Form 6765 Section G Becomes Mandatory in 2026. Most AI-Only R&D Credit Tools Will Not Be Ready.

The IRS finalized the redesigned Form 6765 instructions on February 5, 2026, after a two-year drafting process and an extended public comment period. For tax years beginning in 2026, with returns filed in 2027, most R&D credit filers will be required to complete Section G of the form. Section G itemizes qualified research expenses by business component, breaks employee wages into direct research, supervision, and support categories, and ties every dollar of claimed credit to a specific project narrative the IRS can evaluate against the Four-Part Test under IRC Section 41(d).

This is the same level of detail the IRS has been requesting on examination for the past five years. As of tax year 2026, it is the standard for the original return, not just the audited one.

For taxpayers whose R&D credit was filed through an AI-driven questionnaire platform, the question is whether the file produced by that workflow can survive Section G's reporting requirements without retroactive reconstruction. Across hundreds of second-opinion reviews conducted by Strike Tax Advisory, the answer has typically been no. This article explains why, and what a defensible R&D credit study looks like in 2026.

The Rise of Automated R&D Tax Credit Platforms

Over the past few years, a category of platforms has emerged that promises to compute and file the R&D tax credit largely or entirely through software. Typical workflows ask the taxpayer to complete a questionnaire about their business activities, upload payroll data, and accept a calculated credit amount with minimal human involvement. Some platforms market themselves as a replacement for traditional CPA-led R&D studies. Others position themselves as a tool that allows generalist CPAs to deliver the credit at a lower cost.

The underlying technology is genuine. Large language models are good at parsing forms, summarizing text, classifying transactions, and producing first-draft narratives. What they are not good at is making the legal and technical judgments that determine whether a given research activity qualifies under Section 41. That distinction matters because, on examination, the IRS does not ask whether the questionnaire was thorough. It asks whether the activities and expenses claimed meet the statutory and regulatory tests, and whether the documentation supports the position.

What Has the IRS Actually Said About AI in Tax Preparation?

The IRS has spoken on this topic in three places that matter for any business evaluating an AI-driven R&D credit provider.

1. Internal IRS AI policy directs employees to avoid over-reliance on AI

The IRS published its internal policy on AI governance as Internal Revenue Manual Section 10.24.1. The policy instructs IRS personnel to use large language models as tools to assist and augment their work, not to replace their own critical thinking and judgment, and to avoid becoming overly reliant on these models. The agency that will scrutinize R&D credit claims has explicitly told its own staff that AI cannot replace professional judgment. That is a meaningful signal about how the IRS views AI-generated work product from outside the agency as well.

2. The Taxpayer Advocate Service has warned taxpayers

The Taxpayer Advocate Service, an independent organization within the IRS, has cautioned taxpayers against relying solely on AI-generated tax advice. The Advocate cited research finding that AI chatbots from major tax preparation companies provided inaccurate answers on complex tax questions a significant share of the time. Independent benchmarks have produced similar results. A 2025 benchmark called TaxCalcBench found that frontier AI models correctly computed fewer than one-third of federal-only returns when the document collection stage was already completed correctly. The R&D credit is more complex than a standard federal return, not less.

3. Reliance on AI does not qualify as reasonable cause for penalty relief

Under IRC Section 6662, accuracy-related penalties of 20 percent of the underpayment apply when a taxpayer takes a position without substantial authority, or when an underpayment is attributable to negligence or disregard of rules. A taxpayer can request relief if there was reasonable cause and the taxpayer acted in good faith. Telling the IRS that an AI tool produced the number does not meet that standard. Reasonable cause requires evidence of professional review, contemporaneous analysis of the facts, and reliance on a qualified preparer who has the expertise and access to the relevant information. An AI questionnaire produces none of those things in a form an examiner would accept.

Why Does the R&D Tax Credit Resist Automation?

The R&D tax credit is not a checkbox item. Qualification is determined by applying the Four-Part Test under IRC Section 41(d) to each business component the taxpayer wants to include. Every step in the test requires judgment that depends on facts only a trained professional can elicit and evaluate.

The exclusions are where AI gets caught

Beyond the Four-Part Test, Section 41(d)(4) lists eight categories of activity that are statutorily excluded from the credit. Two of these are common sources of disallowance on examination and require careful analysis that AI questionnaires routinely miss.

Funded research under IRC Section 41(d)(4)(H) excludes research to the extent it is funded by any grant, contract, or otherwise by another person or governmental entity. Whether a contract constitutes funded research depends on whether the taxpayer retains substantial rights in the research and bears the financial risk of failure. This is a contract-by-contract analysis that requires reading the funding agreement, understanding the milestone structure, and applying the Fairchild Industries and Geosyntec Consultants lines of authority. An AI questionnaire that simply asks whether the taxpayer received any grants misses the analysis entirely.

