In 2025, AI startups continue to benefit from the IRS Research Credit under Internal Revenue Code Section 41, designed to incentivize innovation and research activities.
This credit, often called the R&D Tax Credit, allows eligible businesses to reduce their tax liability by claiming a percentage of qualified research expenses (QREs) that exceed a base amount.
For AI startups, which typically involve software development, algorithm testing, and technological experimentation, this credit is particularly valuable as it can offset the high costs of developing new AI models, machine learning systems, and data processing techniques. The IRS emphasizes that the credit applies to activities aimed at resolving technological uncertainties, making it ideal for startups pushing the boundaries of computer science and AI.
The Research Credit has been a staple incentive since its introduction, with ongoing updates to ensure it supports emerging technologies like AI. In 2025, there are no major legislative changes from the prior year, but the IRS has refined guidance through audit techniques and publications to better address software and AI-specific research.
Eligible startups can claim the credit on Form 6765, and qualified small businesses (QSBs) may elect to apply it against payroll taxes, providing crucial cash flow relief for early-stage companies. This blog explores eligibility, calculation methods, application to AI, and compliance tips, drawing from official IRS sources with at least 7 active links.
We'll include 3 examples from IRS to illustrate real-world applications.
To qualify for the Research Credit, activities must meet the four-part test outlined in IRC Section 41. First, the research must be technological in nature, relying on principles of computer science, engineering, or physical sciences - perfect for AI, which involves algorithmic design and data analysis. Second, it must intend to develop a new or improved business component, such as software, processes, or techniques.
For AI startups, this could include creating novel neural networks or improving model accuracy.
Third, the activities must involve a process of experimentation to eliminate uncertainty, such as testing alternative hypotheses or models. This is core to AI development, where startups often iterate on algorithms to resolve performance issues.
Fourth, the research must have a permitted purpose, like enhancing function, performance, reliability, or quality, excluding aesthetic or social science research. Exclusions include research after commercial production, adaptation of existing components, or activities funded by others.
Startups benefit from special rules: if a company has no base period (e.g., new AI firms post-1983), the fixed-base percentage starts at 3% for the first five years, transitioning gradually. QSBs (gross receipts <$5 million, no receipts before 2021) can apply up to $500,000 (in 2025) against payroll taxes via Form 6765. The IRS audit guidelines stress substantiating claims with records of uncertainties and experiments, crucial for AI, where development is iterative.
The Audit Techniques Guide for Research Credit clarifies that software development, including AI, qualifies if it meets the tests, with emphasis on high-risk experimentation.
The regular credit is 20% of QREs exceeding the base amount, calculated as the fixed-base percentage times average gross receipts over the prior four years (minimum 50% of current QREs).
QREs include wages for qualified services (e.g., AI engineers testing models), supplies (e.g., hardware for data training), and 65% of contract research (75% for consortia). For startups, the alternative simplified credit (ASC) is 14% of QREs over 50% of the prior three-year average, easier for fluctuating expenses.
In 2025, the base amount uses updated gross receipts definitions, excluding returns and allowances. The Research Credit page notes that QSBs can offset up to $500,000 against payroll taxes, doubled for certain startups. Aggregation rules treat controlled groups as a single taxpayer, important for AI startups with affiliates.
The Audit Guidelines on Process of Experimentation for Software provide examples of qualifying AI-related activities, like evaluating multiple software alternatives.
AI startups often qualify due to the technological nature of their work. Uncertainties in model design, data integration, or algorithm efficiency meet the experimentation test.
For instance, testing different neural network architectures to improve accuracy qualifies as qualified research. However, routine coding or using pre-existing AI tools without experimentation does not.
The IRS distinguishes risk levels in software research: high-risk for fundamental innovations, moderate for improvements, low for adaptations. AI falls into moderate to high-risk categories. Basic research payments to universities for AI studies also qualify if without commercial objectives.
In 2025, with AI's growth, the IRS monitors cloud computing costs as QREs if used for experimentation. Startups should document alternatives evaluated to substantiate claims, as per Publication 5316, which lists IRC 41 as a key credit.
Maryna Pirska Tax Preparer, Manager
Explore Real Strategies, Trends, and Tips to Help Your Brand Grow.
