Businesses in this industry may conduct activities that potentially meet the qualified-research requirements. Industry alone does not determine eligibility. Explore credits with the R&D Credit Finder.

What R&D can look like

  • Designing new algorithms or data structures to meet performance targets.
  • Architecting distributed systems where consistency, latency or scale requirements are uncertain.
  • Developing novel integrations where interface behavior is undocumented or unpredictable.
  • Machine-learning model development with structured experimentation.

Research topics in software

Software development

Software development is research only to the extent the work seeks to resolve technical uncertainty through a process of experimentation grounded in computer science. Routine feature work built with known methods — even if it is new to the product — generally does not meet that standard.

Algorithm development

Designing algorithms where the achievable accuracy, speed or resource use is uncertain, and evaluating alternative approaches through modeling, benchmarking or prototyping, may involve qualified research.

Architecture, performance and scalability

Architecture decisions made to meet performance or scale targets that the team could not know how to reach at the outset — for example throughput, latency or concurrency limits — may involve experimentation when alternatives are designed and tested.

Security

Developing new security mechanisms or hardening techniques whose effectiveness is uncertain may qualify; applying established configurations, patches or standard libraries generally does not.

Integration

Integrating systems can involve technical uncertainty where interfaces, data models or timing constraints are not documented or conflict. Configuring commercial software using vendor instructions is generally not experimentation.

Internal-use, dual-function and non-IUS software

Software developed primarily for general and administrative functions is internal-use software (IUS) and must also satisfy the high threshold of innovation test. Software developed to be sold, leased or licensed, or to allow third parties to interact with the taxpayer's systems, is generally not IUS. Dual-function software serves both internal and third-party functions and has its own allocation rules. Form 6765 Section G asks for this classification.

Potential business components

The four-part test is applied to each business component. Examples in this industry:

  • A new or improved software product or module held for license
  • A recommendation, matching or ranking algorithm
  • A data pipeline or storage engine redesigned for scale
  • An embedded or device-control software component
  • A developer platform or API used by third parties

Common technical uncertainties

  • Whether an algorithm can meet accuracy or latency targets
  • Which architecture can sustain a required load or concurrency
  • How to achieve compatibility between systems with undocumented behavior
  • Whether a security mechanism resists defined attack methods

See technical uncertainty.

Potential experimentation

  • Prototyping alternative designs and benchmarking them
  • Load, stress and performance testing of architectural alternatives
  • Iterative model training and evaluation against defined metrics
  • Simulation of failure modes and recovery approaches

See process of experimentation.

The four-part test in software

TestApplication
Permitted purposeFunctional or performance improvements — not UI styling.
Technological in natureRelies on computer science principles.
UncertaintyCapability, method or design uncertainty documented before development (e.g., in design docs).
Process of experimentationPrototyping, benchmarking and iterative evaluation of alternatives.

Read the four-part test research

Project examples

Example · Software

API architecture experimentation

A software company needed a public API able to serve 10x its existing request volume with p99 latency under 100 ms. The team was uncertain whether an event-driven design or a synchronous gateway with caching would meet the target.

Business component: Public API platform (software held for license)

Potentially qualifying factors
  • Uncertainty about appropriate design to meet measurable performance targets
  • Alternatives (event-driven vs. cached synchronous) evaluated through load testing
  • Reliance on computer-science principles (concurrency, caching, queuing theory)
Potential exclusion issues
  • Routine endpoint development using established patterns
  • Work after the API was released to customers (post-commercial production)
  • Documentation and developer-portal content
Documentation required
  • Architecture decision records
  • Load-test plans and results
  • Contemporaneous records of the uncertainty at the outset
  • Records of alternatives evaluated and results
  • Time allocation for personnel involved

Illustrative example. Qualification depends on the specific facts of each business component and is not automatic.

Example · Software

Database optimization

Queries on a multi-tenant database degraded as data grew. The team did not know whether partitioning, alternative indexing or a columnar store would restore performance without breaking consistency guarantees.

Business component: Data layer of a commercial SaaS product

Potentially qualifying factors
  • Uncertainty about method and design
  • Benchmarking of alternative schemas and indexes
  • Iterative refinement based on measured results
Potential exclusion issues
  • Applying a vendor's documented tuning guide without uncertainty
  • Routine maintenance and monitoring
  • Data migration after the design was settled
Documentation required
  • Benchmark scripts and results
  • Tickets linking experiments to findings
  • Contemporaneous records of the uncertainty at the outset
  • Records of alternatives evaluated and results
  • Time allocation for personnel involved

Illustrative example. Qualification depends on the specific facts of each business component and is not automatic.

Example · Software

Distributed system architecture

A team designing a multi-region replication system faced uncertainty about achieving consistency and failover targets under network partitions. They modeled three consensus approaches and ran fault-injection tests.

