At Leyton, we work closely with innovative companies that are adopting Generative AI to accelerate research and development. The opportunities are immense, but when it comes to the Scientific Research & Experimental Development (SR&ED) program, it is crucial to understand that not all AI activities are eligible.
The distinction between experimental development and routine application is subtle for this emerging technology. Recognizing it ensures businesses maximize their claims while maintaining compliance.
Training Data: The Critical Differentiator
One of the first questions to assess for eligibility is: how is the data being handled?
This often defines whether a project qualifies:
- Eligible: Creating, curating, or engineering new datasets specifically for training or advancing AI models. This could involve cleaning raw, unstructured data, building synthetic datasets, developing novel labeling techniques, or constructing automated data pipelines. These activities demonstrate systematic experimentation and technological uncertainty, both required under SR&ED.
- Not Eligible: Simply applying publicly available datasets or using off-the-shelf data directly with a pre-trained model. This is considered routine application rather than experimental development.
In short: training and refining your own data to solve a novel problem is eligible; relying on pre-existing data as-is is not.
Model Creation and Iteration
Another key factor is the level of experimentation in model development:
- Eligible: Designing models from scratch or iteratively experimenting with architectures and developing different approaches. These activities involve uncertainty and experimentation, for example, not knowing whether a model will perform as intended and systematically testing different iterations to overcome limitations.
- Not Eligible: Integrating pre-trained generative models through APIs or using them “out of the box” without engaging in deeper experimentation. These are routine deployments, not SR&ED.
Put simply: if the model is being treated as a black box, the work is not eligible. If it is being opened, modified, and pushed beyond its intended capabilities, then SR&ED eligibility applies.
Generative AI as a Supporting Tool
Generative AI does not need to be the primary focus of the R&D project to be eligible. It can also qualify when it supports other core SR&ED activities.
For example, consider a company conducting experimental development in biotech. If Generative AI tools are used to automate complex data cleaning, this supporting activity may also be claimed, because it directly enables the eligible R&D.
This is a critical point: AI activities that contribute to advancing the experimental development, not just the business, can form part of an SR&ED claim.
Documentation and Evidence
Clear records are essential to distinguish experimental development from routine application.
Leyton recommends capturing:
- Iterations: Each attempt, including both successes and failures.
- Training cycles: How datasets and models evolved over time.
- Uncertainties: What was unknown at the start, and how the team resolved it.
Strong documentation not only protects the claim but also builds a repository of internal intellectual property for future projects.
The Business Impact
Generative AI offers organizations a chance to create defensible innovation. By training proprietary datasets, building custom models, and experimenting at the edge of what’s possible, businesses unlock both:
- SR&ED funding to support their R&D.
- Strategic IP advantages that differentiate them in competitive markets.
Conversely, companies that limit AI use to routine applications risk missing out on both financial support and long-term innovation benefits.
Key Takeaways
Generative AI can qualify under SR&ED, but only when approached as experimental development rather than simple implementation.
- Eligible: Training datasets, designing and iterating on models, experimenting with architectures, or using AI to support core R&D such as advanced data cleaning.
- Not Eligible: Plug-and-play use of pre-trained models, off-the-shelf datasets, or deployments without systematic experimentation.
At Leyton, we believe the bottom line is simple: eligibility depends on innovation, not application. Companies that experiment, iterate, and push technological boundaries are not only advancing AI but also unlocking the SR&ED incentives designed to support this type of work.
Contact our experts to find out if your AI project qualifies for SR&ED!
