Software and AI-generated claims

1Summary

Software has long been used to prepare tax returns. In research credit work, some tools gather payroll and project data and calculate the credit. Some draft the technical narrative that describes each project, using templates or generative artificial intelligence. Such tools can be sold for a fixed price or priced as a percentage of the credit they calculate; no official source records which is more common.

This paper does not describe any product or provider, and it does not suggest that automated tools produce worse claims than people do. No official source found for this paper measures that. It asks a narrower question: how do the existing federal rules apply when a claim, or part of one, is produced by software, and in particular when the software is priced as a share of the credit?

The paper finds that the legal requirements for a claim are unchanged by the way it is drafted. A refund claim must identify the business components and research activities it relates to, the IRS treats a recital of the legal definitions as insufficient, and the taxpayer verifies the facts under penalties of perjury.12 It finds that the federal rules on preparers and practitioners were written for people, that the statutory definition of a preparer excludes mechanical assistance, and that Treasury's contingent fee rule reaches practitioners rather than software providers as such.345 A platform priced as a share of the credit may therefore sit outside every federal fee rule. Swanson Reed's position is stated once, in a labeled box.

2What this paper covers

2.1Two kinds of automation

The paper uses two terms.

A percentage-priced claim platform is software, or a service delivered mainly through software, for which the business pays a fee calculated as a share of the credit it produces.
An automated technical narrative is a description of a project's research activities generated wholly or partly by software, whether from a template or by a generative artificial intelligence model.

The two can appear together or apart. A fixed-price tool can draft narratives. A percentage-priced service can use people to write them. The paper discusses each in general terms and does not suggest that any provider uses either in a particular way.

2.2What the paper does not cover

The paper does not assess the quality of any tool. It does not report how many claims are prepared with software, because no IRS form asks. Form 6765 and its instructions do not ask how a claim was prepared or how its preparer was paid.6 The paper describes rules and incentives, not measured outcomes.

3What a claim has to contain

3.1The legal tests

The credit is available only for qualified research, and the test is applied separately to each business component: each product, process, computer software, technique, formula or invention that is held for sale, lease or license, or used in the taxpayer's business.7 The research must be undertaken to discover information that is technological in nature and intended to be useful in developing a new or improved business component, and substantially all of the activities must be elements of a process of experimentation.7 The regulations require the taxpayer to "retain records in sufficiently usable form and detail to substantiate that the expenditures claimed are eligible for the credit".8

These are questions of fact about a particular business: what it tried to do, what it did not know, and how it worked toward an answer. Software can organize those facts. It cannot supply them.

3.2What the IRS requires with a refund claim

Since January 2022, a research credit claim for refund must identify all the business components to which it relates, identify all research activities performed for each business component, and give the totals of qualified wage, supply and contract research expenses.1 Until June 2024 a claim also had to name the individuals who performed each activity and the information each sought to discover. The IRS has waived those two items at the time of filing, but says they "may be requested if a refund claim involving the Research Credit is selected for examination".1

The IRS has said plainly what will not do. "An original submission or a subsequent response which only repeats the research credit legal definitions and requirements under the Internal Revenue Code or Treasury Regulations will be deemed insufficient."1 Its published best practice example describes, for each business component, what the taxpayer was trying to achieve and what it did, in terms specific to that component.1

IRS Chief Counsel's 2021 memorandum on refund claims adds two points. The taxpayer "must provide a declaration signed under the penalties of perjury verifying that the facts provided are accurate", which in most cases is the signature on the amended return.2 And if the taxpayer provides documents, including a credit study, it "must specify the exact page(s) that supports a specific fact. A mere volume of documents will not suffice."2 The memorandum also notes that the IRS may reject a claim "because it was too general".2

3.3What Form 6765 now asks

For original returns, Form 6765 now includes Section G, which asks for information about each business component, reported for at least 80 percent of qualified research expenses or the top 50 business components.6 Section G is optional for tax years beginning before 2026 and required for tax years beginning after 2025, with exceptions for qualified small businesses making the payroll tax election and for certain smaller filers on original returns.6 The business component name is to be "consistent with how you maintain the books and records that substantiate the qualified research activities".6 For amended returns claiming a new or larger credit, column 49(f) requires the information "as shown in Section G for each business component".6

