Ricerca
Get started
IRC §41(d)

The four-part test, applied the way an examiner applies it

§41(d) admits an activity to the credit only if it clears all four tests - and it applies them to each business component, not to your R&D budget as a whole. Here is what each test actually asks, what clears it in software, manufacturing, and life sciences, and what quietly fails.

Get your free credit estimate See what spending counts
Four quadrants of one base, each cleared - the §41(d) four-part test

The test has a unit, and it isn’t “the project”

The statute applies all four tests to a business component: any product, process, computer software, technique, formula, or invention that you hold for sale, lease, or license, or use in your own trade or business (§41(d)(2)(B)). That definition does more work than people expect.

“Process” is in the list. A manufacturing process you never sell is as eligible as the part it produces. §41(d)(2)(C) goes further: a plant process, machinery, or technique for the commercial production of a business component is treated as a separate business component from the thing being produced - which is why so much qualifying work in manufacturing sits on the plant floor rather than in an engineering department, and why a mature product can still host a brand-new qualifying process.

Your internal structure is irrelevant. Cost centers, epics, and grant numbers are not business components. The pump housing, the ingestion service, and the purification step are. And because §41(d)(2)(A) applies the tests separately to each business component, one strong component does not carry a weak one, and one failing component does not sink the rest of the claim.

This is also why the IRS revised Form 6765 to collect information by business component in its Section G - the reporting now follows the statute’s unit of analysis, which is exactly how a study should have been built all along. See documentation & substantiation for what that means on the return.

Part by part

What each test asks - and what quietly fails it

Statutory anchors, plain-English explanation, concrete examples from three industries, and the fact patterns that don’t survive review.

1

Permitted purpose

IRC §41(d)(1)(B)(ii) IRC §41(d)(3)

Was the work undertaken to create a new or improved business component of yours?

“Improved” has a closed definition: improved function, performance, reliability, or quality (§41(d)(3)(A)). The improvement is measured against your prior component, not against the state of the art - you do not have to invent something new to your industry, or even new to your competitors.

Note what is absent from that list: cost. A cost-reduction program qualifies when the cost comes out because you improved the performance or reliability of a process - which is usually how it happens - but “we wanted it cheaper” is not by itself a permitted purpose. §41(d)(3)(B) draws the other boundary: research relating to style, taste, cosmetic, or seasonal design factors never has one.

What fails it

  • A visual refresh, a new palette, or packaging that changes only appearance - §41(d)(3)(B) names this case.
  • Evaluating two off-the-shelf vendors and configuring the winner to spec.
  • Documentation, training material, or collateral for a component that is already built.

What clears it

SaaS & software

Re-architecting a batch ingestion service into a streaming pipeline so p99 latency holds at ten times the volume. The component is the software you license; the purpose is performance.

Industry guide Manufacturing

Changing joint geometry and fixture design so a stamped assembly holds tolerance at a higher line rate. The component is the process - reliability and performance both.

Industry guide Pharma & biotech

Developing a purification step that holds yield and purity moving from a 2 L bench run to a 200 L batch. The component is the process; the purpose is quality.

Industry guide
2

Technological in nature

IRC §41(d)(1)(B)(i) Treas. Reg. §1.41-4(a)(4)

Does the information you were seeking fundamentally rely on the hard sciences?

The regulation lists the eligible disciplines: the physical or biological sciences, engineering, or computer science. “Fundamentally rely” is the operative phrase - the question is which discipline the problem belongs to, not what your company sells or what your team is called. A payments company solving a distributed-consistency problem is doing computer science.

The bar is “relies on,” not “advances.” You are not required to push the frontier of the discipline; well-understood science applied to a problem you have not solved still clears this test. What does not clear it is a problem whose answer comes from a discipline the statute left out.

What fails it

  • Problems that belong to economics, psychology, or market behavior - demand models, churn studies, org design. §41(d)(4)(G) excludes the social sciences, arts, and humanities outright.
  • Preference and aesthetic work: which layout users like, which shade converts.
  • Regulatory or legal analysis, however technical its subject matter.

