AI and machine learning companies develop models, algorithms, and systems that learn from data — including natural language processing, computer vision, predictive analytics, recommendation systems, and autonomous decision systems. The technical challenges can include developing model architectures for performance targets, engineering algorithms, optimizing inference performance, developing evaluation methods, and improving training approaches and system reliability. This page explains what development work may look like in an AI and machine learning business and how it relates to qualified research under Section 41. It is educational and is not individualized advice.
What R&D May Look Like in AI & Machine Learning
AI and machine learning development involves model architecture design, algorithm development, training approach optimization, evaluation method development, and system reliability engineering. Technical development may arise when a company develops a new model architecture for a performance target, engineers an algorithm, optimizes training or inference, develops evaluation methods, or improves system reliability. Work directed at resolving genuine technical uncertainty in these areas — through a structured evaluative process — may warrant review under the four-part test.
Industry-Specific Examples of Technical Development
- Developing model architectures for performance targets where the architecture performance is uncertain.
- Engineering algorithms for accuracy or efficiency where the algorithm performance is uncertain.
- Optimizing training and inference for performance where the optimization performance is uncertain.
- Developing evaluation methods for model assessment where the method performance is uncertain.
- Improving system reliability and latency where the reliability is uncertain.
None of these constitutes qualified research by itself. Each depends on whether the work satisfies all four elements of the four-part test.
Technical Uncertainty Examples
- Whether a new model architecture can achieve the specified accuracy target within the latency constraint.
- Whether an alternative training approach can achieve the specified convergence and generalization targets.
- Whether a new evaluation method can reliably assess the specified performance dimensions.
For more, see our page on elimination of uncertainty.
Process-of-Experimentation Examples
- Implementing and testing alternative model architectures, measuring accuracy and latency, and comparing results.
- Training with alternative approaches, measuring convergence and generalization, and evaluating results.
- Developing and testing alternative evaluation methods, measuring assessment reliability, and comparing results.
For more, see our page on process of experimentation.
Potential Business Components
Potential business components may include a new or improved product (an AI system with improved performance), a new or improved process (a training or inference process), or a new or improved technique (a model architecture or evaluation method).
Employee Work That May Warrant Analysis
Employees whose work may warrant analysis include ML engineers developing architectures, data scientists developing algorithms, research engineers developing approaches, and quality engineers conducting evaluation and reliability testing tied to a development project. For more, see our page on R&D tax credit employee wages.
Contractor Work That May Warrant Analysis
Contractor work that may warrant analysis includes data providers, specialized consultants, and testing service providers — where the company bears the economic risk and retains substantial rights. For more, see our page on R&D tax credit contractor costs.
Supplies and Materials That May Become Relevant
Supplies in AI and machine learning are often limited because the work is primarily computational. Where tangible materials are consumed in specialized hardware testing or prototyping, they may become relevant. For more, see our page on R&D tax credit supplies.
Activities That Generally Require Caution or May Not Qualify
- Simply using an existing AI model or API without developing or improving a business component.
- Routine model fine-tuning using established methods.
- Standard data preparation and cleaning.
- Ordinary bug fixes and maintenance.
- Copying or deploying an existing model for a new application without resolving technical uncertainty.
Documentation That May Help
Records that may help include project descriptions, architecture and algorithm evaluation results, training and inference test data, and records connecting personnel to specific development projects. For more, see our page on R&D tax credit documentation.
Example Hypothetical Project
The following is a hypothetical example for illustration only. It does not represent any actual company and does not state that the work qualifies.
An AI company is developing a computer vision model for an industrial inspection application where the standard model architecture does not achieve the specified accuracy target within the latency constraint. The technical uncertainty is whether an alternative model architecture, a modified training approach with augmented data, and a new evaluation methodology can together achieve the accuracy, latency, and reliability targets. The team implements and tests three model architectures with two training approaches, measures accuracy, latency, and generalization, and evaluates the assessment methodology. Based on the results, the team selects an architecture and training approach and refines the evaluation method. Records of the alternatives, test conditions, and results may help support analysis — but professional review is still needed.
Questions to Ask Internally
- What specific business component was being developed or improved?
- What technical uncertainty existed at the outset?
- What alternatives were evaluated, and how were they tested?
- Who performed or directly supported the work?
- What materials were consumed in the testing?
- How does this differ from simply using an existing AI model?
Relationship to the Four-Part Test
The four-part test applies the same way it does in any industry. The work must be directed at developing or improving a business component (permitted purpose), must fundamentally rely on principles of the physical sciences or engineering (technological in nature), must be intended to eliminate a technical uncertainty (elimination of uncertainty), and must be conducted through a structured evaluative process (process of experimentation). Meeting one element is not enough.
Key Takeaway
AI and machine learning companies may perform activities that warrant analysis under IRC §41 — particularly work involving model architectures, algorithms, training approaches, and evaluation methods. Simply using an existing AI model does not automatically qualify. Professional review is appropriate. For related industries, see our pages on SaaS companies and custom software development.