The R&D Tax Credit is back! Find out how the Big Beautiful Bill Act can fund your R&D. Learn More
%201.svg)
.svg)
.svg)
.svg)
Last Updated: October 2026
AI development rarely follows a straight line. Companies test models, adjust parameters, rebuild data pipelines, compare retrieval methods, measure accuracy, reduce latency, and sometimes abandon an approach entirely when it fails to meet the required performance.
That kind of iteration is one reason AI development can be particularly relevant to the R&D tax credit. But the credit is not available simply because a company builds or uses artificial intelligence. The underlying work still needs to meet the qualified research requirements under Section 41.
For AI companies, the real question is not “Are we building AI?” It is “What technical uncertainties were we trying to solve, and what experiments did we run to solve them?”
This guide explains how the R&D tax credit applies across the AI development lifecycle, from data and model selection through training, evaluation, infrastructure, deployment, and optimization.
Yes. AI development can qualify for the federal R&D tax credit when the underlying activities satisfy the requirements of Section 41.
There is no separate federal "AI tax credit." AI development is evaluated under the same qualified-research framework that applies to software and other forms of technical development. The IRS generally requires qualified research to satisfy the Four-Part Test separately for each business component.
Potentially qualifying work can involve developing or improving:
The technology itself does not establish qualification. A company can use highly sophisticated AI without performing qualified research. Another company can perform qualified research while improving a relatively narrow machine learning system.
The analysis starts with what the technical department was trying to solve and how it tried to solve it.
AI research generally needs to satisfy the same Four-Part Test as other research claimed under Section 41. The test considers the purpose of the research, its technological basis, the uncertainty involved, and whether substantially all of the research activities constitute elements of a process of experimentation.
The research should be intended to develop a new or improved business component related to its function, performance, reliability, or quality.
For an AI system, that might involve improving:
The research must fundamentally rely on principles of the physical or biological sciences, engineering, or computer science.
AI development frequently relies on computer science, software engineering, data engineering, and related technical principles.
Simply using an AI product in a business process does not establish technological research.
The development department should face uncertainty about its capability or method for developing or improving the business component, or about the appropriate design.
For example: Can the system achieve the required retrieval accuracy while keeping response latency within the product's performance threshold?
That presents a technical question.
Compare it with: Which AI-generated headline do customers prefer?
The second may be important to the business, but it is primarily a marketing or preference question rather than a technical uncertainty under Section 41.
The Company should evaluate alternatives in an effort to resolve the technical uncertainty. For AI development, that could involve comparing:
A failed experiment can still be relevant. The R&D credit does not require every research project to produce a successful commercial result.
Qualified research can occur throughout the AI development lifecycle. A useful way to think about that lifecycle is:
Data → Model or architecture selection → Training or configuration → Evaluation → Iteration → Infrastructure → Deployment and optimization
Not every activity within those stages qualifies. Technical experimentation can, however, occur at multiple points.
Developing data pipelines can involve R&D when engineers face uncertainty about how to process, transform, structure, retrieve, or otherwise handle data to achieve a required technical result.
Routine data cleaning using established methods presents a different case.
Companies may test different models or architectures because they do not know which approach can satisfy their technical requirements.
This can become particularly relevant when balancing competing requirements such as accuracy, latency, compute requirements, context size, or reliability.
Research can occur while experimenting with training techniques, fine-tuning approaches, hyperparameters, datasets, or other technical variables.
AI development often requires companies to determine whether a system performs as intended.
Developing an evaluation framework can itself involve technical uncertainty when existing metrics do not adequately measure the required technical performance.
Results from one experiment often create the next technical question.
A company might improve accuracy but increase latency. It may then need to experiment with model configurations, architecture, or inference methods to achieve both requirements.
AI systems can create technical challenges involving training infrastructure, data processing, compute utilization, scalability, reliability, and deployment.
Deployment does not automatically mean R&D has ended. Qualified experimentation can continue when engineers encounter unresolved technical questions involving performance, reliability, scalability, inference, or another qualifying purpose.
Potentially qualifying AI research generally involves technical uncertainty and a systematic effort to evaluate alternatives.
AI activity
Example of potential R&D
Model development
Testing alternative architectures to achieve required technical performance
Fine-tuning
Comparing techniques or configurations to improve model performance
RAG systems
Testing retrieval, chunking, ranking, or embedding approaches
Model evaluation
Developing and testing methods for measuring technical performance
Computer vision
Comparing models or preprocessing methods to improve detection accuracy
NLP
Testing alternative methods for classification, extraction, or generation
Recommendation systems
Experimenting with algorithms to improve prediction performance
Data pipelines
Developing processing methods for technically challenging datasets
Inference optimization
Testing approaches to reduce latency or computing requirements
AI infrastructure
Experimenting with architectures to improve scalability or reliability
These examples are potentially qualifying. They are not automatically eligible.