4Automated technical narratives

4.1The precedent of the prepackaged claim

The concern about generic documentation is older than generative artificial intelligence. In 2008 the IRS told its examiners of "a growing trend among taxpayers, and their representatives, to submit prepackaged material to support research credit claims", usually in multiple binders, which "frequently fail to substantiate that the taxpayer paid or incurred qualified research expenses ('QREs') as claimed".9 The guide advised examiners to consider disallowing or returning a claim that "relies too heavily on high level estimations (i.e. interviews, estimates, judgment sampling, etc.)".9 The guide states that it is not an official pronouncement of IRS position; it is cited here as a record of what examiners were told.9

4.2What changes with generative tools

A template produces the same words for different projects. A generative model can produce different words for each project, fluently and quickly, from limited input. That is useful when the input is accurate: a model can turn an engineer's notes into a clear description. It is a risk when the input is thin, because a fluent narrative can describe uncertainty and experimentation in terms that fit the legal tests without the underlying facts having been established.

Nothing in the IRS requirements depends on who or what wrote the words. The test is whether the claim identifies the business components and activities, and whether the records substantiate them.18 A narrative that restates the four-part test in project-specific language may still be, in substance, a repetition of the legal definitions, which the IRS treats as insufficient.1 A narrative that cannot be traced to records and to the people who did the work leaves the taxpayer exposed if the claim is examined and the IRS asks for the names and information it waived at filing.1

4.3The signature stays with the taxpayer

The taxpayer signs the return or claim and verifies its facts under penalties of perjury.2 The IRS tells taxpayers that although a paid preparer signs the return, "you're ultimately accountable for the accuracy of every item reported on your return".10 If a credit is reduced, the taxpayer repays it with interest and, for an excessive claim for refund or credit, owes a penalty of 20 percent of the excess unless it shows reasonable cause.1112

Reasonable cause can include reliance on professional advice, but the regulations set conditions. Reliance on the advice of a professional "does not necessarily demonstrate reasonable cause and good faith"; it does so only if the reliance was reasonable and in good faith in all the circumstances.13 For the accuracy-related penalty, the regulations define advice as a communication "setting forth the analysis or conclusion of a person, other than the taxpayer".13 This paper found no regulation or published IRS guidance on whether output generated by software, without a person's analysis, is advice for that purpose.

5Percentage-priced claim platforms

5.1Why the fee basis matters more when the work is automated

Analysis (Swanson Reed's view)

Under a fixed fee or an hourly fee, the adviser's income follows the work. Under a percentage fee, it follows the credit. That is true whether the work is done by people or software, and it is a statement about incentives, not about any provider.

Automation sharpens the point. When much of the work is done by software, the cost of producing a larger claim may be little different from the cost of producing a smaller one, while a percentage fee rises in proportion to the credit. As an illustration only: on a claimed credit of $100,000, a fee of 20 percent is $20,000; if the same data were processed into a claim of $150,000, the fee would be $30,000. The additional work, for a tool, may be minimal.

Treasury has described the incentive in general terms. In December 2024 it said that "A practitioner with a direct, financial interest in the tax benefits of a client may be incentivized to increase such tax benefits."14 Treasury's definition of a contingent fee includes a fee "based on a percentage of the refund reported on a return" or "of the taxes saved".5 The definition does not distinguish between fees for human work and fees for automated work.

5.2The employee retention credit experience

The IRS's warnings about the employee retention credit, a different and refundable credit, show that it has treated both features as relevant. Its list of warning signs for that credit includes "Statements that the promoter or company can determine your ERC eligibility within minutes" and "Fees based on a percentage of the refund amount of Employee Retention Credit claimed".15 In 2025 Congress defined an employee retention credit promoter to include "any person" providing aid, assistance or advice on a credit document who charges a fee "based on the amount of the refund or credit", where that work is more than 20 percent of the person's gross receipts.16 The definition is not limited to individuals or to practitioners. It applies only to the employee retention credit, and nothing in this section is evidence about research credit software.

6How the existing rules apply to automated tools

6.1Is a software provider a tax return preparer?

A tax return preparer is any person who prepares for compensation, or employs others to prepare for compensation, any return or claim for refund, and preparing a substantial portion of a return or claim counts as preparing it.3 The statute excludes a person who merely "furnishes typing, reproducing, or other mechanical assistance".3 The regulations illustrate the mechanical assistance exception with a reporting agent who "did not render any tax advice to the client or exercise any discretion or independent judgment on the client's underlying tax positions".4

Software that only performs arithmetic on numbers the taxpayer has classified may fall within that exception. Software that decides which projects are business components, which activities qualify or which wages are qualified research expenses may be exercising the kind of judgment the exception describes as absent. The preparer penalty applies to a preparer whose return or claim understates tax because of an unreasonable position, and is the greater of $1,000 or 50 percent of the preparer's income from the return or claim.17 In 2000 the IRS Office of Chief Counsel noted that the IRS had applied the preparer definition to a company, and its computer programmer, furnishing computerized return preparation services where the program goes beyond mere mechanical assistance, citing revenue rulings from 1985.18 This paper found no IRS guidance applying that position to a provider of research credit software, as distinct from the people who use it.

For a different purpose, the regulations are explicit. For the rules on disclosure and use of tax return information under section 7216, a tax return preparer includes "a person who develops software that is used to prepare or file a tax return".19 A provider of research credit software is therefore bound by those confidentiality rules for the information it receives in connection with return preparation, even where its status for penalty purposes is unclear.19

6.2Circular 230 reaches practitioners, not software

Treasury's contingent fee rule applies to practitioners, a term Circular 230 defines by reference to the attorneys, certified public accountants, enrolled agents and others who may practice before the IRS.520 In 2014 the D.C. Circuit held that Treasury's authority to regulate practice does not extend to tax return preparers as such, and a district court then held that the IRS cannot apply its contingent fee rule to ordinary refund claims prepared before an audit.2122 A software provider that is not a practitioner is outside the rule. A practitioner who sells or uses a percentage-priced platform is subject to it, to the extent the rule can still be enforced.522

The AICPA Code prohibits its members in public practice from preparing an original or amended return or claim for refund for a contingent fee, and applies only to its members.23

6.3Duties of practitioners who use automated tools

For practitioners, Circular 230 already sets standards that apply to work produced with software. A practitioner must exercise due diligence in preparing or assisting in preparing returns and other papers.20 A practitioner is presumed to have exercised due diligence when relying on another person's work product only if the practitioner "used reasonable care in engaging, supervising, training, and evaluating the person".20 A practitioner may not willfully, recklessly or through gross incompetence sign a return or claim containing a position that lacks a reasonable basis or is an unreasonable position under section 6694.20 Competent practice "requires the appropriate level of knowledge, skill, thoroughness, and preparation necessary for the matter".20

Treasury's December 2024 proposal would add to the competence rule that "Competency includes understanding the benefits and risks associated with relevant technology that is used by the practitioner to provide services to clients".24 The proposal says that this standard is based on a comment to the American Bar Association's model rule on competence.24 It has not been finalized.24

6.4Figure 1: how the federal rules apply

Who prepares the claim	Contingent fee rule (Circular 230)	Preparer penalty (section 6694)	Confidentiality (section 7216)
A practitioner using software	Applies, subject to the 2014 court limits	Applies to the practitioner	Applies
A non-practitioner preparer using software	Does not apply	Applies to the preparer	Applies
A software provider performing only mechanical assistance	Does not apply	Excluded by statute	Applies to software developers
A software provider exercising judgment on tax positions	Does not apply unless the provider is a practitioner	Not addressed in guidance found for this paper	Applies to software developers

Sources: 31 CFR 10.27; 26 U.S.C. 6694 and 7701(a)(36); 26 CFR 301.7701-15(f) and 301.7216-1(b)(2); Loving v. IRS; Ridgely v. Lew.51734192122 The table is a simplified summary of general rules and is not advice on any particular arrangement.

7Questions a business can ask

The rules described above suggest some general questions for any business that uses software, or an adviser who uses software, to prepare a research credit claim. They are general information, not advice on any particular tool or engagement.

Can each statement be traced? The IRS asks for the business components and research activities, and may later ask who did the work and what each person sought to discover.1 A narrative is only as strong as the records and people behind it, and the taxpayer must keep those records.8
Who reviewed the output? If a practitioner is involved, Circular 230 requires due diligence and reasonable care in relying on the work of others.20 A business can ask who checked the calculations and narratives, and what they compared them with.
How is the tool or service priced? A fixed price or a subscription does not change with the credit. A percentage of the credit does. Treasury's definition of a contingent fee looks at how the fee is calculated, not at whether a person or a program did the work.5
Who will answer the IRS? If the claim is examined, someone has to explain how it was built. The IRS tells taxpayers to choose a preparer they can contact if the IRS examines the return.10 A business can ask whether the provider supports examinations and on what terms.
What happens to the data? Software developers are treated as tax return preparers for the rules on disclosure and use of tax return information.19 A business can ask how its payroll and project data will be used and retained.

See our disclaimer.

None of these questions assumes that software produces worse claims. They are the same questions a careful business would ask of any adviser, applied to a process in which more of the work is done by a program.

8What the record shows

The record shows that the requirements for a valid claim turn on facts specific to the taxpayer, that the IRS has rejected generic documentation for many years, and that the taxpayer verifies the claim and bears the consequences if it is reduced.1291011 It shows that the federal fee rules were written for practitioners, and that the definition of a preparer was written before software could exercise judgment on tax positions.35

9What the record does not show

The record does not show how many research credit claims are prepared with software or with generative tools, how such claims fare on examination, or how software used for research credit claims is priced. No IRS form asks.6 This paper makes no estimate of any of those things. It found no IRS guidance specific to generative artificial intelligence in research credit claims.

In our view, software can make a claim faster, cheaper and better organized, and for many businesses that is a real benefit. A fixed-price tool used by a diligent adviser carries no fee incentive at all. The concern this paper describes is narrower: a fee that rises with the credit, attached to a process whose cost does not, and delivered by a provider that the federal fee rules may not reach.

Our position

Congress should prohibit contingent fees, commonly called success fees, in federal research and development tax credit engagements under Section 41, so that advisers are paid for the work they perform rather than for the size of the claim they produce. Swanson Reed, which publishes this paper, is a fixed-fee R&D tax adviser. It has a direct commercial interest in the outcome: SR Tax Services LLC does not charge success fees, and a ban would remove a pricing model it competes against.