What clears it

SaaS & software

Consensus behavior under network partition, query-planner cost models, numerical stability in a training pipeline - computer science, squarely.

Industry guide Manufacturing

How a new alloy behaves in the die, how a coolant path removes heat, how a weld’s heat-affected zone changes fatigue life - mechanical, materials, and thermal engineering.

Industry guide Pharma & biotech

Binding kinetics, formulation stability, assay sensitivity and specificity - biology and chemistry doing the work.

Industry guide
3

Elimination of uncertainty

IRC §41(d)(1)(A) Treas. Reg. §1.41-4(a)(3)

At the outset, did you already know how you were going to get there?

Uncertainty exists when the information available to you at the start does not establish one of three things: the capability of developing or improving the component, the method of doing so, or the appropriate design. The third is the one most studies underclaim. Knowing that something can be built, and even how in broad terms, still leaves real design uncertainty - which architecture, which alloy, which excipient, at which parameters - and design uncertainty counts.

Two things this test does not require. You do not have to succeed: abandoned branches, failed candidates, and dead ends are squarely inside the credit. And you do not have to seek information that exceeds, expands, or refines the common knowledge of skilled professionals in the field - the 2003 final regulations say so at §1.41-4(a)(3)(ii), which is where the old “discovery test” went.

§41(d)(1)(A), as amended by the OBBBA, ties the activity to the §174A world: the expenditures must be treated as domestic research or experimental expenditures under §174A. Domestic R&E is immediately deductible again under §174A for tax years beginning after 2024 - the deduction and the credit are separate benefits computed off overlapping, but not identical, cost bases.

What fails it

  • Applying a known solution with established techniques - installing and configuring a vendor module per the implementation guide.
  • Scaling by adding more of what already works.
  • Uncertainty that is commercial rather than technical: whether it will sell, what it will cost, whether the date is achievable, which vendor to hire.

What clears it

SaaS & software

The throughput target is reachable in theory; which partitioning and consistency scheme holds it under multi-tenant load is unknown. Design uncertainty.

Industry guide Manufacturing

Whether the geometry can be produced to tolerance on the presses you already own (capability), and which of three tooling approaches gets there (method).

Industry guide Pharma & biotech

Whether an excipient system holds the drug substance stable at 2-8 °C through the shelf-life study, and what the appropriate formulation is.

Industry guide
4

Process of experimentation

IRC §41(d)(1)(C) Treas. Reg. §1.41-4(a)(5)-(6)

Did substantially all of the activity consist of systematically evaluating alternatives?

The regulation gives the process three elements, and an examiner looks for all three: you identify the uncertainty, you identify one or more alternatives intended to eliminate it, and you identify and conduct a process of evaluating those alternatives - modeling, simulation, or systematic trial and error.

“Substantially all” means 80% or more of the business component’s research activities, measured on a cost or other consistently applied reasonable basis (§1.41-4(a)(6)). The remaining fifth can still produce qualified expenses if it independently satisfies the other tests. Do not confuse this 80% with the other 80% in §41 - the wage gross-up for an individual whose services are substantially all qualified lives at Treas. Reg. §1.41-2(d)(2) and is explained on the qualified research expenses page.

This test now carries most of the examination risk. In the Seventh Circuit’s 2023 Little Sandy Coal decision the court required the taxpayer to show that substantially all of the activities constituted a process of experimentation; generalized testimony about how engineers work did not get there. Alternatives considered and rejected - with the evaluation still attached to them - are what makes this test survivable.

What fails it

  • One approach, implemented, then debugged. Debugging is not the evaluation of alternatives.
  • Validation testing of a design that was already settled - that is confirmation, not experimentation.
  • Routine or ordinary testing for quality control, which §41(d)(4)(D) excludes by name.

What clears it

SaaS & software

Two candidate indexing strategies benchmarked against a recorded workload, with the losing branch and its numbers still in the pull request.