A company still needs to establish the Section 41 requirements for the particular business component.
Fine-tuning can potentially qualify when a development department faces technical uncertainty and evaluates alternatives to achieve a technical result.
For example, a company may not know which combination of the following activities will achieve its required performance.
The company might conduct multiple experiments, compare the results against technical benchmarks, modify its approach, and repeat the process.
That presents a different R&D fact pattern from following a standard fine-tuning procedure where the method and expected outcome are already established.
Fine-tuning is not what makes the activity qualify. The underlying technical experimentation is what matters.
Prompt engineering should not automatically be classified as qualified research.
There is a meaningful difference between routine prompt creation and technical experimentation.
For example, asking employees to try different wording until an AI assistant sounds friendlier may be closer to content optimization than technological research.
A different situation might involve engineers systematically evaluating prompt architectures as part of a broader system where they are trying to resolve uncertainty involving structured outputs, reliability, tool use, or another technical performance requirement.
Even then, the overall activity still needs to satisfy the Four-Part Test.
Companies should therefore avoid treating all prompt engineering as R&D simply because generative AI is involved.
Retrieval-augmented generation, or RAG, development can potentially qualify when engineers are experimenting to resolve technical uncertainty about how a system retrieves and uses information.
A company might evaluate:
For example, imagine a company needs its AI assistant to retrieve information from thousands of technical documents while meeting defined accuracy and latency requirements.
The company does not know which retrieval architecture will work. It builds several approaches, runs evaluations, compares retrieval quality and response times, and changes the architecture based on the results.
That presents a different fact pattern from connecting an application to an existing vector database using a documented implementation.
Potentially. A development department may face technical uncertainty about how to improve the factual reliability of an AI system while maintaining other performance requirements.
It might experiment with:
The relevant tax question is not whether the team calls the problem "hallucination."
The question is whether the work involves a qualifying technical uncertainty and experimentation directed toward improving the function, performance, reliability, or quality of the business component.
AI evaluation can potentially form part of qualified research when it is used to evaluate technical alternatives during development.
AI departments may need to measure several dimensions simultaneously, including:
A company might develop an evaluation framework because existing metrics do not adequately measure whether its system satisfies a particular technical requirement.
It could then use that framework to compare models, retrieval strategies, architectures, or other alternatives.
Routine monitoring of a production model after technical uncertainty has been resolved presents a different case.
Developing AI agents can potentially qualify when engineers need to resolve technical uncertainty involving how a system plans, selects tools, maintains state, processes information, or reliably completes tasks.
Potential experimentation could involve:
Simply connecting an existing agent framework to a set of APIs using documented methods does not automatically establish qualified research.
As with other AI projects, the technical experimentation matters more than the label attached to the product.
Proprietary algorithm development can potentially qualify when the research satisfies Section 41.
For example, a company may be uncertain about how to design an algorithm capable of achieving a required prediction, classification, optimization, or processing result.
The company might:
That process can provide evidence of technical experimentation. The company does not need a patent or an algorithm that is new to the entire world for the activity to potentially qualify. The analysis focuses on the taxpayer's research and whether the Section 41 requirements are met.
Computer vision projects can potentially involve qualified research when developers experiment to solve technical problems involving detection, classification, tracking, recognition, or image processing.
For example, a company may test:
The company should be able to identify the technical performance it was trying to achieve, the uncertainty it encountered, and the alternatives it evaluated.
Routine implementation of an existing computer vision service presents a different case.
Potentially. AI systems often depend on pipelines that ingest, transform, label, store, retrieve, and process large amounts of data.
Pipeline development may involve qualified research when engineers face technical uncertainty involving:
Routine ETL configuration using established tools and known methods generally presents a weaker case. The question is not whether a pipeline is technically sophisticated. It is whether the company was required to experiment to resolve a qualifying technical uncertainty.
AI infrastructure work can potentially qualify when companies experiment to overcome technical limitations involving training, inference, performance, reliability, or scalability.
Examples might include experimentation with:
For example, a company may have a model that meets its accuracy requirement but cannot meet its latency requirement at production scale. Engineers might test quantization methods, model variants, serving architectures, caching approaches, and infrastructure configurations.
That technical experimentation can be relevant to the R&D credit analysis.
Simply using an existing AI model does not automatically constitute qualified research.
A business can use commercially available AI for writing, customer service, coding assistance, marketing, research, data analysis, or internal productivity without performing qualified R&D.
Even incorporating an existing AI API into a software product does not automatically qualify.
