SR&ED AI Companies: What Canadian Founders Should Know

Canadian AI and machine learning companies can claim SR&ED for qualifying R&D work — but only the work that involves genuine technical uncertainty resolved through systematic investigation. Routine API usage and standard fine-tuning generally do not qualify.

What AI/ML work qualifies for SR&ED

AI and ML work qualifies for SR&ED when it involves genuine technical uncertainty that was resolved through systematic investigation. CRA applies the same three criteria to AI companies as to any other claimant.

• Novel model architecture research where the best approach wasn't known in advance
• Training methodology experimentation beyond standard hyperparameter tuning
• Dataset creation and curation for new problem domains where the data requirements weren't established
• Inference optimization that required systematic investigation to achieve performance targets
• Evaluation framework development for novel capabilities where existing benchmarks were inadequate

What AI/ML work does NOT qualify

Not all AI work qualifies for SR&ED. CRA excludes routine development and work that uses standard practice without genuine technical uncertainty. For AI companies, this distinction is particularly important because the field moves fast.

• Routine API calls to OpenAI, Anthropic, or other providers using documented parameters
• Standard fine-tuning of pre-trained models using established techniques and datasets
• UI integration of existing models where the technical approach is well-documented
• Standard MLOps setup using known tools and practices without technical uncertainty
• Data preprocessing and cleaning using established pipelines without novel research
• Model deployment using standard infrastructure and serving frameworks

How AI companies should document SR&ED

AI companies should document their research with the same rigor they apply to their technical work. The documentation should capture: the technical uncertainty that motivated the research, the experiments conducted and their results, the model architecture or training decisions and their rationale, performance benchmarks comparing approaches, and the knowledge gained.