Artificial intelligence is now embedded across nearly every sector of the economy. Companies are using AI in drug discovery, fraud detection, autonomous vehicles, robotics, and financial services. For R&D tax credit purposes, the question is not whether a company “uses AI.” It is whether the company is actually building or improving technology that involves technical uncertainty and a process of experimentation. Here is what qualifies for AI R&D tax credits 2026 and how to document it.
The Four-Part Test for AI R&D Tax Credits 2026
To qualify, your AI activities must meet all four IRS requirements under IRC Section 41.
1. Permitted Purpose
Your work must aim to develop or improve a product, process, software, or system.
Examples:
-
Building proprietary machine learning models
-
Developing new fraud detection algorithms
-
Creating custom recommendation engines
2. Technological in Nature
The work must rely on computer science, engineering, or other technical disciplines.
Examples:
-
Neural network architecture design
-
Data pipeline engineering
-
Algorithm optimization
3. Elimination of Technical Uncertainty
You must face uncertainty about capability, method, or design at the project’s outset.
What creates uncertainty:
-
Unknown whether a model can achieve required accuracy
-
Uncertainty about scalability at expected load
-
Questions about system integration into legacy environments
-
Doubt about whether a technical approach can satisfy speed or cost constraints
What does not create uncertainty: Commercial uncertainty (“Will customers buy this?”).
4. Process of Experimentation
You must show systematic evaluation through testing, modeling, or trial and error.
Qualifying activities:
-
Testing alternative model architectures
-
Comparing different training methodologies
-
Evaluating multiple integration approaches
Failed experiments count. What matters is the systematic attempt to solve a technical problem.
The Core Distinction: Development vs. Deployment
This is the most important distinction for AI R&D credits.
| Category | Qualifying Activity | Non-Qualifying Activity |
|---|---|---|
| Model Development | Building proprietary ML models | Using off-the-shelf AI APIs |
| Algorithm Design | Creating novel algorithms | Implementing existing algorithms |
| Integration | Custom solutions for unique challenges | Vendor tool installation |
| Training | Developing novel fine-tuning techniques (e.g., new LoRA variations) to resolve technical uncertainty | Routine fine-tuning using pre-built scripts or standard transfer learning |
Important distinction: Routine transfer learning using pre-built scripts does not qualify. But fine-tuning that requires developing new parameter-efficient tuning techniques (e.g., novel LoRA variations) to resolve technical uncertainty can qualify.
If the initiative is about building or materially improving technology, it may qualify. If it is mostly about deploying a finished tool, activating existing software, or training people to use a purchased platform, it likely does not.
Understanding AI R&D tax credits 2026 starts with knowing the difference between development and deployment.
What AI Activities Typically Qualify
| Activity Type | Examples |
|---|---|
| Model Development | Creating proprietary ML models, neural networks, deep learning frameworks |
| Algorithm Optimization | Designing novel algorithms, improving model accuracy, reducing computational requirements |
| Custom Data Processing | Developing automated labeling techniques, unique data augmentation strategies |
| System Integration | Engineering custom solutions to integrate AI into production environments |
| Performance Engineering | Optimizing inference speed, scalability, and reliability through systematic testing |
What Usually Does NOT Qualify
-
Using off-the-shelf AI tools: Implementing existing large language models into a chatbot without advancing the technology
-
Routine configuration: Standard customization of vendor platforms
-
Dashboards and reporting: Creating data visualizations without technical development
-
Automation without uncertainty: Automating tasks with pre-trained models where no technical uncertainty exists
-
Generic implementation: Using AI to generate code without addressing technical questions about how the work is integrated, scaled, or validated
How AI Costs Are Treated: Important Differences
Not all AI costs follow the same Section 41 treatment. You should separate expenses into categories:
| Cost Type | Section 41 Treatment |
|---|---|
| Cloud GPU and training compute | 100% included as computer rental expense if the underlying activity qualifies |
| API-based AI tools | High audit risk if structured as purchasing a completed API output service. May qualify under computer rental QREs only if contractually structured around reserved compute capacity (e.g., dedicated GPU compute nodes) rather than pay-per-token usage. The IRS frequently challenges API calls as non-qualifying third-party service fees rather than computer rental expenses under Section 41(b)(2)(A)(iii). |
| SaaS AI subscriptions | Allocate portion tied to qualified research. Software licenses generally excluded. |
| On-premises hardware | Depreciation on owned hardware is NOT a qualified expense. But wages for engineers working on qualifying research may still count. |
| AI-assisted contractors | 65% of expenses can be included under contract research rules, subject to funded-research tests. |
Key principle: Allocate carefully. A team using AI tools for routine bug fixes, support tickets, and prototype work is engaged in qualified research only part of the time. Defensible claims require contemporaneous allocation methods based on usage data or time studies.
