AI + the R&D Tax Credits: What Qualifies, What Doesn’t, and How to Claim It in 2026

Whether you are building AI from the ground up or deploying it to transform your operations, your company’s AI investment may qualify for the federal R&D tax credit. This guide covers what counts, what falls short, and how the rules apply after OBBBA restored immediate R&D expensing.

Key Highlights

• Two paths to qualification: AI builders and AI adopters both qualify under IRC Section 41.
• The Four-Part Test applies naturally: AI work involves inherent technical uncertainty and iterative experimentation, the core of what the IRS looks for.
• Generative AI and vibe coding don’t disqualify: Uncertainty shifts rather than disappears. Human testing and design decisions around AI-generated code can constitute qualifying experimentation.
• 2026 is the strongest year to claim since 2021: OBBBA’s Section 174A restores immediate expensing, so your QREs generate both a full deduction and a credit in the same year.
• Cloud compute is a major QRE: GPU/TPU training costs qualify under Section 41(b)(2)(A)(iii). Tag R&D workloads separately from production.
• Startups can offset payroll taxes: Up to $500K/year against FICA, even with zero income.
• Documentation standards are rising: Form 6765 Section G is mandatory for 2026 (exceptions for QSB payroll filers and ≤$1.5M QREs).
• State conformity varies: Not all states have conformed to Section 174A. Multi-state filers should verify each state’s treatment.

Does AI Qualify for the R&D Tax Credit? Two Types of Companies, One Answer

There are two categories of companies that should evaluate their AI work under IRC Section 41:

Companies that build AI are the obvious candidates. If your core product involves developing ML models, NLP systems, computer vision, or generative AI platforms, most of your development projects will strongly align with qualified research activities. The cycle of hypothesis, experimentation, testing, and iteration is exactly what the credit rewards.

Companies that use AI to transform their operations are where the biggest gap exists. A manufacturer building a predictive maintenance system, a healthcare organization developing an AI diagnostic tool, a financial services firm creating a fraud detection model, or an ag business deploying AI crop analysis can all qualify. The work they did to develop, customize, integrate, and validate AI within their operations meets the threshold if it involved technical uncertainty and experimentation.

One important note: the IRS applies a higher threshold of innovation test for internal-use software under Treas. Reg. §1.41-4(c)(6). If your AI system is used solely for internal operations, you may need to demonstrate that the software is innovative, involves significant economic risk, and is not commercially available. See our software R&D credits guide for more.

The AI Four-Part Test: How the IRS Evaluates Your R&D Claim

Every R&D credit claim must pass the IRS Four-Part Test under Section 41(d). There are no special carve-outs for AI. Here is how each prong applies:

Why AI is a natural fit: Traditional software follows known patterns with predictable outcomes. AI is different. You genuinely do not know at the outset whether your approach will work. That inherent uncertainty, combined with iterative experimentation, is why AI projects across industries are strong R&D credit candidates.

AI Activities That Qualify for the R&D Tax Credit

The following categories pass the Four-Part Test when technical uncertainty and experimentation are present. These apply to AI-native companies and AI adopters alike:

• Model development and training: Designing ML architectures, experimenting with neural networks, building generative models, developing proprietary algorithms. Includes deep learning, reinforcement learning, transformers, and diffusion models.
• Data pipeline and feature engineering: Novel data ingestion, cleaning, transformation, and augmentation methods. Experimenting with feature extraction, building training pipelines at scale, handling noisy or unstructured datasets.
• NLP, computer vision, and multi-modal AI: Language understanding, sentiment analysis, image recognition, video analysis, and applications combining text, image, and structured data.
• Custom AI integration: Embedding AI into existing systems: real-time prediction APIs, decision-automation layers, IoT/robotics integration. This is where non-tech companies often qualify. See our software & technology guide.
• AI-powered product and process innovation: Predictive analytics, diagnostic tools, recommendation engines, conversational interfaces, predictive maintenance, and supply chain optimization.
• Infrastructure and compute optimization: Distributed training, GPU/TPU orchestration, model parallelism, edge deployment, and novel compute routing to reduce costs or improve speed.

AI Activities That Do Not Qualify

• Off-the-shelf AI deployment: Subscribing to an AI SaaS and using it as-is. Turnkey chatbots, pre-built APIs, commercial tools without customization.
• Basic prompt engineering: Writing prompts for LLMs without fine-tuning or proprietary integration. (Prompt work that is part of building a RAG system or fine-tuning effort may qualify as part of the broader project.)
• Routine configuration, QA, and monitoring: Standard deployment, bug fixes on stable systems, and production monitoring are operational, not experimental.
• Market research and business analysis: Social science activities do not qualify under Section 41.
• Employee training on AI tools: Onboarding and change management are administrative.
• Duplicating prior solutions: Repeating a known approach without new uncertainty does not qualify.

