Defending a research credit on exam

The research credit is a statutory entitlement, not a negotiated position — but it is claimed on facts the taxpayer alone possesses, and on exam the burden of proving every element sits with the taxpayer. The IRS has treated the credit as a priority compliance area for two decades, and its current toolkit — refund-claim specificity requirements, the redesigned Form 6765 with business-component reporting, and a string of taxpayer losses on substantiation — has raised the practical bar well above where it stood ten years ago.

This guide describes what actually gets challenged, what the leading cases teach, and what a defensible file looks like. It offers no statistics on exam rates, because reliable ones by claim size and industry are not published; the honest framing is that a research credit of any size is a position the taxpayer should be prepared to prove, and preparation is almost entirely a claim-time activity.

The exam landscape

Research credit issues arrive through three doors. Large corporate claims are examined within LB&I, where the credit has repeatedly appeared in campaign and issue-focus form. Mid-market and small-business claims are examined in SB/SE, often triggered by a large amended-return refund claim. And refund claims themselves now face an administrative screen: a facially deficient claim can be rejected without examination at all.

A typical exam opens with information document requests seeking the credit computation workpapers, the list of business components, project documentation, the wage allocation methodology, and contracts for any contract research — in both directions: contracts under which the taxpayer paid others, and contracts under which others paid the taxpayer. That last request is the funded-research probe, and it surprises taxpayers who thought of themselves as doing internal product development while also performing customer-funded work.

What the IRS challenges

Qualification: the four-part test, applied component by component

Examiners rarely dispute that engineers were doing engineering. They dispute whether the activities claimed satisfy the four-part test at the business-component level — and in particular whether a process of experimentation occurred, and whether substantially all (80% or more) of each component's activities were elements of it. Projects framed as "we built a difficult thing" rather than "we identified these uncertainties, evaluated these alternatives, and here are the test artifacts" are the soft targets. The research-after-commercial-production line is a second front: examiners push development activity into the post-release period, converting it into excluded routine engineering.

Funded research

Any research performed under contract for a customer or grantor draws the Section 41(d)(4)(H) analysis: did the taxpayer retain substantial rights, and was payment contingent on success? Fixed-price contracts generally favor the performer on risk; time-and-materials and cost-plus contracts generally do not, and rights clauses assigning all IP to the customer can be independently fatal. The doctrine, and the Fairchild and Geosyntec case law, are treated in the funded research exclusion.

Wage allocations

Wages are the bulk of most claims, so allocation methodology is the bulk of most adjustments. The pressure points: employees claimed at 100% under the 80% substantially-all rule without support for the sub-20% remainder; executives claimed for "direct supervision" that is actually second-level management; support staff swept in wholesale; and allocation percentages produced years later through interviews alone, with no contemporaneous anchor. The categories and rules are detailed in qualified research expenses.

Substantiation of amounts

Even where qualification is conceded, examiners test the arithmetic: whether supply costs trace to research consumption rather than production, whether cloud costs are segmented between development and production hosting, and — under the ASC — whether base-year QREs were determined on a basis consistent with the credit year. An inflated credit year sitting on an unexamined, understated base is a classic adjustment.

What the cases teach

Little Sandy Coal: substantially all means proven, not assumed

In Little Sandy Coal Co. v. Commissioner (7th Cir. 2023, affirming the Tax Court), a shipbuilder claimed credits for first-of-a-kind vessels, arguing that because the vessels were novel, the work of building them was experimentation. Both courts disagreed. Designing and constructing something new is not, without more, a process of experimentation; the taxpayer had to show that at least 80% of the activities relating to each vessel — measured by cost or another reasonable basis — constituted elements of an evaluative, alternative-testing process, and it could not, because it had not tracked which activities were which. The Seventh Circuit was explicit that treating the entire production effort as experimentation because the end product was uncertain "proves too much."

The lessons: the substantially-all fraction must be demonstrable, activity by activity; novelty of the output is not evidence about the character of the inputs; and the shrink-back rule is a tool the taxpayer must actually use — claiming the qualifying subsystem is defensible where claiming the whole vessel is not.

Siemer Milling: effort without documented method is not experimentation

In Siemer Milling Co. v. Commissioner (T.C. Memo. 2019-37), a flour miller claimed credits for new flour products and process improvements. The Tax Court accepted that the company faced real technical questions but denied the credit because the record did not show a process of experimentation: no documents identifying hypotheses, alternatives evaluated, or a methodology for testing them — only evidence that trials occurred and adjustments were made. Trial and error in the colloquial sense is not the systematic trial and error the regulation describes.

The lesson is narrower and more actionable than Little Sandy Coal's: the taxpayer probably was experimenting, and lost anyway, because nothing written at the time said so. A one-page test plan per project — uncertainty, alternatives, evaluation criteria, results — is the cheapest audit insurance in this area.

Estimation and its limits

Cohan v. Commissioner permits courts to estimate deductible amounts when entitlement is established but records are imperfect, and research credit taxpayers have occasionally benefited (Fudim, and portions of Union Carbide). But estimation presupposes a reliable foundation. Where the qualification of the underlying activities is unproven, there is nothing to estimate — the modern cases decline to guess. Oral testimony from engineers remains admissible and useful, but as corroboration of records, not as a substitute.

The refund-claim specificity requirements

Since January 10, 2022 (per the IRS's October 2021 Chief Counsel memorandum and subsequent guidance, with transition perfection periods now expired), a claim for refund based on the research credit must include, at filing: identification of each business component; for each, all research activities performed; and the total qualified employee wage, supply, contract research, and computer rental expenses for the claim year. The IRS pared back an earlier demand to name individuals and the information each sought to discover, but the component-level core stands. A claim missing these elements can be rejected as deficient — no exam, no appeal rights on the merits, and if the limitations period has closed, no second chance.

The practical consequence: an amended-return credit is now a documentation project before filing, not after. The same component-level discipline appears prospectively on original returns through Form 6765 Section G, which for larger claimants asks for business-component detail annually. Taxpayers should assume the two datasets will be compared.

Building the contemporaneous file

A defensible file is assembled in the credit year, keyed to business components, and boring. The elements:

• Component inventory. A list of business components with a one-paragraph description of each: the uncertainty at the outset, the alternatives considered, how they were evaluated, and the outcome — including failures, which are affirmative evidence.
• Experimentation artifacts. The documents engineering already produces, preserved and indexed: design review minutes, test plans and results, simulation runs, sprint records and tickets referencing the technical problem, lab notebooks, prototype disposition records.
• Wage allocation basis. Time tracking where it exists; otherwise a written allocation methodology built during the year from project assignments, with the substantially-all rule applied only where the sub-20% residual can be described.
• Cost tracing. GL detail tying supplies and cloud costs to components; a written cloud-allocation method separating research environments from production.
• Contracts file. Every research-adjacent contract, inbound and outbound, with a memo on rights and risk for each.
• Computation consistency memo. Documentation that base-period QREs use the same definitions as the credit year — the requirement most often skipped and most easily tested.

An honest word on risk

The research credit is not an aggressive position; it is a congressionally intended, permanent incentive that the IRS itself acknowledges most claimants are entitled to. But it is also a self-scored credit in an area where the taxpayer holds all the facts, and the enforcement architecture is built accordingly. Two postures are defensible: claim it with component-level records made in real time, or do not claim it. The middle path — a round-number credit assembled from memory when the return is on extension — is the one the case law was written about.