Industry guide Manufacturing

A designed experiment across feed rate, temperature, and dwell, with first-article measurements and scrap counted by run.

Industry guide Pharma & biotech

Dose-ranging and stability arms, assay qualification, and the protocol deviations that forced re-runs - recorded as they happened.

Industry guide

The patent safe harbor. Treas. Reg. §1.41-4(a)(3)(iii) treats the issuance of a U.S. patent - other than a design patent - as conclusive evidence that the taxpayer discovered information that is technological in nature and intended to eliminate uncertainty. It settles two of the four tests for that component. It does not settle the other two, and it is not a prerequisite: the overwhelming majority of qualifying work is never patented.

Shrinking back

When a component fails, you shrink it - you never grow it

Treas. Reg. §1.41-4(b)(2) contains the rule that decides most real studies. If a business component as a whole does not clear the four tests, you do not stop. You apply the tests to the most significant subset of elements of that component. If the subset fails, you shrink again, and you keep going until a subset qualifies or you reach the smallest divisible unit. The rule runs in one direction only: you may never aggregate a failing component into a larger one to make it pass.

In practice this is the difference between a defensible study and an all-or-nothing one. A fourteen-module release where three modules carried genuine uncertainty is not a failed claim - it is a claim on three modules. A production line where only the vision-inspection station was developmental is a claim on that station. The same logic rescues components caught by an activity-level exclusion: a largely domestic project with a sliver of offshore QA is not zeroed by §41(d)(4)(F), it is shrunk back to the domestic elements.

Two disciplines make shrink-back hold up under examination. The qualifying subset has to be identified as a subset, with a written basis for why those elements and not the others. And the resulting share has to be applied to that component’s expenses exactly once - our platform records the subset determination and applies its share at a single point in the computation, because a haircut applied twice states a claim nobody determined.

Shrink-back, one release at a time

Illustrative

  1. 1 The release as a whole - fourteen modules, most of them configuration and UI work. Substantially all of the activity is not experimentation. Fails.
  2. 2 The most significant subset - the scheduling engine, the sync-conflict resolver, and the query planner. Each had a stated uncertainty and evaluated alternatives.
  3. 3 The claim - three components, tested individually, with the wages, supplies, and contract dollars traced to each and the basis for the subset written down.

Illustrative example only. Whether a subset is “most significant,” and what it captures, depends entirely on your facts.

IRC §41(d)(4)

Eight activities the statute takes off the table

Clearing the four-part test is necessary, not sufficient. These eight exclusions apply however well an activity scores.

§41(d)(4)(A)

Research after commercial production

Activity after the component is ready for commercial sale or use. Post-release bug fixes and routine maintenance are out; a genuinely new development effort on the next version is a new component, not a continuation of this one.

§41(d)(4)(B)

Adaptation

Adapting an existing business component to a particular customer’s requirement or need. Configuration and customization work is the classic trap for agencies and integrators.

§41(d)(4)(C)

Duplication

Reproducing an existing component from a physical examination of it, or from plans, blueprints, or other published information. Reverse engineering to reach parity is excluded; going past parity may not be.

§41(d)(4)(D)

Surveys, studies, and routine work

Efficiency surveys, management function or technique, market research and testing, routine data collection, and routine or ordinary testing or inspection for quality control.

§41(d)(4)(E)

Internal-use software

Software developed primarily for your own general and administrative functions must additionally clear the high threshold of innovation test at Treas. Reg. §1.41-4(c)(6) - innovative, significant economic risk, and not commercially available. Software you sell, lease, license, or that customers interact with is generally outside the exclusion, and dual-function software has its own rules.

§41(d)(4)(F)

Foreign research

Research conducted outside the United States, Puerto Rico, or a U.S. possession. This turns on where the work is performed - not on who performs it, and not on who pays.

§41(d)(4)(G)

Social sciences, arts, humanities

Excluded however rigorous the method. Economics, behavioral research, and design-preference work land here.