Internal-use software regulations under Treasury Regulations Section 1.41-4 apply a heightened three-part high threshold of innovation test for software developed primarily for the taxpayer's own internal operations. Whether software qualifies as internal-use or non-internal-use, and whether the dual-function safe harbor applies, requires reading code, interviewing developers, and understanding customer-facing functionality. None of this fits in a questionnaire field.

The math is judgment, not arithmetic

Even once qualifying activities are identified, the credit calculation requires choices that change the outcome materially. Taxpayers can elect the regular credit method or the Alternative Simplified Credit. They must decide whether to make the Section 280C(c)(2) reduced credit election on a timely filed original return, a choice that for pass-through entities at the 37 percent federal rate plus state often produces a better outcome than the full credit because the preserved deduction value exceeds the credit reduction cost. Members of a controlled group must aggregate gross receipts and QREs under Section 41(f) and the related rules under Section 52, then allocate the group credit among members. For tax years beginning after December 31, 2024, the interaction between Section 41 and the new Section 174A immediate expensing rules added by the One Big Beautiful Bill Act introduces additional planning decisions. These are strategic elections, not data entry, and the wrong choice can leave six or seven figures of credit on the table or trigger problems on examination.

Where AI Belongs in an R&D Credit Study, and Where It Does Not

The argument here is not that AI has no role in R&D credit work. AI is a legitimate tool for the same reason calculators, document parsers, and time-tracking software are legitimate tools. The question is what AI does well and where the judgment has to remain with a licensed professional.

AI does well at ingesting and classifying transaction-level general ledger data, parsing engineering documentation for technical uncertainty signals, generating first-draft business-component narratives from interview transcripts, and identifying potential controlled group members from corporate disclosure databases. Strike Tax Advisory uses AI for all of these tasks. It accelerates the work, reduces cost, and lets our technical staff spend more time on the parts of the engagement that require judgment.

AI does not do well at making the legal determination of whether a specific activity satisfies the elimination-of-uncertainty prong, applying the funded research exclusion to a specific contract, deciding between the regular credit and the Alternative Simplified Credit for a given client, coordinating the Section 41 credit with the new Section 174A immediate expensing election, or evaluating Section 280C(c)(2) against a client's specific marginal rate. These are judgment calls. They require a CPA who can sign the return and a technical reviewer who can defend the position on examination.

The distinction is whether the AI is a tool used by a professional, or whether the AI is the professional. The IRS's own AI governance policy (IRM 10.24.1) draws the same line for its own staff. Strike operates on the same standard.

The 2026 Documentation Bar: Section G in Practice

The case for human-led R&D credit studies became substantially stronger when the IRS finalized the redesigned Form 6765 and its accompanying instructions on February 5, 2026. The agency has been clear that detailed business-component reporting is the new floor.

Once Section G applies, the taxpayer must report each business component generating the credit, describe the research activities performed and the information sought to be discovered, and break out qualified employee wage expenses, qualified supply expenses, and qualified contract research expenses by component. Wages must be subdivided into direct research, supervision of research, and support of research. Taxpayers must report business components representing 80 percent of the total QRE in descending order, capped at 50 components.

The detail required by Section G is the same detail the IRS has been requesting on examination for years. It is also the same detail required for a valid research credit refund claim filed on an amended return. IRS published guidance states that a refund claim must identify all business components to which the Section 41 claim relates for that year, identify all research activities performed for each business component, and provide the total qualified employee wage expenses, supply expenses, and contract research expenses. Claims that fail to meet this standard are not valid claims at all.

The 45-day claim perfection extension is a transition mechanism. It is not a permanent grace period, and it does not change the underlying substantiation standard.

An AI-generated questionnaire output is not the same artifact as the contemporaneous business-component documentation that Form 6765 Section G requires. Producing the required detail at the level the IRS now expects requires interviews, code review where applicable, project-level cost tracking, and a tax professional who can map the facts to the statute. That is work, not a software output.

The Section 174A Window: A Door That Is Closing

The R&D credit conversation in 2026 is incomplete without a separate analysis of Section 174A. The One Big Beautiful Bill Act (Public Law 119-21) enacted Section 174A, which restores immediate expensing for domestic research and experimental expenditures and provides retroactive relief for small businesses.