Internal Use Software Rules for AI
For AI R&D tax credits 2026, cost treatment depends on how your AI expenses are structured. If your AI tool is built purely for internal business operations—not for commercial sale or external customer interaction—it may face stricter rules under the Internal Use Software (IUS) framework.
Under Treasury Regulations, IUS must meet the High Threshold of Innovation test:
1. Substantial Innovation: The software must result in a significant, measurable reduction in cost or improvement in speed.
2. Significant Economic Risk: The company must commit significant resources with substantial uncertainty about recovery.
3. Not Commercially Available: The software cannot be purchased, leased, or licensed and used without significant modification.
The Third-Party Subset Exception: If your AI tool allows third parties to interact with your systems (e.g., customer-facing APIs, partner portals), you may carve out those external-facing functions and avoid the IUS test. Code and development costs dedicated solely to enabling third-party interaction are evaluated under the standard four-part test instead.
How AI Affects Qualification: Two Key Issues
1. Technical Uncertainty
When AI tools successfully generate code or designs, it becomes harder to argue that the taxpayer faced the same technical uncertainty that existed before these tools were available.
Uncertainty doesn’t disappear. It shifts to different questions: Which model to use? How to assess reliability? Can the result be validated in real-world conditions?
2. Process of Experimentation
An engineer who simply accepts an AI-generated answer without comparing options has a weaker case. A team that uses AI to generate several approaches, tests them, measures performance, and rejects some on technical grounds has a stronger experimentation narrative.
The New Reality: IRS Is Using AI to Audit R&D Claims
The IRS has deployed 129 AI use cases, up from 54 just two years ago. Returns are scored on multiple criteria, and AI systems run six times per tax year, learning with each iteration.
For R&D credit filers specifically:
-
Claims are evaluated in the context of financial history, industry peers, and patterns identified across thousands of similar filers
-
Revenue agents use generative AI to draft information document requests and exam reports
-
AI-assisted document analysis can surface inconsistencies faster than manual review
The implication: Documentation must be specific, contemporaneous, and defensible. Generic narratives and AI-generated summaries carry significant audit risk.
Documentation Best Practices for AI R&D
The IRS expects contemporaneous records—not reconstructed studies. For AI projects, that means:
What to document:
-
The technical objective
-
What was uncertain at the outset
-
What alternatives were tested
-
How AI was used in the process
-
Why one approach was selected over others
How to document it:
-
Design specifications
-
Project plans and sprint records
-
Architecture decks and release notes
-
Testing results and model evaluation files
-
Code repositories and Jira tickets
Strong documentation explains: What was tried, what failed, what was learned, and what ultimately worked. That demonstrates a process of experimentation.
A team using AI to generate approaches, test them, measure performance, and reject unsatisfactory options has a strong narrative. A team that simply accepts AI-generated output without comparison does not.
Bottom Line
AI R&D tax credits are real—but the same rules apply. Using AI does not automatically create a credit.
The strongest opportunities arise when companies build or improve technology through technical uncertainty and a systematic process of experimentation. Model development, algorithm optimization, custom integration, and performance improvement may all qualify.
If the initiative is primarily about deploying a finished tool or activating existing software, the answer is different.
Call (844) 463-2400 or email hello@indagotax.com to find out what qualifies in your AI development work.