax results: four fuller workpapers (cases 1–4) dated October 18, 2026; the fifth is not preserved as a CiteMax report.

Exploratory responses: shorter case prompts preserved in the “17 question test” collection, including anonymized specialist services A and B. These abbreviated questions were not equivalent to the complete CiteMax research assignments.

The specialist identities remain confidential. Only anonymized A and B appear in published tables; other specialist providers are listed as not available, not scored.

No cross-cohort “best overall” claim is justified: the five-case diagnostic and October 31 wash-sale study used different questions, grading histories and response versions. Scores should be compared only within the same exhibit and with the stated limitations.

Scorecard: CiteMax fuller workpapers

Scorecard: shorter exploratory responses

October 31 | Matched IRA wash-sale comparison

A separate October 31, 2026 exercise evaluated two presentations of the same wash-sale issue: a simplified question and a comprehensive prompt. The fact pattern was an IRA purchase Monday, IRA loss sale Tuesday, taxable-account purchase Wednesday and taxable loss sale Thursday, with no shares held after Thursday. This exhibit reproduces the numerical assessments recorded in the earlier comparison conversation; the numbers have not been independently regraded from verbatim transcripts in preparing this edition.

Interpretation: These are historical blended reviewer assessments, not independent accuracy-only or workpaper-quality grades. The comprehensive wording materially changed several outcomes; Copilot and Perplexity showed especially large differences. “<5 sec” is an upper bound, not an exact runtime. CiteMax durations are distinct runs, not median latency. No separate workpaper-quality scores can be inferred from this table.

Reading the October 31 scores correctly

The recovered historical scorecard measures overall response performance under two prompt styles. It cannot establish that CiteMax scored 84/100 for legal accuracy or that Gemini scored 65/100 for workpaper quality. The comparison instead supports a narrower observation: prompt specificity changed the quality of several answers, and one well-structured research deliverable may offer additional review value beyond the conclusion alone.

A model that accurately identifies the effect of IRC §1091 should receive credit even if its answer is brief.

The separate workpaper audit should check whether the same response preserves transaction chronology, sources, limits of Rev. Rul. 2008-5, factual assumptions, and an actionable conclusion.

For a fair production-format comparison, give every model the same instruction to provide a professional memorandum; also preserve an ordinary-answer test to measure out-of-the-box behavior.

Observed weaknesses documented during the comparison

Grok: treated Revenue Ruling 2008-5 as more directly decisive for the exact transaction sequence than its holding warrants.

Copilot and Perplexity: simplified-prompt responses altered or misunderstood material transaction facts; comprehensive prompts markedly improved their historical scores.

Claude: relied on excerpts/snippets without sufficiently demonstrating inspection of controlling primary-source text.

Gemini: treated a disputed deductibility issue as if conclusively settled rather than distinguishing statutory inference from a directly decided fact pattern.

CiteMax: showed limitations in secondary-source discovery, citation presentation and workpaper rendering; its legal interpretation likewise required careful qualification.

Why the wash-sale question is revealing

The question tests whether a research system tracks acquisition and disposition dates across an IRA and a taxable account, applies IRC §1091, distinguishes Revenue Ruling 2008-5 from an express holding on this exact four-day pattern, and preserves uncertainty. A fast answer that silently changes the dates or asserts an unexamined holding is less useful than a qualified, sourced analysis. The principal legal issue in the exact four-day pattern was not independently re-adjudicated from complete model transcripts for this edition. The historical scores are descriptive, not validated legal-accuracy grades.

October 20 | Five-question general-purpose AI record

An earlier conversation records receipt of 20 completed responses to five short tax-law questions from four general-purpose products: Grok, Microsoft Copilot, Gemini and Perplexity. The recorded turnaround was approximately 15–20 seconds per response, and Perplexity was provisionally considered strongest in procedural completeness. These are historical tracking facts, not authenticated scores or reproduced answer texts.

Comparability and verification rules

The October 18 five-case protocol is distinct from the October 31 single-issue wash-sale exercise. Do not average across those studies or claim that all products answered all five detailed October 18 cases.

The October 31 scores are reconstructed from a contemporaneous prior scorecard. Independent reproduction requires original model outputs and score sheets, dated prompts, model versions and verification of links/citations; these materials are not all available in this evidence-limited publication.