Companies with average annual gross receipts of $31 million or less can amend their 2022, 2023, and 2024 returns to immediately expense R&E expenditures that were previously required to be capitalized and amortized under the prior version of Section 174. This is a refund opportunity that is separate from the R&D credit itself. A company that previously claimed no credit, or claimed a credit through any provider, may still have a Section 174A refund available.

The Section 6511 statute of limitations on refund claims means the amendment window for tax year 2022 closes in mid-2026 for most calendar-year taxpayers. Tax years 2023 and 2024 remain open longer, but the planning is most powerful when all three years are evaluated together because the cumulative deduction recapture and the credit coordination both depend on the multi-year position.

AI-only providers were not built for this work. The Section 174A amendment is not a credit calculation, it is a tax accounting method change with multi-year coordination implications. It requires a CPA to prepare and sign the amended returns, a technical review of which expenditures qualify under Section 174 versus Section 174A versus pure Section 162 deductibility, and coordination with the Section 41 credit position for the same years. Strike conducts Section 174A eligibility reviews as part of every new engagement, including for prospects who are already working with another R&D credit provider, because the amendment is a separate filing that does not displace the existing credit work.

If your current advisor has not raised Section 174A with you in 2026, it is worth asking why. For sub-$31 million companies, this can produce refunds independent of the credit itself.

Who is Liable When an R&D Tax Credit Filing is Wrong?

This is the part of the analysis most prospects do not see until it is too late. When an R&D credit claim is reduced or disallowed on examination, the consequences fall on the taxpayer.

• Repayment of the disallowed credit, plus statutory interest from the original due date of the return under IRC Section 6601.
• Accuracy-related penalty of 20 percent of the underpayment under IRC Section 6662, applied when there is negligence, disregard of rules, or a substantial understatement.
• Civil fraud penalty of 75 percent of the underpayment under IRC Section 6663 in cases involving intentional misstatement.
• Professional time and cost defending the examination, often a multi-year process the original software vendor does not participate in.

Software providers typically include limitation-of-liability clauses in their terms of service that cap the vendor's exposure at the fees paid. The taxpayer bears the rest. There is no version of the law in which the algorithm pays the penalty.

Where AI-Only R&D Credit Studies Tend to Fall Short

Across hundreds of credit studies and second-opinion reviews, several failure patterns repeat in claims produced by AI-only platforms.

Activities are claimed without an experimentation analysis. A questionnaire that asks whether a team worked on improving a product or process returns a yes from almost every operating business. That answer does not establish that the work satisfies the elimination-of-uncertainty and process-of-experimentation prongs. On examination, the activity is removed.

Funded research is missed. Government contractors, SBIR recipients, and companies performing research under cost-plus arrangements often have funded research exclusions that reduce or eliminate the eligible QRE base. Without a contract review, the exclusion is not applied. The credit is overstated.

Contract research is captured at 100 percent rather than 65 percent. Section 41(b)(3) limits contract research expenses to 65 percent of amounts paid to third parties for qualified services performed in the United States. AI tools that ingest accounts payable detail without classification routinely include the full invoice amount.

Wages are not properly nexused. Qualified wages under Section 41(b)(2)(A) are limited to amounts paid for qualified services, which means direct research, direct supervision of research, or direct support of research. A questionnaire asking what percentage of time employees spent on R&D does not produce the nexus documentation an examiner expects.

Offshore wages are included improperly. Qualified wages are limited to amounts paid for services performed in the United States. Companies with significant offshore development teams (India, Eastern Europe, Latin America) often see AI tools include those wages because the questionnaire does not separate by location. The credit is materially overstated and the position is indefensible on examination.

Controlled group rules are not applied. Section 41(f) requires aggregation of QREs and gross receipts across a controlled group. AI tools that work entity-by-entity miss the aggregation entirely, producing credits that the regulations do not allow, or missing credit-multiplying opportunities in the other direction.

State R&D credits are left on the table. Over 35 states offer their own R&D credits with different qualification rules, base-period definitions, and refundability provisions. Federal-only AI tools do not generate the state filings that often produce additional cash benefit.

What to Ask Any R&D Credit Provider Before You Sign

Whether the provider is a national firm, a generalist CPA, or a software platform, these questions surface the differences quickly. There are no wrong answers here, only honest ones.