Real-World Scenarios: Qualifies vs. Doesn’t

Generative AI R&D Tax Credit: Vibe Coding, Agents, and the New Frontier

Developers now use tools like GitHub Copilot, Cursor, and ChatGPT to generate code from natural language. Practices like “vibe coding” and “agentic coding” are reshaping the development process. The IRS has not issued specific guidance, but the Section 41 framework does not change based on tools used.

Recent practitioner analysis confirms that when developers use AI to generate code, uncertainty shifts rather than disappears. Instead of struggling with syntax, the developer faces higher-order questions: Which AI-generated architecture scales? Will the code integrate with existing systems? How do you validate quality across edge cases? The developer’s testing, debugging, validation, and design decisions constitute the real experiment.

Qualifies

• Fine-tuning LLMs on proprietary data for domain-specific use cases
• Building RAG systems with proprietary data pipelines, retrieval logic, and output validation
• Developing custom AI agents with multi-step workflows and decision logic
• AI-assisted development where the developer retains ownership of design, testing, and uncertainty resolution

Does Not Qualify

• Generating boilerplate code for routine tasks with no experimentation
• Prompt engineering without model modification or proprietary integration
• Automating simple tasks with an off-the-shelf LLM API

The IRS audits the process, not the tool. Document the human layer: design choices, alternatives evaluated, tests performed, failures encountered, and the rationale behind your final approach.

R&D Tax Credit for AI Companies: Qualified Research Expenses

Here is how QREs under Section 41(b) apply to AI work:

Cloud costs are a growing blind spot. If your AWS/Azure/GCP billing lumps R&D and production compute together, you are leaving money on the table or creating audit risk. Tag R&D workloads at the infrastructure level and tie invoices to specific qualified projects.

How OBBBA and Section 174A Changed the Math on AI Spending

The One Big Beautiful Bill Act, signed July 4, 2025, created Section 174A, restoring immediate expensing for domestic R&E costs for tax years beginning after December 31, 2024. From 2022 through 2024, the TCJA required five-year amortization, which crushed the value of both R&D deductions and credits. Section 174A eliminates that constraint.

For AI-heavy companies, this creates a compounding benefit: full immediate deduction plus the credit in the same year. Companies under the $31M gross receipts threshold can also apply Section 174A retroactively to 2022 through 2024 via amended returns (see Rev. Proc. 2025-28). Larger businesses can accelerate unamortized deductions over 2025 and 2026. For more, read our OBBBA R&D expensing breakdown.

With Section 174A restored, the Section 280C election matters more: full gross credit with add-back vs. reduced credit (~79%) with full deduction. Model both before filing. Details in our Section 280C guide.

State conformity warning: Not all states have conformed to Section 174A as of early 2026. Multi-state filers should verify each state’s treatment. Your state R&D credit eligibilitymay vary independently of federal changes.

Documentation That Survives an Audit

The IRS expects business-component-level substantiation. Form 6765 Section G is optional for 2025 and mandatory for 2026, with exceptions for QSB payroll-only filers and taxpayers with ≤$1.5M in QREs (per IRS IR-2025-99). For AI projects, capture:

• Project descriptions: What you built, the technical goal, what uncertainty existed, and why off-the-shelf was insufficient.
• Experiment logs: Architectures tested, configurations evaluated, results, and reasons for pivots.
• Failure records: Failed model runs, abandoned approaches, and the reasoning. In AI, well-documented failure is gold.
• Time tracking by project: Individual-level allocation between qualified research and non-qualifying work.
• Expense records: Cloud invoices, contractor agreements, and supplies tied to specific R&D projects.
• Version control: Git logs, code reviews, and experiment tracking tools (MLflow, W&B) as contemporaneous evidence.

Five Mistakes Companies Make

Mistake 1: Thinking the credit only applies to “AI companies.” Manufacturers, healthcare organizations, logistics firms, and retailers building or integrating AI all qualify. The test is the activity, not the industry.

Mistake 2: Missing cloud computing costs. GPU training for R&D is a fast-growing QRE category. If your cloud billing doesn’t separate R&D from production, fix that now.

Mistake 3: Not tracking time at the project level. The IRS expects individual-level allocation by project. Department-level tracking is not sufficient.

Mistake 4: Overlooking integration work. Data pipelines, APIs, and system integrations needed to make AI work often involve significant experimentation. That’s a QRE source.

Mistake 5: Documenting retroactively. Contemporaneous records (experiment logs, commits, time entries) are stronger than reconstructed ones. With Section G going mandatory, start now.

How to Get Started

Step 1: Review your AI projects against the Four-Part Test. Include development and integration work across all teams.

Step 2: Map wages, cloud costs, supplies, and contractor payments to qualifying activities. Separate R&D from production spend.

Step 3: Assess whether your records support each activity at the business-component level. Start building documentation systems now.

Step 4: Engage a specialist. The intersection of AI and Section 41 is nuanced. An experienced advisor can identify missed expenses, validate documentation, and model the optimal credit strategy.