The October 20 summary establishes that testing occurred, but does not expose enough individual answer text to assign fresh scores. The original submissions should remain primary evidence when retrievable.

Generic AI systems appear by public name; specialized tax research services remain anonymized. Historical branding in source artifacts is presented as CiteMax throughout the main text.

Question 1 | S corporation shareholder health-insurance correction

Benchmark question

After 2025 S corporation and shareholder returns were filed, an omitted $14,400 shareholder health-insurance reimbursement was discovered. What reporting and return corrections are actually necessary?

Legal benchmark

Box 1 inclusion generally differs from FICA Boxes 3 and 5. W-2c is needed, while 941-X, 1120-S amendment, California corrections and 1040-X depend on what the original returns reported. Do not double-count an already deducted premium.

Retained-output assessment

What the result actually demonstrates

S corporation income-tax wages and FICA wages must be analyzed separately; mandatory amendment lists without comparing filed returns lose points.

The specialist responses use stronger categorical assertions than the facts warrant; the retained CiteMax report handles contingencies more carefully.

Evidence limitations

The CiteMax and shorter specialist outputs were not generated from the same complete prompt. Numerical differences within this exhibit therefore describe response quality observed in each record, not statistically controlled relative performance. Any missing vendor response remains unscored.

Question 2 | LLC S-election termination and entity classification

Benchmark question

A partnership-taxed LLC acquired an interest in an S-elected LLC on May 31, 2025. Did S status end, and does the operating LLC become a partnership or remain a corporation?

Legal benchmark

An ineligible partnership shareholder generally terminates S status. The association classification continues absent a valid separate entity-classification change. A short-year Form 1120, rather than automatic Form 1065, is the primary corrective starting point.

Retained-output assessment

What the result actually demonstrates

A partnership-owned shareholder is disqualifying for S status; it does not, standing alone, convert the operating LLC into a partnership for federal tax purposes.

Service B receives a critical-error flag because its default-reversion analysis conflicts with Treas. Reg. §301.7701-3(c)(1)(v)(C).

Evidence limitations

The CiteMax and shorter specialist outputs were not generated from the same complete prompt. Numerical differences within this exhibit therefore describe response quality observed in each record, not statistically controlled relative performance. Any missing vendor response remains unscored.

Question 3 | Hawaii bonus depreciation and suspended passive loss

Benchmark question

Does a 2025 Hawaii bonus-depreciation adjustment create current state taxable income when the federal rental loss is entirely suspended under §469?

Legal benchmark

Hawaii does not adopt federal bonus depreciation. Separately compute state depreciation and passive-activity limitations; a federal suspended loss does not itself determine current Hawaii inclusion. Distinguish the worksheet adjustment from income actually recognized.

Retained-output assessment

What the result actually demonstrates

A source-limited refusal to assert a Hawaii rule avoids invented authority, but cannot earn full legal-conclusion points.

Hawaii requires a separate depreciation adjustment; the relationship to §469 and state suspended amounts must be explained, not skipped.

Evidence limitations

The CiteMax and shorter specialist outputs were not generated from the same complete prompt. Numerical differences within this exhibit therefore describe response quality observed in each record, not statistically controlled relative performance. Any missing vendor response remains unscored.

Question 4 | Inherited IRA annual distributions and ten-year rule

Benchmark question

For an IRA owner who died in 2022 after the required beginning date, when must a non-eligible designated beneficiary take annual RMDs and empty the account?

Legal benchmark

For a 2022 death after the required beginning date, a non-eligible designated beneficiary generally has annual RMDs and a Dec. 31, 2032 full-distribution deadline. Transition excise-tax relief covers specified 2023–24 missed RMDs, but not 2025.

Retained-output assessment

What the result actually demonstrates

The age and account-value facts in the fuller benchmark must be used if a system received them; the retained CiteMax report requests some already-supplied information.

A ten-year end date does not eliminate annual beneficiary RMD requirements after an owner dies after the required beginning date.

Evidence limitations

The CiteMax and shorter specialist outputs were not generated from the same complete prompt. Numerical differences within this exhibit therefore describe response quality observed in each record, not statistically controlled relative performance. Any missing vendor response remains unscored.

Question 5 | Rejected e-file at extension deadline / IP PIN

Benchmark question

A 2025 Form 1040 transmitted on extended deadline Oct. 15, 2026 was rejected for an IP PIN. What electronic retransmission and paper-filing deadlines apply?

