Using AI vs. Building AI: What Actually Qualifies for the R&D Tax Credit in 2026
Quick answer: Simply using AI tools, such as subscribing to a chatbot or plugging an off-the-shelf model into a workflow, generally does not qualify for the R&D tax credit. Building AI often does. Training, fine-tuning or significantly modifying models, developing new algorithms, and writing proprietary software around AI can qualify when the work meets the IRS four-part test: a permitted purpose, technological uncertainty, a process of experimentation, and reliance on computer science or engineering.
Almost every company now says it “does AI.” Marketing teams use generative tools to write copy. Finance teams use AI assistants to summarize reports. Engineering teams build models from scratch. When tax season comes around, the natural question is whether any of this counts toward the R&D tax credit.
The honest answer is that some of it does and a lot of it doesn’t. The dividing line is usually whether you are using AI or building it.
Why the Distinction Matters
The federal R&D tax credit under Section 41 rewards qualified research. It doesn’t reward adopting technology, even impressive technology. The IRS looks at what your team actually did, what they were uncertain about, and how they worked through that uncertainty.
That’s why two companies with the same AI budget can have completely different credit outcomes. One paid for licenses and trained staff to use them. The other hired engineers to design, test and refine a new model. Only the second is likely to have qualifying research.
What "Using AI" Usually Looks Like
Activities like these generally don’t qualify on their own:
- Subscribing to AI writing, coding or chat assistants
- Using a vendor’s model through its standard interface without modification
- Writing prompts for everyday business tasks
- Configuring settings in an AI feature built into existing software
- Training employees to use AI tools
- Running routine analytics with an existing AI product
In each case, the company is a user of someone else’s technology. There’s no new or improved business component being developed through experimentation.
What "Building AI" Usually Looks Like
These activities are much more likely to qualify:
- Training a model on proprietary data to solve a specific technical problem
- Fine-tuning a foundation model and testing different architectures, parameters or training approaches
- Developing new algorithms for prediction, classification, vision or language tasks
- Building data pipelines where the approach to cleaning, labeling or structuring data was uncertain
- Creating retrieval, orchestration or evaluation systems where the design wasn’t known upfront
- Optimizing model performance for speed, accuracy or cost in ways that required testing alternatives
- Integrating AI into a product when the method of integration involved real technical uncertainty
A trained or modified AI model may be part of a software business component when the underlying development activity meets the applicable business-component and qualified-research requirements.
For a closer look at how the credit works across sectors, see our industries served page.
Applying the Four-Part Test to AI Work
1. Permitted Purpose
The work must aim to create a new or improved business component, meaning better function, performance, reliability or quality. A model that detects defects more accurately or processes claims faster fits here.
2. Technological Uncertainty
At the start, your team must have been uncertain about the capability, the method or the appropriate design. With AI, uncertainty is common. Will the model reach the needed accuracy? Which architecture works with limited data? Can inference run fast enough on your hardware?
3. Process of Experimentation
This is where AI projects often shine. Training runs, hyperparameter searches, A/B tests of model versions and evaluation against benchmarks are classic experimentation. Keep the records.
4. Technological in Nature
Machine learning relies on computer science, mathematics and statistics, so this prong is usually satisfied for genuine AI development.
Our research credit page explains the test in more depth.
The Internal-Use Software Trap
Not all AI development is treated the same. If your team builds an AI tool mainly for internal back-office functions, such as HR screening, accounting automation or internal reporting, it may be classified as internal-use software. That triggers an additional three-part high threshold of innovation test, which is much harder to meet.
AI built into a product customers use, or software that lets third parties interact with your systems, is generally not internal-use software and is evaluated under the standard four-part test.
What Costs Can Count?
Qualified research expenses for AI projects typically include:
- Wages for engineers, data scientists and technical leads who perform, directly supervise or directly support the research
- Supplies consumed in research, which is less common in pure software work
- Computer use costs, such as amounts paid to use cloud computing for development and testing, which may qualify in some cases
- Contract research, generally at 65% of amounts paid to U.S. contractors performing qualified research on your behalf
Research must be performed in the United States. Work done by offshore teams doesn’t qualify for the federal credit.
Where Section 174A Fits In
The One Big Beautiful Bill Act restored immediate expensing of domestic research and experimental expenditures under new Section 174A for tax years beginning after 2024. That’s a separate benefit from the credit, but the two interact, so AI companies should model both together.
Documentation Tips for AI Teams
AI teams often create excellent records without realizing it. Preserve:
- Experiment tracking logs (for example, MLflow or Weights & Biases runs)
- Git history, pull requests and code reviews
- Jira or Linear tickets describing technical problems
- Model evaluation reports and benchmark comparisons
- Design documents describing alternatives considered
- Time allocations tying engineers to specific projects
The IRS wants to see the uncertainty, the alternatives you tested and the costs tied to that work. Records created during the project carry far more weight than summaries written later.
A Quick Example
A logistics software company uses an AI chatbot for customer support. That alone isn’t research. The same company also builds a proprietary route optimization model, tests several neural network approaches, and iterates to reduce delivery time estimates by a meaningful margin. The engineers’ time on that model, plus qualifying cloud computing used for training and testing, may count toward the credit.
Questions to Ask About Each AI Project
Before claiming any AI work, walk through a few simple questions with your technical leads. What exactly were we trying to build or improve? What didn’t we know at the start? Which approaches did we try, and how did we measure them? Who worked on it, and for how long? If the answers are specific and backed by records, the project is a strong candidate. If they’re vague, the project may be routine adoption rather than research.
Key Takeaways
- Using AI tools usually doesn’t qualify; building or modifying AI often does.
- AI work must still pass the four-part test.
- Internal AI tools may face the stricter internal-use software rules.
- Experiment logs, code history and tickets make strong documentation.
Frequently Asked Questions
Does using ChatGPT or other AI tools qualify for the R&D tax credit?
Generally no. Using off-the-shelf AI tools for business tasks isn’t qualified research. The credit rewards developing new or improved technology through experimentation.
Does fine-tuning an AI model qualify for the R&D tax credit?
It can. Fine-tuning or training a model to solve a technical problem, when the outcome is uncertain and your team tests alternatives, may meet the four-part test.
Are data scientists’ wages eligible for the R&D credit?
Yes, the portion of wages for time spent performing, directly supervising or directly supporting qualified research can count as qualified research expenses.
Do internal AI tools qualify for the R&D credit?
They may, but AI built mainly for internal back-office functions can be treated as internal-use software, which must also pass the high threshold of innovation test.
Does AI research done overseas qualify?
No. Only research performed in the United States qualifies for the federal R&D tax credit.
Building AI? Get a free estimate from TaxDrone.AI to see what your work could be worth.
This article is general information and not tax advice. Consult a qualified tax professional about your situation.