The relevant question is what the taxpayer itself is developing or improving and whether that development satisfies Section 41.
A straightforward API implementation using a known method presents a different fact pattern from a company testing several models, retrieval architectures, evaluation methods, and inference strategies because the appropriate technical design is uncertain.
This distinction is increasingly important as AI becomes a standard component of ordinary software development.
AI-related work is less likely to qualify when it does not involve the technical research required under Section 41.
Examples can include:
Section 41 also excludes research after commercial production, certain adaptations, duplication, surveys or studies, research conducted outside the United States or a U.S. territory, and funded research under the applicable rules.
The distinction is simple but important:
AI adoption is not the same thing as AI research.
It can. AI software developed primarily for a company's internal use may be subject to the additional rules for internal-use software.
Under the current Form 6765 instructions, internal-use software generally needs to satisfy the ordinary qualified-research requirements and the High Threshold of Innovation, or HTI, test.
The HTI test considers whether:
For example, an AI product developed for customers can present a different software-classification issue from an AI system developed primarily to support a company's internal finance or human resources function.
The IRS rules also distinguish internal-use software, dual-function software, software excepted from internal-use treatment, and non-internal-use software.
Once qualified AI research has been identified, the company needs to determine the associated Qualified Research Expenses, or QREs.
For AI companies, potentially relevant categories include employee wages, certain supplies, qualifying computer rental or lease costs, and contract research expenses. The expense still needs to meet the applicable Section 41 rules.
Wages can qualify when employees perform qualified services, including conducting qualified research or directly supervising or supporting it. Depending on what they actually do, potentially relevant roles could include:
Job title alone does not determine qualification. A machine learning engineer who spends part of the year conducting qualified experimentation and the rest on routine development should not automatically have 100% of their wages included simply because "machine learning" appears in the job title.
Certain amounts paid to outside parties to perform qualified research can potentially qualify under the contract-research rules. Companies should preserve contracts, statements of work, invoices, and records showing what research the contractor actually performed.
Other QRE categories may also be relevant depending on the research and the nature of the expense.
Some computing costs can potentially qualify, but companies should not automatically treat the entire cloud or GPU bill as a QRE.
The current Form 6765 instructions provide a QRE category for amounts paid or incurred for the rental or lease of computers used in qualified research.
The instructions impose specific requirements, including that the computers be located off the taxpayer's premises and that the taxpayer not be the operator or primary user of the computers.
For AI businesses, that makes the facts surrounding the computing arrangement important. It can also be useful to distinguish computing associated with the following activities from ordinary production computing and general business use.
Detailed cloud billing, environment tags, project records, and experiment tracking can make the analysis easier.
Research performed outside the United States or a U.S. territory is excluded from qualified research for the federal Section 41 credit.
That can be especially important for AI companies with distributed technical teams.
A startup might have machine learning engineers in California, a technical founder in Texas, contractors in New York, and a development team overseas.
The company should not assume all of those development costs receive the same federal R&D tax credit treatment.
Where the research is performed matters.
Potentially, yes. An AI startup does not need to be profitable to generate an R&D tax credit.
An eligible Qualified Small Business can elect up to $500,000 per year of its research credit against qualifying employer payroll taxes, subject to the QSB requirements and other applicable limitations.
This can be particularly relevant for early-stage AI companies investing heavily in U.S.-based technical talent before generating taxable income.
The company still needs to establish qualified research and calculate its QREs before determining how the resulting credit can be used.
AI companies use the same Section 41 research-credit framework as other businesses. The calculation is based on QREs, not the company's total AI budget.
Two AI companies spending the same amount on technical development can therefore generate very different credits because their:
For example, a $2 million "AI development" budget does not automatically mean the company has $2 million of QREs.
The company first needs to identify qualified activities and then connect eligible expenses to that research.
Section 174A and Section 41 address related but different aspects of R&D tax treatment.
Section 41 governs the R&D tax credit.
Section 174A governs the treatment of domestic research or experimental expenditures.
For tax years beginning after December 31, 2024, Section 174A generally permits taxpayers to deduct domestic research or experimental expenditures currently. Taxpayers can instead elect to capitalize those expenditures and amortize them over a period of at least 60 months. The provision specifically includes domestic amounts paid or incurred in connection with software development.
For AI businesses, the distinction matters because Section 174A expenditures and Section 41 QREs are not automatically the same population of costs.
An AI software-development expenditure may fall within the research-expenditure rules without necessarily generating an R&D tax credit. The Section 41 activities still need to meet the qualified-research requirements.
AI companies often generate particularly useful technical records because modern development workflows can track experiments in detail.