Legal benchmark

The IRS 2026 Form 1040 MeF schedule sets Oct. 20, 2026 for electronic retransmissions of rejected returns on extension. A separate paper fallback allows ten calendar days after notice of rejection (where applicable), with explanation and proof.

Retained-output assessment

What the result actually demonstrates

Electronic correction due October 20, 2026 and the ten-calendar-day paper-return fallback are distinct procedures.

Service B receives a critical-error flag for presenting the ten-day rule as the electronic correction deadline. Service A produced no substantive answer.

Evidence limitations

The CiteMax and shorter specialist outputs were not generated from the same complete prompt. Numerical differences within this exhibit therefore describe response quality observed in each record, not statistically controlled relative performance. Any missing vendor response remains unscored.

Appendix A | Reproducible scoring ledger

Component grades are shown in rubric order (legal, authority, uncertainty, date/jurisdiction, implementation, usability). Totals equal 30% / 25% / 15% / 10% / 10% / 10% of the listed component grades.

Case 1

Case 2

Case 3

Case 4

Case 5

Appendix B | Evidence inventory

Five-case diagnostic comparison packet (Oct. 18, 2026), approved detailed facts and 22 planned runs.

Five-case comparison workbook (retained Run Tracker and Scorecard remain unfilled; the present grades have not been inserted into those original experimental cells).

CiteMax case workpapers 1, 2, 3 and 4, dated Oct. 18, 2026; branding standardized in this manuscript.

“17 question test.pdf”: shortened five-question comparisons by anonymized specialist tax services; only external specialist services are included in the published scorecards.

October 20 conversation record: 20 complete general-purpose AI responses to five short questions were reported as received (Grok, Copilot, Gemini and Perplexity). The recoverable summary does not supply their verbatim text or case-by-case grades. October 31 conversation record: paired wash-sale scores and durations were recorded for CiteMax, Grok, Copilot, Perplexity, Claude and Gemini. Verbatim comparator responses have not been re-extracted for this version. No retained CiteMax report for detailed case 5.

Selected primary-source checks

IRS, “S corporation compensation and medical insurance issues”: https://www.irs.gov/businesses/small-businesses-self-employed/s-corporation-compensation-and-medical-insurance-issues

Treas. Reg. §301.7701-3(c)(1)(v)(C): https://www.ecfr.gov/current/title-26/section-301.7701-3

Hawaii Department of Taxation, 2025 Form N-11 Instructions, depreciation and gain adjustments: https://files.hawaii.gov/tax/forms/current/n11ins.pdf

IRS, Notice 2024-35 (2024 RMD transition relief): https://www.irs.gov/irb/2024-19_IRB

IRS, 2025/2026 Form 1040 MeF due dates: https://www.irs.gov/e-file-providers/tax-year-2025-processing-year-2026-form-1040-mef-due-dates

IRS, Publication 1345 (rejected paper-return fallback): https://www.irs.gov/pub/irs-pdf/p1345.pdf

CiteMax capabilities outside the scored comparisons

Product capabilities should not be confused with independently measured performance. CiteMax's research workflow is designed to support source restrictions defined by the practitioner, drafts of client-facing responses for professional review, structured research workpapers, and redaction of sensitive information in appropriate document-upload workflows. Staff-only document upload/redaction is distinct from the public research workflow. Feature availability, safeguards and reliability were not assessed through these historical benchmark scores; no comparison in this paper establishes that competing products lack equivalent features.

Versioning, practitioner approval and research correction workflows are separate product goals or operating features to test operationally, not evidence of superior legal conclusions. Any proposed capability should be validated on the specific production deployment before being described as generally available.

Editorial conclusion

The recovered historical wash-sale scorecard permits a readable, named-product comparison of the same two prompt variants. The five-case diagnostic and specialist exhibits provide complementary observations about tax-law workpaper quality, but are not a controlled full-roster ranking. A subsequent source-level validated edition would require the verbatim October 20 and October 31 responses and independent verification of material authorities and score assignments. CiteMax is owned and developed by ARNCO Development LLC; this is sponsor-authored research, not an independent third-party endorsement.

© 2026 ARNCO Development LLC. CiteMax is a trademark of its owner. This review is educational and not tax advice. Website publication edition: November 5, 2026. Historical scores remain provisional where original transcripts are unavailable.