SR&ED for AI companies —
the complete 2026 guide

What SR&ED is — and why most AI companies qualify without knowing it

SR&ED — Scientific Research and Experimental Development — is Canada's largest R&D incentive program. CRA distributes over $4.5 billion annually. For Canadian-controlled private corporations (CCPCs) doing qualifying R&D, the federal government refunds 35 cents of every eligible dollar spent, up to $2.1M per year after the Bill C-15 expansion.

The program has been running since 1985, but the eligibility framing matters more for AI companies than most sectors. The critical test isn't about what you build — it's about whether you faced genuine technological uncertainty in building it. And in AI work, technological uncertainty is almost universal.

Training a model on your domain's sparse data and not knowing if it will converge is technological uncertainty. Trying to make a retrieval-augmented system reliable at sub-100ms latency when existing approaches fail is technological uncertainty. Adapting a known architecture to a new modality where the outcome isn't predictable is technological uncertainty. Building a product using a pre-trained API with no novel technical challenge is not.

The Bill C-15 expansion — what changed in March 2026

The SR&ED program received its biggest overhaul in over a decade when Bill C-15 received Royal Assent on March 26, 2026. The changes apply to tax years beginning after December 16, 2024. Here's what changed:

The gross revenue election is particularly useful for fast-growing AI companies. If your taxable capital has grown rapidly due to investment rounds, you can elect to use prior-year gross revenue for the phase-out calculation instead — whichever gives you the higher limit. This keeps more scaling companies at the full 35% rate for longer.

Eligibility — the three criteria

To claim SR&ED, both your corporation and your work must meet three requirements.

1. Corporation type

The full 35% refundable rate applies to Canadian-controlled private corporations (CCPCs) — incorporated in Canada, controlled by Canadian residents, not publicly listed. Most Canadian AI startups qualify. After C-15, Eligible Canadian Public Corporations (ECPCs) now also access the 35% rate — relevant if you've completed a public raise.

2. Technological uncertainty

Your R&D must have involved advancing knowledge or achieving a technological advancement that wasn't achievable through routine engineering. For AI companies, the clearest qualifying work:

• ✓Novel model architecture or training approach where outcomes were genuinely uncertain
• ✓Adapting existing ML techniques to a domain where their behaviour was unknown and required systematic investigation
• ✓Solving data scarcity, distribution shift, or latency constraints that current techniques couldn't reliably address
• ✓Developing custom training pipelines, evaluation frameworks, or fine-tuning approaches for novel use cases
• ✓Any work where you tried multiple approaches, compared results, and iterated based on technical findings

Work that typically does not qualify:

• ✗Calling an AI API and building a product wrapper without novel technical work
• ✗Standard software engineering — UI development, database design, DevOps infrastructure
• ✗Fine-tuning a model using vendor-provided tools with documented, predictable outcomes
• ✗Market research, business analysis, or routine data collection and labelling

3. Contemporaneous documentation

Your technical work must be documented as it happens — not reconstructed after the fact. This is where the majority of AI companies lose money, and where CRA has focused its scrutiny since 2024.

What costs are eligible and at what rate

Documentation — why so many AI claims get cut

CRA reviewed SR&ED documentation practices in 2024 and 2025 and flagged AI/ML claims as a high-scrutiny category. The industry produces genuinely eligible work at scale, but documents it the way software companies always have — which isn't what CRA wants to see.

The three failures that cut AI claims most often

Retroactive documentation. The project finishes, SR&ED is filed, someone writes up what they worked on. CRA calls this reconstructed documentation and it's the most common audit trigger. The fix is simple but must start before you spend: weekly technical journals, sprint-level records of what was attempted and what was learned, time logs linked to specific technical challenges rather than product features.

Product narrative instead of technical narrative. The T661 Technical Narrative is the document that carries your claim. Most companies write it like a product spec: "We built a document classification system that categorises PDFs." CRA doesn't care what the product does. They need to know: what specific technical obstacle did you face, what approach did you try, what was the outcome, and what did you learn? The framing is always an investigation, never a feature description.

Round-number time allocations. "Developer X spent 50% of their time on SR&ED in Q1." Round numbers are a flag — they signal estimates, not records. Time logs need to be specific, weekly, and tied to named technical challenges.

Real math: what a Toronto AI startup recovers

The key is that IRAP and SR&ED cover different cost buckets. IRAP is paid as a non-repayable grant against labour as you go — better for cash flow. SR&ED then applies its 35% credit to the eligible labour that IRAP didn't cover at year-end. The rule is no double-counting of the same dollar, but the stack itself is entirely legal and the standard approach among companies that know about it.

Ontario adds an 8% refundable credit (OITC) and a 3.5% non-refundable credit (ORDTC) on eligible Ontario R&D spend. These are claimed as additional schedules on your T2 — no separate application beyond what you're already filing for federal SR&ED.

Other programs that stack with SR&ED

SR&ED is almost never the only program a Canadian AI company qualifies for. The most common combinations: