Your sprint board is a research record. It has just never been read as one.
New architectures, novel algorithms, AI/ML systems, scale and security work. The credit was written for exactly this, and software teams are among the most likely to underclaim it. A study maps the engineering you already did to the four-part test.
Photo by Christina @ wocintechchat.com on Unsplash
Why software development qualifies
Each new or improved release is a business component - a product, software, or technique you’re trying to make better. The work qualifies when your team faces genuine technical uncertainty about architecture, algorithms, performance, scale, or security and resolves it through a systematic process of experimentation: designing alternatives, prototyping, and testing.
Every qualifying activity must pass the IRC §41 four-part test - permitted purpose, technological in nature (here, the principles of computer science), elimination of uncertainty, and a process of experimentation. Routine configuration, content changes, and bug fixes after release don’t qualify; the genuinely uncertain engineering does.
Software firms underclaim for predictable reasons: the work feels like “just building the product,” time isn’t tracked against the four-part test, and generalist preparers skip it. Mapping each activity to the statute - with contemporaneous evidence out of the systems you already run - is exactly what turns everyday engineering into a defensible claim. See documentation & substantiation for what that evidence looks like.
Seven things that happened on your roadmap last year
Not activity categories - situations. If any of these read like a sprint you actually ran, there is very likely a qualified business component underneath it.
-
“p99 latency would not come down.”
You suspected the query planner, then the cache, then the write path. Three indexing and caching strategies were benchmarked against production-shaped traffic before one held.
Why it can qualify: The information you had did not establish the right method. The benchmark runs are a documented process of evaluating alternatives.
-
“The model was accurate offline and useless in production.”
Feature pipelines, retraining cadence, and drift detection were designed and re-evaluated against held-out and live data until the deployed metric matched the offline one.
Why it can qualify: Modeling and generalization uncertainty resolved by systematic experimentation - not by reading the documentation.
-
“Tenant isolation had to survive a security review.”
You prototyped schema-per-tenant against row-level security, load-tested both, and measured the blast radius of a misconfigured policy before committing.
Why it can qualify: Two candidate designs, tested against requirements. Design uncertainty is one of the three flavors §41 recognizes.
-
“There was no known-good path off the monolith.”
Strangler-fig increments, dual writes, shadow traffic, and a rollback design that had to be proven before the first cutover.
Why it can qualify: Capability and design uncertainty about your own system. Re-architecture is a new or improved business component, not maintenance.
-
“Two systems disagreed and the connector had to decide.”
Idempotency keys, ordering guarantees, and conflict resolution were designed and replayed against real payloads until reconciliation stopped drifting.
Why it can qualify: Interface uncertainty resolved through iterative testing against actual data, not a configuration exercise.
-
“Usage tripled and infrastructure cost quadrupled.”
Storage tiering, query paths, and batching strategies were each measured under load; two were abandoned after the numbers came in.
Why it can qualify: Improving performance is a permitted purpose. The abandoned approaches are part of the experimentation, not wasted effort.
-
“You shipped it, then rewrote it six months later.”
The first implementation worked but could not be extended; the second was built on a different abstraction after both were prototyped.
Why it can qualify: Qualification turns on the process, not the outcome. Work on the version you replaced can still be qualified research.
Illustrative situations, not client work. Whether any of them qualifies for you depends on your facts, your contracts, and your evidence.
What the evidence looks like when it works
A defensible software claim is not assembled from memory at filing time. It is assembled from the record your team already produced: design documents that name the alternatives, benchmark runs with dates and numbers, tickets that say what was still unknown, and the branch that got abandoned.
We pull that record out of the systems you already run, tie each business component to the four-part test, and keep a line from every dollar back to the source row it came from.
The software work that commonly qualifies
Representative activities we see meet the four-part test. The label matters less than the underlying technical uncertainty and experimentation.
