Tax + FundingJune 1, 2026 · 10 min read

SR&ED for AI Projects in 2026: When LIFT-Funded Work Qualifies (and When It Doesn't)

SR&ED is one of Canada's most generous tax incentives — refundable credits, 35% rate at the federal level for CCPCs, expanded ceilings in 2026. The catch: most AI projects an SME would actually fund don't qualify. Here's the honest read on when to ask the SR&ED question, and when to skip it.

If you've started any conversation about funding an AI project in Canada this year, someone has eventually said the magic phrase: "don't forget you can claim SR&ED on that." It's said with the same casual confidence as "don't forget to set aside HST" — as if it were a routine bookkeeping line. For most AI projects, it isn't. Most production AI work isn't SR&ED-eligible at all, and pretending otherwise gets SME owners into preparation costs and audit risk they didn't need.

The 2026 expansion to SR&ED is real, and for the small subset of projects that do qualify, the credits are meaningfully larger than they were last year. So this is a useful conversation to have — but only if we're honest about which projects sit on which side of the line. Below: what SR&ED actually is in 2026, what it takes to qualify, the contractor 80% rule that lets a Creatrixe invoice count, and why LIFT and SR&ED almost never stack on the same dollar of work.

1. What SR&ED actually is

SR&ED — Scientific Research and Experimental Development — is a federal tax-incentive program administered by the Canada Revenue Agency. It is the largest single source of federal R&D support in Canada, with roughly $4 billion in annual support flowing to claimants. The structure has two parts:

• An investment tax credit (ITC) applied to qualified SR&ED expenditures — labour, contractor payments, materials consumed in the R&D, and overhead. Canadian-controlled private corporations (CCPCs) get a 35% federal rate on the first portion of qualified expenditures, with provincial top-ups on top (BC, for example, adds 10%). Other corporations get 15% federally.
• Partial refundability for CCPCs — meaning the ITC is paid out in cash even if the corporation has no tax payable. This is what makes SR&ED the closest thing Canada has to a true cash grant for genuine R&D.

The claim is filed via form T661 alongside the corporate tax return (T2), typically within 18 months of the relevant tax year. It is not a grant application — you don't apply for it in advance. You spend the money, complete the work, and claim it after the fact when you file taxes. CRA reviews a subset of claims; the review process can be quick (a desk review) or extended (a technical review with a CRA science advisor).

2. The 2026 expansion — bigger ceiling, more refund

On March 26, 2026, Bill C-4 received Royal Assent, formalising the SR&ED expansion that had been promised in the 2024 Fall Economic Statement and clarified through the 2025 budget cycle. Two changes matter for SMEs:

The headline is the doubled enhanced expenditure limit: a CCPC can now claim up to $6M in qualified SR&ED expenditure at the enhanced 35% refundable rate, for a maximum refundable federal ITC of $2.1M in a single tax year. Add provincial credits and you can comfortably reach $2.5M+ in a strong year for a growing R&D-heavy CCPC. That's real money — for the right project, in the right corporate structure.

3. What qualifies as SR&ED — the three-part test

This is the part that decides most AI projects out. SR&ED isn't "we built something technical." It's a specific test that has been refined through three decades of CRA jurisprudence. There are three pillars, and all three must be present for work to qualify:

1. Technological uncertainty. The work begins with a question that cannot be resolved through routine application of existing knowledge. A competent professional in the field, given the available information, would not know the answer.
2. Systematic investigation. The work follows a structured method — hypothesis, experiment, observation, conclusion — documented at the time it happens. Not after-the-fact storytelling.
3. Technological advancement. The work either produces new knowledge or extends existing knowledge in a way that contributes to the broader technical field, not just to the claimant's commercial product.

Most production AI work fails on pillar one. If you are deploying an existing model — Claude, GPT, Gemini, Whisper, an off-the-shelf computer-vision model — to a known use case, you are doing routine application of existing knowledge. The vendor has solved the technological uncertainty for you. You are integrating, configuring, prompting, evaluating, and shipping. That is engineering work; it is not experimental development in the SR&ED sense.

This is the gap that catches SME owners off-guard. The work feels novel — it's new to your business, it uses cutting-edge models, it required real technical skill to ship. But "new to your business" is not the SR&ED test. The test is "new to the field," or at least "the technical answer was not obvious to a competent professional with access to the literature." Deploying a known LLM to a known workflow does not meet that bar, no matter how much craft went into the deployment.

What does meet the bar:

• Genuine fine-tuning where the base model behaviour is materially different from off-the-shelf, and the path to get there required hypothesis-testing rather than recipe-following.
• Novel model architecture or training procedures where the literature didn't already supply the answer.
• Solving a technical problem that other people in the field have tried and failed at, or that the literature treats as open.
• Cross-disciplinary work — physics-informed ML, novel domain adaptation, low-resource language work — where the existing knowledge wasn't sufficient.

Routine prompt engineering, agent orchestration with off-the-shelf frameworks, RAG over your documents, deploying a known TTS model into a phone workflow, integrating Claude into a CRM — none of that qualifies, regardless of how well it was done.

4. The contractor 80% rule

For the subset of projects that do qualify, there's an important wrinkle for SMEs that hire an integrator (like Creatrixe) to do the work: the 80% contractor rule.

When an SME hires an arm's-length Canadian contractor to perform SR&ED work on its behalf, 80% of the contract payment is treated as qualified SR&ED expenditure on the SME's claim. The remaining 20% is treated as a non-eligible margin. The contractor's labour, materials, and overhead are bundled into that 80% figure — the SME doesn't itemise the contractor's internals.