§41(d)(4)(H)

Funded research

Research funded by grant, contract, or otherwise by another person or a governmental entity. It turns on two questions - do you bear the economic risk, and do you retain substantial rights - and it decides whole claims.

Not all eight behave the same way. Several are activity screens rather than component killers - a project with some routine QC, some offshore work, or some post-production activity keeps the elements that qualify, through the shrinking-back rule above. Adaptation, duplication, and the internal-use-software classification describe the component itself, so an affirmative answer there removes it. Which category each exclusion removes from your dollars - wages, supplies, contract research, or cloud - is set out on the qualified research expenses page, where funded research gets the rights-and-risk treatment it deserves.

Passing the test is step one. Then it has to be provable.

A four-part-test conclusion is only worth what supports it. For every business component a study should be able to state, in writing and from contemporaneous sources, what the uncertainty was, which alternatives were evaluated, how they were evaluated, and which people, supplies, contracts, and compute went into that evaluation. That is the same record Form 6765 Section G now asks about, and the same record an examiner opens first.

From there the arithmetic takes over: qualified activity determines qualified research expenses, QREs feed the regular and ASC calculations, and - for a qualified small business - the result can be taken as a payroll tax offset instead of an income-tax credit. Start at the R&D credit overview if you want the whole picture in one page.

If the IRS examines a Ricerca study, we prepare the response, assemble the documentation, and defend the technical determinations behind our work alongside your tax preparer; representation before the IRS, where it is needed, is performed by credentialed practitioners. That is what Audit Protection covers, and it is included with every study.

FAQ

Four-part test questions we get every week

Do we have to invent something new to our industry?
No. The improvement is measured against your own prior business component, and Treas. Reg. §1.41-4(a)(3)(ii) says explicitly that you need not seek information exceeding the common knowledge of skilled professionals in the field. Novel to you, not novel to the world.
Does a failed project still qualify?
Yes - qualification turns on the process, not the outcome. An abandoned architecture, a formulation that never stabilized, or a tooling approach that never held tolerance can all produce qualified expenses, provided the four tests were met while the work was underway.
What exactly is a “business component”?
Any product, process, computer software, technique, formula, or invention that you hold for sale, lease, or license, or use in your own trade or business (§41(d)(2)(B)). A manufacturing process you never sell is a business component. A cost center, a Jira epic, and a grant number are not.
How much of a project has to be experimental?
For a business component to qualify, substantially all - 80% or more, on a cost or other consistently applied reasonable basis - of its research activities must be elements of a process of experimentation. If the component as a whole cannot meet that, the shrinking-back rule applies the tests to its most significant qualifying subset instead.
Our engineers don’t write much down. Can we still claim?
Often, yes - but the evidence usually already exists in systems nobody thinks of as documentation: tickets, commits and pull requests, test results, DOE runs, travelers, batch records, and lab notebooks. What matters is that the record is contemporaneous. See documentation and substantiation for what examiners actually ask for.
Is software used only inside our company eligible?
Possibly, but under a stricter standard. Software developed primarily for internal general and administrative use must clear the high threshold of innovation test at Treas. Reg. §1.41-4(c)(6) in addition to the four-part test. Software you sell, license, or expose to customers is generally not internal-use software at all.
Did §174A change the four-part test?
No. §41 and its four-part test are unchanged. What changed is the deduction: new §174A restores immediate expensing of domestic research and experimental costs for tax years beginning after 2024. The two interact through §41(d)(1)(A) and the §280C election, not through the qualification tests.
Who decides whether an activity passes?
You do, on your return - which is why the determination has to be documented at the time it is made rather than reconstructed later. A study is the record of that determination: each business component tested, each conclusion supported, and each dollar traced to it.

Let’s test your components, not your patience

Send us what your engineers worked on last year. We’ll tell you which business components look qualified under §41(d), what a study would likely capture, and what it costs - before you commit to anything.

[email protected] We typically reply within one business day.
Get your free credit estimate Talk to our team