Three prospective clients have walked into our office over the last six weeks with cost segregation studies they generated themselves using generative artificial intelligence (two with ChatGPT, one with Google Gemini), and asked our licensed engineering firm to "validate" the work. We declined in each case.
The questions those requests raised reach beyond the sales conversation to the substance of what constitutes a defensible cost segregation study under the IRS Audit Technique Guidelines, and to the exposures a taxpayer faces under Section 6662 of the Internal Revenue Code and a preparer faces under Section 6694 when the underlying work product does not satisfy the framework.
The framework
The ATG is explicit that no methodology is mandated: "Neither the Internal Revenue Service nor any group or association of practitioners has established any requirements or standards for the preparation of cost segregation studies." What it does is identify six methodologies ranked by reliability, with the Detailed Engineering Approach from Actual Cost Records and the Detailed Engineering Cost Estimate Approach at the top, and describe 13 "principal elements of a quality cost segregation study" that an IRS examiner is trained to look for on audit.
Those elements include preparation by a qualified individual with expertise; description of methodology; use of appropriate documentation; interviews with the property owner and facility manager; legal analysis with citations to authority; unit costs and engineering take-offs; reconciliation to actual cost records; identification of Section 1245 property; and treatment of indirect costs under Section 263A. Element 1 (qualified preparer) is binary. Elements 8, 10 and 12 (take-offs, reconciliation and Section 1245) are analytical work products, not document possession.
Three studies against the framework
The Gemini output consisted of six rows in a spreadsheet: five heuristic categories summing to $100,000 of bonus-eligible basis on a $400,000 property. No methodology, no preparer, no documentation. To be fair, Gemini did not represent the output as a study.
The ChatGPT report is analytically more interesting. The client is the developer of a $4.7 million new-construction athletic facility. He uploaded the general contractor's cost breakdown, including trade-level invoices. The output is an Excel workbook: a Trade Allocation Matrix reclassifying $1.24 million (26.1% of basis) to short-life assets, plus five supporting tabs holding the contractor's invoices as embedded PNG images. Document possession is satisfied. No methodology is stated. No engineer of record. No take-offs. No reconciliation between the invoices and the trade percentages. No Section 1245 statutory analysis. Substantial trade-level amounts, including tilt-wall erector ($480,000), plumbing ($237,000), roofing ($201,000), painting ($118,000), architect fees ($99,000) and combined construction management ($440,000), are reclassified at 0% to short-life assets, totaling approximately $1.59 million of basis for which a competent engineering analysis would identify partial reclassification opportunities.
The comparator is a 139-page engineering-based study our firm prepared on a $43.9 million high-rise apartment property. It includes the following: A named engineer of record. Firm certification. Florida P.E. license CA-27824 and Texas P.E. license F-12634. The methodology is named explicitly: both top-ranked ATG methodologies, applied together. A legal analysis citing Hospital Corp. of America & Subsidiaries v. Commissioner, 109 T.C. 21 (1997), IRC Section 263A, and the Tax Reform Act of 1986. Component-level breakdown; indirect cost allocation; Section 1245 property identified separately from Section 1250; unit of property analysis; and disposition analysis.
Scored against the 13 ATG elements: ETS 13. ChatGPT 1. Gemini 0. ChatGPT satisfies element 3 (documentation) because the invoices are in the file, but fails the 12 elements that require analytical work performed by a qualified individual on those documents.
Taxpayer exposure under Section 6662
The 20% accuracy-related penalty (40% under Section 6662(h)) applies on substantial understatements. The reasonable cause defense under Section 6664(c) turns on Neonatology Associates, P.A. v. Commissioner, 115 T.C. 43 (2000), aff'd 299 F.3d 221 (3d Cir. 2002), and its three-prong test: the advisor was a competent professional with sufficient expertise; the taxpayer provided accurate information; and the taxpayer relied in good faith. An AI tool fails the first prong. It is not a "competent professional" within the meaning of Neonatology. There is no licensure, no Circular 230 obligation, and no professional judgment to rely upon. The defense fails at the threshold, and the taxpayer carries the underpayment plus penalty.
Preparer exposure under Section 6694
The greater of $1,000 or 50% of the fee derived attaches when a return position lacks substantial authority (or, if disclosed, a reasonable basis). AICPA Statement on Standards for Tax Services No. 1 sets the practitioner's own standard at a realistic possibility of being sustained on the merits, and Circular 230 §10.22 imposes a due diligence obligation on representations made to the IRS. A CPA signing a return that reflects cost segregation classifications drawn from a work product that satisfies one of 13 ATG quality elements will find substantial authority difficult to articulate. Disclosure on Form 8275 shifts the standard to reasonable basis but does not eliminate exposure. Several professional liability carriers have issued updated 2025 and 2026 guidance addressing generative AI in return preparation. The practical implication when a client brings a CPA an AI cost segregation study is that the appropriate response is rarely "we'll file it." It is more often "we'll need an engineering-based study from a qualified firm before I sign this return."
The frame
This is not anti-AI. Our firm uses generative AI substantially for research, drafting and analytical support, and expects the role to expand. The line the ATG framework surfaces is professional accountability for a tax position. AI tools cannot be the preparer of a tax position. They cannot be subject to Circular 230, cannot defend a study under exam, cannot be professionally disciplined. They can be used by licensed professionals who are the named preparer. A cost segregation study prepared by a qualified engineering firm may legitimately use AI in document review or analytical support, so long as the engineering work is performed by qualified individuals whose names appear on the report.
The IRS examiner's opening question on audit will not be whether AI was involved in preparing the position. It will be the five document requests that have opened cost segregation examinations for a decade: the engineering report with classifications and reasoning; timestamped photographic evidence; the engineer of record's credentials; reconciliation to actual cost records; and documentation of the methodology applied. Whether AI was used in preparing those work products is largely irrelevant. Whether the work products exist, and whether a qualified preparer's name stands behind them, is dispositive. For practitioners advising clients on cost segregation positions in 2026, the conversation must begin and end there.
