Artificial Intelligence (AI) is transforming industries at an unprecedented pace. From predictive analytics and natural language processing to autonomous systems and intelligent automation, AI is no longer a futuristic concept — it’s a competitive necessity.
Yet while many companies, software-based or otherwise, are investing heavily in AI, few are fully leveraging one of the most powerful financial incentives available to them: the Research & Development (R&D) Tax Credit.
At Leyton USA, we specialize in helping companies across industries identify and claim qualifying R&D activities. In this article, we’ll explore how AI-related work can qualify for the R&D Tax Credit, clarify common misconceptions, and highlight the types of AI development that do, and don’t, meet the IRS’s criteria.
What Is the R&D Tax Credit?
The R&D Tax Credit is a federal (and often state-level) incentive designed to reward companies for investing in innovation. It provides a dollar-for-dollar reduction in tax liability for qualified research expenses (QREs), including wages, supplies, contract research, and cloud service provider costs.
To qualify, a project must meet the IRS’s Four-Part Test:
- Permitted Purpose : The activity must aim to create or improve a proprietary product, process, technique, formula, computer software, or invention.
- Technological in Nature: The activity must rely on principles of a hard science, such as physical or biological sciences, engineering, or computer science.
- Elimination of Uncertainty: The company must attempt to eliminate uncertainty about the capability, method, or design of the product or software.
- Process of Experimentation: The company must evaluate one or more alternatives through modeling, simulation, systematic trial and error, or other methods.
In simple terms, qualified R&D involves developing or improving intellectual property (IP) through experimental, hard science-based methods to enhance a measurable aspect of that IP or its outputs.
And here’s a bonus: for software written from scratch, you own the IP the moment you start coding. Even better, QREs invested towards awarded patents benefit from safe harbor qualification.
AI and the R&D Tax Credit: What Qualifies?
AI development often involves complex technical challenges, making it a strong candidate for the R&D Tax Credit, provided the work meets the Four-Part Test. Here are some examples of AI-related activities that typically qualify:
Training and Fine-Tuning AI Models
If your company is training, re-training, fine-tuning, or otherwise modifying an AI model to meet specific business needs, this work likely qualifies. These activities involve technical uncertainty and require experimentation to optimize performance, accuracy, or efficiency.
Developing Proprietary AI Software
Writing proprietary software to support or integrate AI functionality, such as custom data pipelines, model orchestration frameworks, or inference engines, can also qualify. This includes building tools to manage model deployment, monitor performance, or ensure compliance with data privacy regulations.
Algorithm Development and Optimization
Creating new algorithms or significantly modifying existing ones to improve AI performance (e.g., reducing latency, increasing accuracy, or enabling real-time processing) is a classic example of qualified R&D.
Integrating AI into Novel Applications
If your team is integrating AI into a new or significantly improved product or platform, and the integration requires overcoming technical hurdles, those efforts may be eligible.
What Doesn’t Qualify?
Not all AI-related work meets the IRS’s criteria. It’s important to distinguish between using AI and developing or improving it. Here are some common activities that do not qualify:
Using Pre-Trained Models “Out of the Box”
Simply using an existing AI model (e.g., OpenAI’s GPT, Google’s BERT, or Meta’s LLaMA) without modifying it does not qualify. This includes:
- Connecting a pre-trained model to your application via an API
- Using AI tools for content generation, summarization, or translation
- Running inference on a model without altering its architecture or training data
Prompt Engineering Alone
While prompt engineering can be creative and valuable, it typically doesn’t involve technical uncertainty or a process of experimentation as defined by the IRS. More specifically, the success of the activity is not rooted in a hard science but rather speculation of the existing model’s output. Unless it’s part of a broader effort to fine-tune or retrain a model or systematically prompt a model to work with your proprietary software, prompt engineering on its own is not a qualified activity.
Workflow Automation Using AI Tools
Creating workflows that leverage AI tools to fully automate a task in lieu of software development (e.g., automating document classification using a large language model (LLM)) is generally considered application of existing technology, not development. These activities simply utilize the existing functionalities of the software, whether third-party or proprietary.
Practical Examples
Let’s look at how different approaches to building a customer support chatbot affect R&D eligibility:
Example 1
Not Qualified: You use OpenAI’s GPT-4 API to answer customer questions and write prompts to improve responses.
Qualified: You fine-tune a transformer model on your proprietary support data to improve accuracy, develop a custom intent recognition engine, and build a backend system to manage model updates and compliance.
The difference lies in whether your team is creating or improving technology, versus simply configuring and applying it.
Example 2
Even when the AI model itself isn’t modified, the surrounding proprietary software may still qualify. Consider how the following set of approaches to building the chatbot affect R&D qualification:
Not Qualified: You configure connections between Google’s Gemini 2.5 large language model (LLM), HubSpot, and Zendesk using Zapier to automate support responses.
Qualified: You develop a proprietary retrieval-augmented generation (RAG) system and model management infrastructure, including:
- A Python-based data ingestion and transformation pipeline that extracts, cleans, and formats internal company data (e.g., product documentation, support tickets, service logs).
- A retrieval system that dynamically selects relevant documents based on user inquiries, chat history, and metadata (e.g., user profile, order history, subscription tier).
- A prompt construction engine that assembles this context into structured input for GPT-4, ensuring the model has the most relevant information.
- A post-processing layer that validates and filters the model’s output to align with internal support policies and tone guidelines.
- Experimentation to optimize retrieval relevance, reduce latency, and maintain response accuracy across diverse user scenarios.
In this case, the development of the proprietary software system is the qualified activity, not the use of the AI model itself. This example illustrates that you don’t have to alter the model to qualify for the R&D Tax Credit, as long as the software you build around it involves technical challenges and innovation.
Business Components and the Four-Part Test
In short, when AI-related work qualifies, it’s typically claimed under the category of a software business component, which is the product or intellectual property being developed or improved.
If your company is training, re-training, fine-tuning, or otherwise modifying the model, then the model itself may be claimed as a software business component. Additionally, if you’re writing proprietary software related to the AI (third-party or otherwise), that software may also qualify.
Of course, business component projects must still satisfy the other three criteria of the Four-Part Test. That’s where we come in, helping you determine what qualifies and how to document it.
How Leyton USA Can Help
At Leyton USA, we specialize in helping software companies navigate the complexities of the R&D Tax Credit. Our technical consultants understand both the software development lifecycle and the nuances of IRS compliance. We work closely with your engineering and finance teams to:
- Identify qualifying AI-related activities
- Document the technical challenges and experimentation involved
- Calculate and substantiate your qualified research expenses
- Maximize your credit while minimizing audit risk
Whether you’re building AI from the ground up or enhancing existing platforms, we’re here to help you turn innovation into savings.
Ready to Claim Your Credit?
AI is reshaping the future of software, and your company’s investment in AI deserves to be rewarded. Don’t leave money on the table.
Contact Leyton USA today to find out how much of your AI development work qualifies for the R&D Tax Credit.
Book a meeting now, and unlock your latent savings with Leyton!
