AI Projects and SR&ED: What Changed, What Didn't
AI and ML work can qualify for SR&ED — but not all of it. The CRA's eligibility test hasn't changed. Here's what that means for model training, fine-tuning, and architecture decisions.
The question that sounds simple but isn't
AI founders have it especially complicated. Their work looks obviously innovative to everyone around them — neural networks, transformers, model architectures — but 'looks like research' and 'is SR&ED-eligible' are different things. And the reverse is also true: some genuinely qualifying ML work doesn't look exotic at all.
The CRA hasn't created a special AI eligibility framework. The two-part test is unchanged: technological advancement plus systematic investigation. Using a transformer doesn't make work SR&ED. Neither does fine-tuning. The question is whether you faced a technical obstacle that publicly available knowledge couldn't resolve — and whether you investigated it systematically.
The CRA evaluates whether the work achieved technological advancement beyond the state of publicly documented knowledge, and whether it was carried out through systematic investigation. The presence of AI or ML in a project is not itself the test.
Where AI and ML work commonly qualifies
- Novel model architecture work for a domain where existing architectures demonstrably failed on your specific data characteristics and constraints
- Investigating training approaches for sparse, noisy, or domain-specific datasets where standard fine-tuning produced insufficient results and no documented solution existed
- Resolving trade-offs between model accuracy, latency, and resource consumption for specific deployment constraints using systematic experimentation
- Building proprietary data preprocessing or evaluation methods for novel domains where existing methods weren't applicable
A legal tech company develops a document analysis tool. They use GPT-4 via API for initial extraction — not SR&ED. But they discover that no standard benchmark exists for evaluating legal reasoning accuracy in their jurisdiction and document type. They design a novel evaluation framework, run 200+ annotated test cases, and discover that their prompting approach fails on a specific clause type. The evaluation framework investigation qualifies. The API usage doesn't.
Where it doesn't qualify — even if it sounds technical
- Calling a third-party API (GPT-4, Claude, Llama via API) to build a product feature — this is procurement, not technical investigation
- Standard fine-tuning of a pre-trained model using documented techniques for a well-covered domain
- Prompt engineering, even sophisticated chain-of-thought approaches — currently not SR&ED under CRA guidance
- Benchmarking existing models against each other for vendor selection
- Building a RAG pipeline following published RAG architecture for a standard document retrieval use case
Training runs generate lots of data — loss curves, accuracy charts, parameter logs. But a training log by itself doesn't show that investigation was systematic. CRA reviewers need to understand why each experiment was chosen, what the hypothesis was, and what the finding meant. The documentation challenge specific to ML is adding narrative to the numbers.
The lab notebook problem in ML
Traditional scientific investigation has a tool for this: the lab notebook. It records not just what happened, but why you tried it, what you expected, and what you learned. ML teams almost never keep one — which means their evidence base is often technically rich and narratively empty.
Useful SR&ED evidence for ML projects: experiment configs with written rationale for parameter choices, hypothesis-to-experiment-to-finding records, architecture decision rationale explaining what was tried and why alternatives were rejected, meaningful records of what failed and why.
For each ML project you think might qualify: what technical obstacle did you face that couldn't be solved by following existing documentation? If you can point to a paper, library, or API that directly addressed the problem, it likely doesn't qualify. If the path required genuine experimentation and produced unexpected findings — document those carefully, and consider whether a SR&ED specialist should review before filing.
This guide is for general information only. SR&ED eligibility for specific AI/ML projects depends on the facts of each project and is determined by the CRA. All claims should be reviewed by a qualified CPA or SR&ED preparer before filing.
Turn this guide into a claim package
SREDY.IO walks you through eligibility, project narratives, supporting costs, and evidence so your CPA has a cleaner file to review.
More guides
What Actually Makes Software Work SR&ED-Eligible?
Most software companies doing real R&D leave SR&ED credits on the table because they don't recognize their own qualifying work. Here's the eligibility distinction that matters.
T661 WritingHow to Talk About Uncertainty Without Overclaiming
The language you use in a T661 narrative matters. Overstating uncertainty reads as speculative. Understating it undersells the challenge. Here's how to write about technical uncertainty in CRA-ready language.
DocumentationThe Evidence Habit: What to Save Before Year-End
SR&ED claims are significantly easier to defend when you've been saving the right artefacts throughout the year. A practical pre-year-end checklist for Canadian tech companies.
Ready to check if your work qualifies? Take the free eligibility assessment.
Check eligibility — free →