Useful documentation can include:
The goal is not to save every experiment forever. The documentation should help establish:
What technical problem existed → What alternatives were tested → What happened → Who performed the research → What the qualified research cost
For example, a record showing that one model outperformed another is useful. But the broader documentation should also explain why the alternatives were being evaluated and what technical uncertainty the team was trying to resolve.
AI development can produce a strong contemporaneous record of experimentation.
Experiment-tracking platforms and internal systems may automatically preserve:
These records can help show that alternatives were actually evaluated rather than reconstructed after the fact.
But experiment logs alone do not prove Section 41 qualification.
The company still needs to connect the experiments to a qualifying business component, technical uncertainty, qualified purpose, employees or contractors performing the research, and claimed expenses.
That is where technical and financial records need to work together.
Form 6765 is used to calculate and claim the federal research credit.
For tax years beginning after 2025, Section G is required for many taxpayers, subject to specific exceptions. Taxpayers required to complete Section G generally report business components representing at least 80% of total QREs, but no more than 50 business components, with remaining components reported in aggregate.
AI companies should pay particular attention to how their business components are identified.
A company should not necessarily treat "AI development" as one broad project when it is actually developing several distinct products, processes, software systems, techniques, formulas, or inventions.
The current instructions also tell taxpayers to use a business-component name or identifier consistent with the books and records used to substantiate the research and QREs.
If the AI business component is software, Section G also requires the applicable software classification, such as internal-use, dual-function, excepted from internal-use treatment, or non-internal-use software.
Potentially. Many states offer research incentives in addition to the federal R&D tax credit.
This can be particularly relevant for AI companies because technical departments are often distributed across the US.
A company headquartered in New York could have machine learning engineers in California, software developers in Texas, and contractors elsewhere.
State rules vary, so companies need to consider where the research was actually performed and which state's requirements apply, rather than automatically assigning all QREs to the
The biggest mistake is assuming AI equals R&D. It does not.
Other common problems include:
A stronger analysis begins with the research itself:
What technical result were we trying to achieve? What did we not know? What alternatives did we test? How did we evaluate them? What did that research cost?
AI companies should not need to create a parallel tax workflow for every experiment.
Instead, companies can build the R&D process around records their technical teams already generate.
A practical quarterly process could be:
This is much easier than asking a technical team at year-end to reconstruct hundreds or thousands of experiments from memory.
Start with the AI projects where the technical team did not already know the answer.
Look for development where engineers, data scientists, or researchers had to test different ways to achieve a technical result.
Then ask:
That project-first approach is much stronger than starting with an "AI budget" and assuming some percentage qualifies.
TaxTaker helps AI and software companies identify qualified research, calculate QREs, organize technical documentation, and prepare R&D tax credit claims in coordination with their CPA.
Book a call with TaxTaker to see which parts of your AI development may qualify.
AI development can qualify when the underlying research satisfies Section 41. The company generally needs to establish the Four-Part Test for the relevant business component, including the technological nature of the research and a process of experimentation.
Not automatically. Straightforward implementation of an existing API using known methods may involve little technical uncertainty. A broader development project involving experimentation with models, architectures, retrieval systems, or other technical approaches can present a different R&D analysis.
It can potentially form part of qualified research, but prompt engineering is not automatically R&D. The activity still needs to satisfy Section 41, including the technological and experimentation requirements.
Potentially. A RAG project can involve qualified research when engineers face technical uncertainty and experiment with retrieval architectures, embeddings, chunking, ranking, or other approaches to achieve required technical performance.
Certain costs associated with qualified AI research may potentially be QREs, but companies should not automatically classify all model-training or cloud costs as qualified. The nature of the expense, computing arrangement, and connection to qualified research need to be evaluated.
Yes, a pre-revenue company can potentially generate an R&D credit if it performs qualified research and incurs QREs. An eligible Qualified Small Business may also elect up to $500,000 of research credit per year against qualifying employer payroll taxes, subject to the applicable requirements.
.png)
Rachel Darrough is a Sr. R&D Manager with nearly 10 years of experience conducting federal and state R&D tax credit studies across various industry types, e.g., manufacturing, software, engineering, and construction. Rachel received a Bachelor's Degree in Managerial Finance and brings a strong technical foundation to evaluating qualified research activities, technical uncertainty, and experimentation under IRC §41. At Tax Taker, Rachel manages R&D engagements by collaborating with technical and finance teams to identify qualified expenditures, substantiate eligibility, and optimize credit outcomes. She applies an analytical approach to documentation and methodology while ensuring compliance with IRS guidance. Rachel is committed to helping clients leverage innovation-driven incentives to reduce tax liability and reinvest in continued R&D.
.svg)