Business component: Replication subsystem

Potentially qualifying factors
  • Modeling and simulation of alternatives
  • Fault-injection experiments with recorded outcomes
  • Design iterations driven by test results
Potential exclusion issues
  • Deploying the chosen design to production
  • Operations and on-call work
  • Configuration of managed cloud services without experimentation
Documentation required
  • Design documents comparing alternatives
  • Fault-injection test logs
  • Version-control history
  • Contemporaneous records of the uncertainty at the outset
  • Records of alternatives evaluated and results
  • Time allocation for personnel involved

Illustrative example. Qualification depends on the specific facts of each business component and is not automatic.

Example · Software

Performance optimization

A mobile application's rendering pipeline could not meet frame-rate targets on low-end devices. Engineers evaluated alternative rendering strategies with profiling.

Business component: Rendering engine

Potentially qualifying factors
  • Measurable performance objective
  • Systematic profiling and evaluation of alternatives
Potential exclusion issues
  • Cosmetic UI changes
  • Bug fixes after release unrelated to uncertainty
Documentation required
  • Profiling data
  • Benchmark comparisons
  • Contemporaneous records of the uncertainty at the outset
  • Records of alternatives evaluated and results
  • Time allocation for personnel involved

Illustrative example. Qualification depends on the specific facts of each business component and is not automatic.

Example · Software

Security architecture

A company designed a new key-management scheme for tenant data isolation and was uncertain whether it could meet latency constraints. It prototyped and tested two cryptographic designs.

Business component: Encryption and key-management component

Potentially qualifying factors
  • Technical uncertainty on design and performance
  • Prototype evaluation of alternatives
Potential exclusion issues
  • Compliance audits and policy writing
  • Routine penetration testing of a finished system
  • Implementing a standard library as documented
Documentation required
  • Threat models and design comparisons
  • Performance test data
  • Contemporaneous records of the uncertainty at the outset
  • Records of alternatives evaluated and results
  • Time allocation for personnel involved

Illustrative example. Qualification depends on the specific facts of each business component and is not automatic.

Example · Software

Machine-learning experimentation

A team sought to build a model to predict equipment failure from sensor data. It did not know which features or architectures would achieve the required precision, and ran structured experiments across candidates.

Business component: Predictive-maintenance model (software)

Potentially qualifying factors
  • Experiment tracking of model architectures and features
  • Evaluation against defined metrics
Potential exclusion issues
  • Routine data collection and labeling
  • Model retraining in production on a fixed pipeline
  • Market research on customer demand
Documentation required
  • Experiment-tracking logs
  • Evaluation reports
  • Cloud GPU usage allocated to experiments
  • Contemporaneous records of the uncertainty at the outset
  • Records of alternatives evaluated and results
  • Time allocation for personnel involved

Illustrative example. Qualification depends on the specific facts of each business component and is not automatic.

QRE examples

Wages

  • Developers and architects performing experimental development
  • First-line engineering managers directly supervising it

Computer use

  • Cloud compute for development and test environments

Contract research

  • U.S. contractors building prototypes where the company bears risk and retains rights

QRE rules: qualified research expenses.

Documentation

  • Architecture decision records
  • Tickets and pull requests linked to experiments
  • Benchmark and test results
  • Time tracking by project

Potential non-qualifying activities

The statutory exclusions in IRC §41(d)(4) — such as research after commercial production, adaptation or duplication of an existing component, routine quality-control testing, foreign research and funded research — apply to software projects in the same way as any other industry. The qualified research page lists each exclusion with its source.

Relevant federal rules: four-part test, qualified research, funded research, research location. Build a documentation checklist.

Federal R&D credit considerations

  • Apply the four-part test to each software business component.
  • Classify software for internal-use software rules and Form 6765 Section G.
  • Software development costs are research expenditures for §174 and §174A purposes; that treatment is separate from credit qualification.
  • Cloud costs may be QREs only where they are paid for the right to use computers in qualified research — see computer use.

Industry-specific caveats

  • Bug fixing, routine maintenance, UI styling and configuration of commercial software generally are not qualified research.
  • Work after a component is released for general availability may be post-production activity.
  • Development funded by a customer may be excluded as funded research depending on the contract terms.

Start with the federal R&D tax credit and qualified research expenses; estimate with the federal R&D credit calculator.

Common mistakes

  • Claiming routine feature work, bug fixes after release or configuration of third-party tools.
  • Overlooking internal-use software requirements for back-office systems.
  • Including production hosting costs as computer-use QREs.

State considerations

  • State credits require development to be performed in the state; remote developers create multi-state allocation questions.
  • For 2025+ federal purposes, domestic software development costs are deductible under §174A; state conformity varies.

Potentially relevant active state programs: California, Texas, Massachusetts, Utah.

Calculators: California R&D Tax Credit Calculator, Texas R&D Tax Credit Calculator, Massachusetts R&D Tax Credit Calculator, Utah R&D Tax Credit Calculator.

Frequently asked questions

Does all software development qualify?

No. Only development that meets the four-part test qualifies; internal-use software must also meet the high threshold of innovation test.

Official sources

General research and educational information, not legal or tax advice. Rules vary by tax year and taxpayer-specific facts affect results. See the disclaimer.