New features & architectures
Novel algorithms & data structures
AI/ML model development
Performance & scalability engineering
Security engineering
DevOps & infrastructure-as-code
Complex integrations
Prototypes & POCs
Platform re-architecture
Typical QRE categories for software
What spending counts toward the credit - tailored to how software teams actually spend.
| Expense category | What goes into the base |
|---|---|
| Technical wages§41(b)(2)(A)-(B) | W-2 wages for engineering, QA, DevOps, and technical product staff for time spent on qualified development, supervision, and direct support. |
| Cloud & compute§41(b)(2)(A)(iii) | Amounts paid to rent cloud and compute for development, test, staging, and model-training environments used in qualified research. |
| Contract development (65%)§41(b)(3) | 65% of amounts paid to U.S. third-party developers and agencies for qualified development performed on your behalf. |
| Supplies§41(b)(2)(C) | Limited for software teams - tangible property consumed in research, not depreciable equipment or overhead. |
What the base usually looks like
IllustrativeA directional shape for a software company, not a benchmark. Wages carry the claim; the categories below them are the ones most often left out entirely.
- Technical wages - Engineering, QA, DevOps, and technical product time on qualified work.
- 82%
- Cloud & compute - Rented dev, test, staging, and model-training capacity.
- 10%
- U.S. contract development - Agencies and contractors, in the base at 65% of amounts paid.
- 6%
- Supplies - Rarely material for a software team.
- 2%
Where the line sits
Production hosting for a released product is not research compute. Neither is the seat cost of the tools your team uses - software licenses and SaaS subscriptions are not a §41 expense category at all.
The defensible version of a cloud QRE is an allocation you can point at: tagged accounts, separate development and training projects, or environment-level billing. If nobody can tell which spend was research, an examiner will make the same observation.
A credit you can use before you owe income tax
§41(h) lets a qualified small business elect to apply up to $500,000 of its research credit per year against payroll taxes instead of income tax - the employer share of Social Security tax first, and the Medicare share above that. For a venture-funded software company with a large engineering payroll and no taxable income, that is the difference between a carryforward and cash.
The definition is where companies get caught: it turns on gross receipts under $5 million for the credit year and on not having had gross receipts before the five-year window ending in that year. A company that took its first dollar of revenue six years ago is out, however small it still is. The election is made on a timely-filed return, claimed on Form 8974 with your quarterly employment tax return, and cannot be made for more than five tax years.
Rough QSB screen
- Gross receipts under $5M in the credit year
- No gross receipts before the five-year window ending in that year
- A real U.S. payroll to offset
- Election made on a timely-filed return, not after the fact
- Five tax years is the maximum, ever
Summary only - the statutory definition and the aggregation rules decide it. We test them explicitly.
What a software study can look like
A hypothetical scenario to show how the pieces fit together. It is not a quote, projection, or promise of results.
- Engineering payroll
- $3.2M
- Share on qualified development
- ~60%
- Cloud & compute (dev/test/training)
- $300K
- Estimated QRE
- ~$2.2M
- Illustrative federal credit
- ≈ $130K-$220K
Plus the full §174A first-year deduction on domestic R&E, including software development.
Illustrative only. Figures are hypothetical and rounded; the federal credit commonly works out to roughly 6-10% of QRE depending on method, filing history, and the §280C election. Your result depends entirely on your facts. This is not a quote or a guarantee.
Software development is fully deductible again
IRC §174A restores immediate, full expensing of domestic research & experimental costs for tax years beginning after December 31, 2024 - and domestic software development is explicitly included. Captured alongside the §41 credit, you get the deduction and the credit, correctly.
SaaS & software - frequently asked questions
Does building features for our own product count?
Does AI/ML development qualify?
Do cloud and compute costs count?
Do offshore developers count?
A customer paid us to build it. Can we still claim it?
Does §174A really cover software development?
We’re pre-revenue - can we still benefit?
Can we claim prior years we already filed?
What does the §280C election change?
Next
The four-part test, applied the way an examiner applies it
Permitted purpose, technological in nature, uncertainty, experimentation - and the shrink-back rule when only part of a release qualifies.
Also relevant
If your team ships firmware or silicon alongside the service, the hardware page carries the supply and funding rules that page does not.