The practical effect: if you hire Creatrixe to do $200K of work, and the work genuinely qualifies as SR&ED, $160K becomes qualified expenditure on your T661. At a 35% federal CCPC rate that's $56K of federal ITC, plus provincial top-up. The contractor invoice can count — but only if the underlying work is genuinely SR&ED, not routine deployment dressed up in technical language.

5. When LIFT and SR&ED stack — almost never

The most common question we get from SMEs reading about both programs: can the same project be funded by BDC LIFT and claimed for SR&ED at year-end?

The honest answer is: almost never on the same project, and rarely on the same SME at all.

The reason is structural. BDC LIFT is designed to fund adoption — the deployment of mature, known AI technology into Canadian SME operations. The whole product thesis is "off the sidelines": take what already exists in the market and put it to work in your business. The projects LIFT funds are, by design, the projects SR&ED disqualifies. A trades business automating dispatch with a known LLM and a known CRM is a textbook LIFT-eligible project — and it is precisely the kind of project that fails the SR&ED technological-uncertainty test.

The same applies in reverse. The narrow set of projects that do qualify for SR&ED — genuine research, novel architectures, unresolved technical problems — are usually pre-revenue or early-stage, run by R&D-heavy teams that BDC LIFT was not designed to underwrite. They are also a much better fit for NRC IRAP, which is structured specifically to fund early-stage Canadian R&D with the kind of advisor and program-management overhead that genuine research benefits from.

So in practice: the same SME might take LIFT for the production adoption project and run SR&ED on an entirely separate genuine-research workstream (perhaps an internal R&D group experimenting with novel approaches on the side). What they almost never do is take LIFT for project X and claim SR&ED on project X. The two programs are looking for different shapes of work.

6. A concrete example of qualifying SR&ED

Here's a worked example, fictional but based on real conversations we've had.

A Surrey HVAC business operates across the South Asian and Middle Eastern diaspora communities of the Lower Mainland. Their customer base regularly code-switches between English, Punjabi, and Arabic, sometimes mid-call. Existing off-the-shelf voice models handle single-language calls well; they fail on cross-lingual code-switching, especially in the Punjabi-Arabic-English triad common to this customer base. There's a real technical question: can we adapt a base voice-recognition stack to handle code-switching reliably in this specific linguistic blend, where the existing literature doesn't supply a clear answer?

If the project is structured as genuine investigation — hypothesis (which adaptation method works on this corpus), systematic experiment (controlled training runs against a held-out evaluation set), observation, conclusion — and if the resulting work materially extends what off-the-shelf models could do, that might qualify as SR&ED. The contractor performing the work (us) would charge a normal engagement fee; the SME would file a T661 claiming 80% of that fee plus any of their own qualifying internal labour.

Contrast this with a routine English-only AI receptionist deployment for the same SME. Same business, same vendor (us), same form factor. The English-only deployment is good engineering work using mature models; it doesn't qualify for SR&ED. The code-switching adaptation might. The difference is not the form of the deliverable — it's the presence or absence of genuine technological uncertainty that required investigation rather than configuration to resolve.

7. What we don't do

One important honest note: Creatrixe is not an SR&ED claim-preparation firm. We will not file your T661 for you. We will not represent you in a CRA science review. We are a technical AI consultancy that builds the systems; the SR&ED claim is filed by your accountants, ideally in coordination with a specialist SR&ED firm (in BC, names that come up regularly include MNP, BDO, Boast, and a handful of independent consultants — we have no commercial relationship with any of them).

What we will do, when the project genuinely sits on the qualifying side of the line: keep contemporaneous technical documentation as we work, because the documentation is what the science review actually reads. After-the-fact reconstruction of "what we tried and why" is much weaker than a real engineering journal kept during the work. If you intend to file SR&ED on a project we deliver, tell us at the start so we keep the right artefacts as we go.

8. The honest closing

SR&ED is excellent — for the projects it was designed for. The 2026 expansion to a $6M enhanced limit and $2.1M maximum refundable federal ITC is a real improvement, and for Canadian SMEs running genuine R&D it is one of the most valuable tax instruments in the federal toolbox. The official program landing page at canada.ca/sr-ed is the canonical source for current rules.

But SR&ED is not a default funding mechanism for AI adoption. Most AI projects an SME would actually fund are adoption, not research — and adoption is what LIFT and CDAP-style programs exist for, not what SR&ED exists for. Pretending otherwise creates real risk: claim-preparation fees on a claim that won't survive review, an unhappy CRA file, and a hit to the SME's standing for future legitimate claims.

The honest rule we use internally: don't even start the SR&ED conversation unless the project has a real technical question that the public literature does not already answer. If the project is "we want to deploy a known model into a known workflow as quickly as possible," it isn't SR&ED — and it doesn't need to be. LIFT funds that shape of work at a 2.25% rate with a Canadian integrator, and the math is fine without ever opening a T661.

If the project genuinely sits on the research side of the line — and a handful of the projects we see each year do — then SR&ED is real and worth the work. Just engage a specialist on the claim, keep contemporaneous documentation, and don't bundle adoption work into the same envelope. The line matters.

About this post

Creatrixe is a Canadian AI consultancy based in Burnaby, BC. We build production AI workflows for SMEs and named-partner with SR&ED specialists when the underlying work genuinely qualifies. We are independent of CRA and earn nothing from claim-prep referrals. Program details accurate to the official CRA SR&ED page as of publication; specific case treatment varies and this post is not tax advice.