1. Will a licensed CPA review and sign off on the final study? The credit study should be defensible by a licensed professional who exercises independent judgment.
2. Does the team include engineers, software developers, or scientists with subject-matter expertise in my industry? The Four-Part Test cannot be applied without technical fluency.
3. How will you document the elimination of uncertainty and the process of experimentation for each business component? If the answer is that the taxpayer fills in a questionnaire, the documentation will not survive a Form 6765 Section G review.
4. How do you evaluate the funded research exclusion under Section 41(d)(4)(H)? The provider should describe a contract review process, not a yes-or-no question.
5. How do you handle controlled group aggregation under Section 41(f)? If the provider does not ask about related entities, the credit may not stand.
6. Have you evaluated my company's Section 174A amendment opportunity for 2022 through 2024? For sub-$31 million companies, this is a separate refund opportunity from the credit itself, and the Section 6511 window for 2022 is closing.
7. Will you defend the credit through an IRS examination, and at what cost? Look for a clear answer in writing, not a marketing line.
8. What is your professional liability insurance coverage, and is your audit defense contractual? CPAs and tax counsel carry policies. Software vendors typically do not, and their liability is capped at fees paid.
9. Can I review the complete credit file, including the business-component narratives and the Section G data, before it is filed? If the answer is no, the provider is not building an examination-ready file.

What a Defensible R&D Credit Study Looks Like in 2026

Every Strike engagement combines four roles that AI alone cannot replace, plus a contractual audit-defense layer that AI-only providers structurally cannot offer.

A licensed CPA reviews and signs off on every study. Strike's CPAs are the same professionals who will represent the client if the credit is examined. They carry professional liability coverage and can speak to the file under oath. This is a structural difference from software-only providers whose terms of service cap liability at the software fee.

A technologist or engineer with industry-specific experience conducts the qualifying-activity interviews. For software clients, this is typically an engineer who has shipped production code. For manufacturers, an engineer with NADCAP, ISO, or AS9100 process exposure. For life sciences, a scientist with regulatory submission experience. The technical interviewer is the person who determines, in real time, whether an activity satisfies the elimination-of-uncertainty and process-of-experimentation prongs. That determination cannot be outsourced to a questionnaire.

A tax counsel layer coordinates the credit with Section 174, Section 174A, and state positions. This includes applying the Section 41(f) controlled group rules, evaluating the Section 280C(c)(2) reduced credit election against the client's specific marginal rate, coordinating the Section 174A retroactive amendment opportunity for tax years 2022 through 2024, and confirming state credit positions in the more than 35 jurisdictions that offer them. For a pass-through entity at the 37 percent federal rate plus state, the Section 280C reduced credit election is often superior to the full credit because the preserved deduction value exceeds the credit reduction cost. AI questionnaires default to the full credit and leave money on the table.

A documentation specialist builds the business-component narratives to the Section G standard. Each business component is mapped to qualifying activities, time-tracked wages are nexused to direct research, supervision, and support categories, and contract research is properly limited to 65 percent under Section 41(b)(3). The deliverable is an examination-ready file from day one, not a number with reconstruction work waiting to happen.

STRIKE Shield audit defense and fee guarantee

The structural difference that matters most on examination is what happens when the IRS asks questions. Strike's audit defense product, STRIKE Shield, is contractual, not promotional.

Under STRIKE Shield, if the IRS examines the credit, Strike defends the position at no additional cost to the client through the full administrative process. If any portion of the credit is disallowed on final assessment after exhaustion of administrative appeals, Strike refunds the corresponding fee. This is the skin-in-the-game commitment that AI-only providers structurally cannot offer, because their terms of service cap liability at the software subscription. The provider's incentive structure changes when the provider has economic exposure to the position taken on the return.

Strike has delivered more than $300 million in credits to over 1,100 clients across software, manufacturing, engineering, life sciences, and other technical industries. Engagements are success-based, which means no fee is owed unless a credit is delivered. Every position taken is anchored to IRS guidance, the Internal Revenue Code, and Treasury regulations, never to third-party interpretation.

Two Questions Worth Answering Before Your Next R&D Filing

If the article above has raised any concern about the durability of an existing credit position, there are two questions worth answering quickly.

1. Has your current advisor evaluated whether your company qualifies to amend 2022, 2023, and 2024 returns under Section 174A? For sub-$31 million companies, this is a refund opportunity independent of the credit itself, and the Section 6511 amendment deadline for tax year 2022 expires in mid-2026.
2. Does your current advisor's work product meet the Section G reporting standard required for tax years beginning in 2026? The deliverable that satisfies the new IRS standard looks very different from a questionnaire output, and reconstruction after the fact is not a reliable strategy.

Strike offers a no-cost second-opinion review for both questions. If your current credit position is well-documented and your Section 174A position is optimized, we will tell you. If there are gaps, we will quantify them.

Disclaimer: This article is for general informational purposes only and does not constitute tax, legal, or financial advice. The R&D tax credit involves entity-specific facts, controlled group analysis, state rules, and other variables that affect outcomes. Consult a qualified tax professional for guidance tailored to your situation.