AI Development and the R&D Tax Credit: What Businesses Should Know

Businesses are investing in artificial intelligence to improve products, automate internal processes, analyze data, support customers, and build new software capabilities. Some of that work may qualify for the federal research credit under Section 41.

The credit is not available simply because a company uses artificial intelligence or pays for an AI platform. Eligibility depends on the technical work performed, the uncertainties the company attempted to resolve, the alternatives it evaluated, and the costs connected to those activities.

Companies that wait until tax-return preparation to evaluate their AI projects may find that the strongest evidence no longer exists. Project records should be created while the technical work is underway.

AI projects must satisfy the four-part test

An AI project does not need to produce a scientific breakthrough or an entirely new technology. It must satisfy the established four-part test for qualified research at the applicable business-component level.

The project generally must meet these requirements:

Permitted purpose: The work must seek to improve the function, performance, reliability, or quality of a product, process, software system, technique, formula, or invention.
Technological in nature: The experimentation must fundamentally rely on principles of computer science, engineering, physical science, or biological science.
Elimination of uncertainty: The company must face uncertainty about whether or how it can achieve the intended technical result or determine the appropriate design.
Process of experimentation: The team must evaluate alternatives through modeling, simulation, testing, prototyping, trial and error, or another systematic technical process.

Testing alone does not establish eligibility. Substantially all—generally at least 80%—of the research activities for the business component, measured on a cost or other consistently applied reasonable basis, must constitute elements of a process of experimentation. If a broader component does not satisfy the requirements, a smaller component or subset may still qualify under the applicable rules.

Potentially qualifying work may include developing proprietary models, fine-tuning an existing model, building retrieval or data-processing systems, integrating AI into a proprietary platform, and testing competing technical architectures.

Technical uncertainty may involve accuracy, hallucination rates, latency, scalability, model drift, data quality, security, reliability, or the ability to operate within defined computing constraints.

Developing AI is different from using AI

Purchasing an AI subscription, configuring a standard chatbot, or asking employees to use generative AI for routine business tasks generally will not establish qualified research. A stronger credit position may exist when a company’s developers and engineers must determine how to achieve a technical result that is uncertain at the beginning of the project.

For example, a SaaS company may test several retrieval architectures to determine whether its product can return accurate responses within an acceptable latency threshold. A business developing an internal fraud-detection tool may compare models, data pipelines, and validation methods because it is uncertain whether the system can meet required accuracy and reliability standards.

Companies must also distinguish experimental development from routine implementation, maintenance, and quality-control testing. Activities that merely adapt an existing product or process to a particular customer’s needs, duplicate an existing product or process, or occur after the relevant business component is ready for commercial use generally are excluded. However, later work to develop a new or improved feature or component may qualify if that work independently satisfies the research-credit requirements.

The project should be evaluated based on the technical process, not the use of the term “AI” in a proposal or budget.

Internal AI tools may face additional limitations

Some software developed primarily for a company’s internal general and administrative functions is subject to an additional high-threshold-of-innovation test. That test generally requires the company to establish that the software is innovative, involves significant economic risk, and is not commercially available for the intended use without modifications that themselves satisfy the innovation and significant-economic-risk requirements.

Software developed for sale, license, or use by customers may not be treated as internal-use software. Systems that allow customers or other third parties to initiate functions or access data may also receive different treatment.

A company building an AI feature into its commercial SaaS platform may therefore have a different credit analysis from a company customizing an AI system solely for finance, human resources, or internal administrative support.

Which AI development costs should be reviewed?

Qualified research expenses may include wages paid to employees who perform, directly supervise, or directly support qualified research.

Relevant employees may include:

Software engineers and machine-learning engineers
Data scientists and technical architects
Engineering managers directly supervising the work
Employees testing or directly supporting experimental development

Job titles alone do not determine eligibility. Companies should identify the activities each employee performed and develop a reasonable method for allocating time when an employee worked on both qualifying and nonqualifying tasks.

Certain contractor costs may also qualify. Generally, 65% of eligible contract research payments may be included in qualified research expenses for the federal credit. The agreement generally must be entered into before the research begins, provide for research on the company’s behalf, give the company rights to the research results, and require it to bear the expense even if the research is unsuccessful. Exclusive ownership of the results is not required.

