SR&ED for AI & machine learning
An AI label tells a reviewer almost nothing about eligibility.
The useful question is what the team could not achieve with the available models, methods or data, and what it learned by testing alternatives.
Using an API or following a documented fine-tuning recipe is generally routine. Review any separate investigation in which the baseline failed and the team used experiments or analysis to pursue a technological advance.
Start where the baseline stopped working
- Architecture or training-method experiments undertaken after standard approaches failed under the actual conditions.
- Work on accuracy, latency or resource constraints that known techniques could not meet for the defined problem.
- Tests addressing scarce labels, noisy inputs or distribution shift where the result could not be predicted from available knowledge.
- Competing hypotheses evaluated through controlled runs, including results that ruled an approach out.
Routine AI work stays routine
Model integration, standard data labelling, pipeline plumbing, ordinary prompt iteration and fine-tuning by an established method generally fall outside SR&ED. A new commercial use does not change that. The technical record needs to show an advancement sought and a systematic investigation, not simply a capable product.
Treat the training record as a project diary
Experiment logs, evaluation sets, notebooks, model versions, ablation results and the commits behind them can reconstruct the sequence from baseline to conclusion. Keep the metric definition and test conditions with each result; a chart without its setup is easy to misread.
SREDlog organizes approved records by project and prepares an editable narrative draft around the hypotheses and tests they contain. The preparer decides what the evidence supports.
Frequently asked questions
No. Review the specific investigation separately from the interface, deployment, integration and other product work around it. The project still has to meet both CRA requirements.
Fine-tuning by an established method is generally routine. Related work may qualify only when the facts show a technological advancement purpose and a systematic investigation or search by experiment or analysis.
Keep the hypothesis, dataset and evaluation definition, configuration, result and conclusion for each material run. Model history, ablations and source changes help connect the sequence. See evidence management.
Keep reading
Put the failed baselines beside the final model
Start a free trial and organize one project's training history, evaluations and source records for review.
7-day free trial · No credit card required
These industry examples are prompts for a fact-specific review, not categories guaranteed to qualify. Work must be conducted in Canada and meet both current CRA requirements; support work must directly support and be commensurate with eligible work. This is general information, not tax advice, and has not been reviewed by an independent qualified tax professional. See our editorial policy.