When Does LLM Work Qualify for SR&ED?
Nov 19, 2026

Canadian technology companies are increasingly using large language models (LLMs) to develop new products, improve processes, and solve complex problems. But does all this work qualify for Scientific Research and Experimental Development (SR&ED) tax incentives?
Not necessarily. While some LLM development activities may qualify, others fall under routine implementation. The key to determining eligibility is the technological uncertainty being addressed, not simply the technology being used.
Start with Technological Uncertainty
The Canada Revenue Agency's (CRA) core question is whether, at the start of the work, there was a technological uncertainty that standard practice could not resolve. If a published method or the vendor's documentation already provides the answer, applying it is generally considered routine implementation, regardless of the effort involved. SR&ED starts when known approaches fall short of your requirements and your team has to investigate why.
That investigation must be systematic: define the problem, form a hypothesis, test it, and draw conclusions. A project does not have to fail to qualify. A well-run project that closes a real technical gap can be eligible, and so can one that demonstrates why an approach does not work.
What Usually Does Not Qualify
Not all work involving LLMs meets the CRA's SR&ED eligibility criteria. Activities that rely on established methods without addressing technological uncertainty generally do not qualify.
Examples include:
Calling a model through an API and shipping the output
Refining prompts for tone, style or format
Fine-tuning with a documented recipe that works as expected
Setting up a standard retrieval (RAG) pipeline
Labelling data for normal day-to-day operations
Where Eligible Work Often Hides
LLM development may qualify when existing approaches cannot achieve a required technical outcome, and systematic experimentation is needed to overcome that limitation.
Potential examples include:
Cutting hallucinations below a defined threshold when retrieval and re-ranking alone do not get there
Meeting strict latency, memory or on-device limits without an unacceptable drop in accuracy
Building new evaluation methods because no existing methods measure what you need
Accuracy on messy, industry-specific data that off-the-shelf models and standard fine-tuning cannot reach
Necessary supporting work that directly contributes to eligible experimental development may also qualify.
The bottom line
Not all LLM development qualifies for SR&ED, but projects involving genuine technological uncertainty and systematic experimentation may be eligible.
If your team is going beyond standard implementation to overcome technical challenges, some of your development activities may be worth a closer look.
Talk to R&D Partners
Not sure where your AI work falls? R&D Partners helps Canadian tech companies identify eligible SR&ED activities, prepare documentation that holds up during CRA reviews, and maximize their claims.
Book a meeting with Cameron Cunningham to discuss your LLM projects and explore potential SR&ED opportunities.