Companies performing AI development for customers must separately evaluate whether their work is excluded as funded research. Customer payments do not automatically prevent a company from claiming a credit, but eligibility depends on the contract terms and circumstances, including whether payment is contingent on successful research and whether the company retains substantial rights to the results.

Cloud computing charges also warrant review. Certain payments for the right to use off-premises computers in qualified research may qualify under the computer rental rules, depending on the arrangement and use. Companies should distinguish computing charges attributable to experimental development from production hosting and other operating costs. AI subscriptions, model API fees, and data-storage charges do not automatically qualify simply because they support a development project.

The location of the research also matters. Research performed outside the United States, Puerto Rico, and other U.S. possessions generally is excluded from the federal credit. Companies using offshore developers or globally distributed teams should identify where the work was performed and allocate costs accordingly. Paying a U.S.-based contractor does not, by itself, establish that the underlying research occurred in an eligible location.

Supplies used in qualified research may also qualify, but purchased depreciable equipment, such as servers and GPUs, is excluded from qualified research supplies.

Documentation should begin during development

Current Form 6765 reporting places greater emphasis on identifying research by business component. For tax years beginning after 2025, Section G business-component reporting is generally required, subject to the form’s exceptions.

Useful project records may include:

Project charters describing the intended technical improvement
Records of technical uncertainties identified at the outset
Design documents and architecture alternatives
Sprint tickets, engineering notes, and version-control histories
Model evaluation criteria and test results
Records of failed approaches and design changes
Employee time allocations and reporting structures
Contractor agreements, invoices, and work descriptions

The documentation should explain what the company did and why experimentation was necessary. A collection of payroll reports and development tickets without a clear technical narrative may not support the full claim.

The research credit is separate from the R&E deduction

The Section 41 research credit and the deduction treatment of research expenditures are related but separate tax provisions.

For tax years beginning after 2024, Section 174A generally permits businesses to deduct domestic research and experimental expenditures currently. Foreign research expenditures generally remain subject to 15-year capitalization under Section 174.

A cost may be relevant to the deduction analysis without qualifying for the research credit. The credit applies a narrower set of activity and expense requirements. Companies claiming both should also address the Section 280C adjustment, which generally requires either reducing the related deduction or electing a reduced research credit.

Aura’s discussion of Section 174 capitalization relief provides additional background on the deduction rules.

Eligible small businesses may benefit before becoming profitable

Eligible qualified small businesses may elect to apply up to $500,000 of federal research credit annually against certain employer payroll taxes, even when they have no current federal income-tax liability. Eligibility generally requires less than $5 million in gross receipts for the credit year and no gross receipts before the five-tax-year period ending with that year, subject to aggregation and other rules. The election generally must be made on a timely filed original income-tax return, including extensions, and is available for no more than five tax years.

California requires a separate calculation

The California research credit is based on the federal credit with significant modifications. Qualified activities generally must be conducted in California, and California now permits an alternative simplified credit election for taxable years beginning on or after January 1, 2025.

The California credit is generally nonrefundable. For taxable years beginning in 2024 through 2026, the use of business credits is also generally limited to $5 million for the taxable year, with the limitation applied at the combined-reporting group level where applicable.

A company may generate a California research credit without receiving an immediate cash-tax benefit. Location of personnel, entity structure, projected California tax, credit carryovers, and current utilization limits should be modeled separately from the federal calculation.

How Aura Advisors can help with AI development research credit

AI projects should be evaluated while engineers, product leaders, and financial teams can still identify the technical uncertainties, experiments, and costs associated with the work.

Aura Advisors helps technology companies assess whether AI and software projects warrant a detailed research-credit review, organize project and payroll information, coordinate with engineering teams, and connect the tax calculation to Section 174A treatment and state reporting. We can also help companies establish documentation practices that fit their development process rather than attempting to reconstruct the record after year-end.

The objective is not to label every AI initiative as research. It is to identify the projects that meet the statutory tests, separate qualifying work from routine implementation, and build a supportable position from records created when the work occurred.

For More Reading:

Aura Advisors is a boutique tax consulting and compliance firm working with start-ups, emerging growth companies, and affluent individuals. Making it safer for good people and good companies to continue to do good things.