References

Internal Revenue Service, Research credit claims (Section 41) on amended returns frequently asked questions, introduction and questions 4, 7, 20 and 21 (page last reviewed or updated 1 October 2026). https://www.irs.gov/businesses/corporations/research-credit-claims-section-41-on-amended-returns-frequently-asked-questions (accessed 11 October 2026). Back
IRS Office of Chief Counsel, Memorandum 20214101F, I.R.C. 41 Research Credit Refund Claims, released 15 October 2021, Short Answers 2 and Law and Analysis part A. https://www.irs.gov/pub/irs-lafa/20214101f.pdf (accessed 11 October 2026). Back
Office of the Law Revision Counsel, 26 U.S.C. 7701, Definitions, subsection (a)(36)(A) and (B), United States Code 2024 edition published by the Government Publishing Office. https://www.govinfo.gov/content/pkg/USCODE-2024-title26/html/USCODE-2024-title26-subtitleF-chap79-sec7701.htm (accessed 11 October 2026). Back
Treasury Department, 26 CFR 301.7701-15, Tax return preparer, paragraphs (a), (f)(1)(viii) and (f)(6), Example 1. https://www.ecfr.gov/current/title-26/section-301.7701-15 (accessed 11 October 2026). Back
Treasury Department, Circular No. 230, 31 CFR 10.27, Fees, paragraphs (b) and (c). https://www.ecfr.gov/current/title-31/subtitle-A/part-10/subpart-B/section-10.27 (accessed 11 October 2026). Back
Internal Revenue Service, Instructions for Form 6765 (Rev. December 2025), Section G, "Requirements To Complete Section G", "How To Complete Section G" and columns 49(c) and 49(f); the instructions do not ask how a claim was prepared or how its preparer was paid. https://www.irs.gov/instructions/i6765 (accessed 11 October 2026). Back
Office of the Law Revision Counsel, 26 U.S.C. 41, Credit for increasing research activities, subsection (d)(1) and (d)(2), United States Code 2024 edition published by the Government Publishing Office. https://www.govinfo.gov/content/pkg/USCODE-2024-title26/html/USCODE-2024-title26-subtitleA-chap1-subchapA-partIV-subpartD-sec41.htm (accessed 11 October 2026). Back
Treasury Department, 26 CFR 1.41-4, Qualified research for expenditures paid or incurred in taxable years ending on or after December 31, 2003, paragraph (d). https://www.ecfr.gov/current/title-26/section-1.41-4 (accessed 11 October 2026). Back
Internal Revenue Service, Large and Mid Size Business Division, Research Credit Claims Audit Techniques Guide (RCCATG): Credit for Increasing Research Activities Section 41, LMSB-04-0508-030 (May 2008), introductory note, Chapter I "Background" and Chapter II. https://www.irs.gov/businesses/research-credit-claims-audit-techniques-guide-rccatg-credit-for-increasing-research-activities-section-41 (accessed 11 October 2026). Back
Internal Revenue Service, Topic no. 254, How to choose a tax return preparer (page last reviewed or updated 7 October 2026). https://www.irs.gov/taxtopics/tc254 (accessed 11 October 2026). Back
Office of the Law Revision Counsel, 26 U.S.C. 6676, Erroneous claim for refund or credit, subsections (a) and (b), United States Code 2024 edition published by the Government Publishing Office. https://www.govinfo.gov/content/pkg/USCODE-2024-title26/html/USCODE-2024-title26-subtitleF-chap68-subchapB-partI-sec6676.htm (accessed 11 October 2026). Back
Office of the Law Revision Counsel, 26 U.S.C. 6601, Interest on underpayment, nonpayment, or extensions of time for payment, of tax, subsection (a), United States Code 2024 edition published by the Government Publishing Office. https://www.govinfo.gov/content/pkg/USCODE-2024-title26/html/USCODE-2024-title26-subtitleF-chap67-subchapA-sec6601.htm (accessed 11 October 2026). Back
Treasury Department, 26 CFR 1.6664-4, Reasonable cause and good faith exception to section 6662 penalties, paragraphs (b)(1) and (c)(1) and (2). https://www.ecfr.gov/current/title-26/section-1.6664-4 (accessed 11 October 2026). Back
Treasury Department and Internal Revenue Service, Regulations Governing Practice Before the Internal Revenue Service, notice of proposed rulemaking, REG-116610-20, 89 FR 104915 (26 December 2024), Explanation of Provisions part B, page 104918. https://www.federalregister.gov/documents/2024/12/26/2024-29371/regulations-governing-practice-before-the-internal-revenue-service (accessed 11 October 2026). Back
Internal Revenue Service, Employee Retention Credit, section "Beware of ERC scams" (page last reviewed or updated 8 October 2026). https://www.irs.gov/coronavirus/employee-retention-credit (accessed 11 October 2026). Back
Public Law 119-21, 139 Stat. 72 (4 July 2025), section 70605(b) and (c). https://www.govinfo.gov/content/pkg/PLAW-119publ21/html/PLAW-119publ21.htm (accessed 11 October 2026). Back
Office of the Law Revision Counsel, 26 U.S.C. 6694, Understatement of taxpayer's liability by tax return preparer, subsection (a), United States Code 2024 edition published by the Government Publishing Office. https://www.govinfo.gov/content/pkg/USCODE-2024-title26/html/USCODE-2024-title26-subtitleF-chap68-subchapB-partI-sec6694.htm (accessed 11 October 2026). Back
IRS Office of Chief Counsel, Memorandum WTA-N-115554-00, Paid Preparer Issues: Internet Returns (19 October 2000), page 4, citing Rev. Rul. 85-187, Rev. Rul. 85-188 and Rev. Rul. 85-189. https://www.eitc.irs.gov/pub/lanoa/pmta00402_7127.pdf (accessed 17 October 2026). Back
Treasury Department, 26 CFR 301.7216-1, Penalty for disclosure or use of tax return information, paragraph (b)(2)(i)(B) and paragraph (b)(3)(ii), Example 1. https://www.ecfr.gov/current/title-26/section-301.7216-1 (accessed 11 October 2026). Back
Treasury Department, Circular No. 230, 31 CFR part 10, sections 10.2(a)(5), 10.3, 10.22(a) and (b), 10.34(a)(1) and 10.35(a). https://www.ecfr.gov/current/title-31/subtitle-A/part-10 (accessed 11 October 2026). Back
US Court of Appeals for the District of Columbia Circuit, Loving v. Internal Revenue Service, No. 13-5061, decided 11 February 2014, slip opinion pages 2 to 3 and 17 to 18. https://media.cadc.uscourts.gov/opinions/docs/2014/02/13-5061-1479431.pdf (accessed 11 October 2026). Back
US District Court for the District of Columbia, Ridgely v. Lew, Civil Action No. 1:12-cv-00565 (CRC), Memorandum Opinion, 16 July 2014, pages 1 and 14 to 15. https://www.govinfo.gov/content/pkg/USCOURTS-dcd-1_12-cv-00565/pdf/USCOURTS-dcd-1_12-cv-00565-1.pdf (accessed 11 October 2026). Back
American Institute of CPAs, AICPA Code of Professional Conduct (updated for all official releases through September 2026), ET 1.510.001 Contingent Fees Rule, paragraph .01. https://pub.aicpa.org/codeofconduct/ethicsresources/et-cod.pdf (accessed 11 October 2026). Back
Treasury Department and Internal Revenue Service, Regulations Governing Practice Before the Internal Revenue Service, notice of proposed rulemaking, REG-116610-20, 89 FR 104915 (26 December 2024), Explanation of Provisions part I and proposed section 10.35(a), pages 104920 and 104929. https://www.federalregister.gov/documents/2024/12/26/2024-29371/regulations-governing-practice-before-the-internal-revenue-service (accessed 11 October 2026; Federal Register search on that date showed no final rule). Back

