Introduction
The SR&ED tax credit program has always rewarded Canadian companies that advance science and technology — not routine development, not incremental upgrades, but genuine technological uncertainty resolved through systematic investigation. The rapid expansion of AI as a development tool, and AI as the subject of development, creates new questions about what qualifies — and some of those questions do not yet have settled answers from the CRA.
The SR&ED Test Has Not Changed
The eligibility criteria for SR&ED have not been updated to address AI specifically. The three-part test remains:
1. Technological advancement: The work must attempt to achieve a scientific or technological advancement — to generate new knowledge or to achieve something that was not previously achievable with existing knowledge.
2. Technological uncertainty: There must be genuine uncertainty that cannot be resolved by standard practice or by the application of known techniques. If the outcome of the work was predictable from existing knowledge, it does not qualify.
3. Systematic investigation: The work must be conducted through hypothesis, experiment, and analysis — the scientific method applied to technological problems. Ad hoc tinkering does not qualify.
These criteria apply equally to AI projects as to any other development work.
When AI Development Qualifies for SR&ED
Training novel architectures: A company developing a new neural network architecture — designing the model structure, training procedures, and loss functions — and doing so with genuine uncertainty about whether the architecture will perform as hoped is likely performing qualifying SR&ED. The work is attempting to advance the state of neural architecture design.
Domain-specific model adaptation with genuine uncertainty: A company attempting to fine-tune or adapt a large language model for a specific domain where standard fine-tuning methods produce inadequate results — and where resolving that inadequacy requires investigating new techniques — may be performing qualifying SR&ED. The key is genuine uncertainty, not the fact of using AI.
Novel AI-hardware integration: Developing software that exploits specific hardware characteristics in ways that require investigating undocumented or poorly understood hardware behaviour is a traditional area of SR&ED eligibility that extends naturally to AI accelerator hardware.
When AI Use Does Not Qualify
Using a pre-existing AI tool to build a product: A company that uses a commercial LLM API (GPT-4, Claude, Gemini) to build a product — even a sophisticated one — is not performing SR&ED. The technological uncertainty lies in the model, which was developed by the API provider. The integrator is applying a known tool, not advancing technology.
Standard machine learning applications: Applying well-established ML techniques — logistic regression, random forests, convolutional neural networks on image data — to a new dataset or business problem is not SR&ED. The techniques are known; the application is routine engineering.
Prompt engineering: Improving outputs from a pre-existing model through prompt design or fine-tuning is not SR&ED. It is an application layer activity.
The CRA's Evolving Position
The CRA has acknowledged that AI-related SR&ED claims are a growing area and has indicated that existing eligibility criteria apply — but has not issued specific administrative guidance for AI. The area where the most questions arise is foundation model fine-tuning for domain-specific applications, where the line between "applying known techniques" and "investigating new methods to resolve genuine uncertainty" is genuinely blurry.
The NRC's Industrial Research Assistance Program (IRAP), which works alongside SR&ED, has also expanded its AI advisory capacity — and IRAP consultations can provide informal feedback on eligibility before a formal SR&ED claim is prepared.
The Documentation Requirement for AI SR&ED Claims
SR&ED claims require contemporaneous documentation of the technological uncertainty, the experimental approach, and the results. For AI projects, this documentation needs to clearly articulate:
What was known before the project began (existing AI techniques)
What was unknown (the specific uncertainty the project addressed)
What experiments were run and what they showed
What advancement or knowledge was produced
AI development teams that do not maintain project documentation in terms of technological uncertainty and experiment results — common in fast-moving product teams — face challenges in preparing SR&ED claims that survive review. The documentation habit must exist during the project, not be reconstructed after.
When to Speak With a CPA
For tech companies spending significant resources on AI development — and uncertain about SR&ED eligibility — a preliminary review with a CPA experienced in SR&ED is the right first step. The 18-month claim window applies regardless of when the eligibility question is resolved.