Cite this paper

Summary

1Summary

2What this paper covers

2.1Two kinds of automation

2.2What the paper does not cover

3What a claim has to contain

3.1The legal tests

3.2What the IRS requires with a refund claim

3.3What Form 6765 now asks

4Automated technical narratives

4.1The precedent of the prepackaged claim

4.2What changes with generative tools

4.3The signature stays with the taxpayer

5Percentage-priced claim platforms

5.1Why the fee basis matters more when the work is automated

5.2The employee retention credit experience

6How the existing rules apply to automated tools

6.1Is a software provider a tax return preparer?

6.2Circular 230 reaches practitioners, not software

6.3Duties of practitioners who use automated tools

6.4Figure 1: how the federal rules apply

7Questions a business can ask

8What the record shows

9What the record does not show

References

About this paper

Cite this paper

Summary

1Summary

2What this paper covers

2.1Two kinds of automation

2.2What the paper does not cover

3What a claim has to contain

3.1The legal tests

3.2What the IRS requires with a refund claim

3.3What Form 6765 now asks

4Automated technical narratives

4.1The precedent of the prepackaged claim

4.2What changes with generative tools

4.3The signature stays with the taxpayer

5Percentage-priced claim platforms

5.1Why the fee basis matters more when the work is automated

5.2The employee retention credit experience

6How the existing rules apply to automated tools

6.1Is a software provider a tax return preparer?

6.2Circular 230 reaches practitioners, not software

6.3Duties of practitioners who use automated tools

6.4Figure 1: how the federal rules apply

7Questions a business can ask

8What the record shows

9What the record does not show

References

About this paper

Related papers