Key takeaway: SR&ED for edtech startups runs into an exclusion most software companies never have to think about: social science or humanities research. A learning platform can spend a whole year testing pedagogical theories, running efficacy studies, and refining curriculum sequencing — genuinely hard work, and almost none of it SR&ED. What does qualify is the computer science underneath: an adaptive algorithm that has to personalize content from noisy, sparse student data; an NLP model grading free-text answers without a rubric it can pattern-match against; a recommendation engine sequencing content when the "right next lesson" isn't a solved problem. Edtech sits in the same software category that already earns 42.6% of all credits allowed nationally — but only for the engineering half of the business.
Ask an edtech founder what their R&D was last year and you'll usually hear about the platform: which instructional design changed, what the pilot with a school board showed, how retention improved after a curriculum revamp. All of that might be true and valuable. Almost none of it is what the CRA means by scientific or technological research, and that gap trips up more edtech claims than any dollar limit or filing deadline ever will.
The reason is structural, not a matter of the CRA disliking education. Two of its named exclusions — social science or humanities research, and market research or sales promotion — sit almost directly on top of what an edtech company spends most of its year doing. Below is where the line actually falls, which parts of an adaptive-learning or assessment stack clear it, and where teams lose the evidence for work that genuinely did.
Does SR&ED cover education technology
Yes, for the computer science and engineering inside an edtech product. Not for pedagogical or curriculum research. The CRA applies the same two-part eligibility test to every industry: was there a technological uncertainty that standard practice couldn't resolve, and did the company try to resolve it through systematic investigation rather than trial and error until something worked well enough.
Computer programming, mathematical analysis, engineering, and testing all appear among the CRA's eight recognized categories of support work, and none of them is scoped by industry. A natural-language model that grades open-ended student responses gets evaluated the same way a natural-language model for legal document review would. What makes edtech different is which exclusion tends to catch the work instead — and it's usually not the one software founders expect.
The exclusion edtech runs into that software rarely does
Social science or humanities research is explicitly excluded from SR&ED, and it's the exclusion doing the most damage to edtech claims. A learning-efficacy study — does this instructional approach improve test scores more than that one — is education research. It might be rigorous, peer-reviewable, genuinely valuable science. It's also not the kind of science this program funds, because the CRA's definition of scientific research is scoped to the natural and applied sciences, and pedagogy sits outside that scope regardless of how systematic the methodology is.
Market research or sales promotion does similar damage from a different angle. A pilot deployment with a school board, run mainly to generate outcome data for a sales deck or a case study, isn't SR&ED just because the underlying software is complex. The tell is intent: was the pilot designed to answer an open technical question, or to prove the product works well enough to close the next contract? Plenty of edtech pilots do both at once, and the claim has to separate the two honestly.
Neither exclusion is edtech-specific guidance — there's no education-sector bulletin, and the same rules rule out prospecting, style changes, and routine data collection in every other industry. Edtech just happens to generate an unusual amount of activity that looks exactly like these two exclusions from the outside.
Where the real technical uncertainty shows up
Real uncertainty in edtech clusters around problems where the input data is messy, sparse, or genuinely ambiguous, and where no off-the-shelf method reliably solves it. A few shapes recur.
Adaptive learning algorithms are the clearest case. Personalizing a learning path from a student's incomplete, noisy interaction history — a handful of quiz answers, some skipped exercises, inconsistent time-on-task — is a different problem than recommending a movie from a dense ratings matrix, and most published recommendation techniques don't transfer cleanly.
Automated assessment of open-ended work sits right next to it. Grading free-text answers, code submissions, or essays without a fixed answer key means building a model that has to judge quality and correctness under genuine ambiguity, not pattern-match against a known set of right answers. Learning analytics raises a related but distinct question: predicting which students are at risk of disengaging or failing from behavioral signals, where the relationship between those signals and the outcome isn't established for your population or your platform.
Speech and language technology for accessibility deserves its own mention, because it fails in ways general-purpose models weren't built to handle. Speech-to-text or text-to-speech tuned for children's speech patterns, non-native speakers, or students with specific learning disabilities routinely breaks assumptions baked into off-the-shelf models, and getting it to work reliably is a real engineering problem — not a configuration tweak.
Knowledge-graph or content-sequencing engines look deceptively simple from the outside. Figuring out prerequisite relationships between concepts and ordering material accordingly, when the "correct" sequence isn't documented anywhere, sits closer to information retrieval and graph algorithms than to curriculum design, even though the output resembles a curriculum.
One more shape has gotten more common with every model release since large language models became a classroom fixture: detecting AI-generated or plagiarized student work. A detector trained on last year's writing patterns tends to degrade as the underlying generation models change, and distinguishing a student's own edited draft from AI-assisted or AI-authored text without a reliable ground-truth label is an open problem, not one you can license off the shelf. Building and validating a detection approach against your own institution's writing samples, where published detectors demonstrably fail, is exactly the kind of work the two-part test was written for — provided the team can show what it tried, what failed, and why the failure wasn't predictable in advance.
Two edtech companies, two different claims
Two edtech companies can both call last year "heavy on R&D" and mean two entirely different things, and the difference usually comes down to what the uncertainty actually was.
Company A builds a reading platform on a well-established adaptive algorithm licensed from a research paper, and spends the year running the platform in twelve classrooms, adjusting the UI based on teacher feedback, and comparing reading-score gains against a control group to prove the product works. That comparison study is well-designed and the results are real. But the adaptive technology itself works as published, the classroom deployments were run to generate proof points for sales and renewal conversations, and the uncertainty being resolved is pedagogical effectiveness, not a technological unknown. Good marketing collateral, thin SR&ED claim.
Company B is building an automated grading model for open-ended math proofs, where a correct answer can be expressed a dozen structurally different ways and no existing model reliably distinguishes a valid alternative approach from a wrong one. The team tries several representations of the problem, documents where each fails against real student submissions, and only converges on an approach after multiple rounds of structured experimentation against a technical hypothesis. Company B's claim holds up because the uncertainty is squarely computational — it would look the same whether the subject were math, chemistry, or a foreign language.
Most edtech companies run both kinds of work in the same year, sometimes on the same feature. The claim has to draw that line project by project rather than treating a busy year as one undifferentiated research program.
Contractors, pilots, and the 80% rule
Only 80% of what an edtech company pays an arm's-length contractor for SR&ED work enters the qualified expenditure pool. That's the same rule, in effect since 2012, that applies to any software claim, and it covers a contracted ML engineer building a grading model exactly as it would a contracted backend developer.
What it doesn't cover, at any percentage, is a payment to a school, school board, or research partner to run a pilot or efficacy study — not because of the contractor discount, but because that work usually isn't SR&ED to begin with, for the reasons above. If a contract genuinely requires the partner to help resolve a technical unknown — say, a university lab helping validate a novel assessment model's scoring against expert human graders — get that scope written into the agreement itself. A reviewer looks at what was actually commissioned, and "run our program in your school for a semester" reads very differently from "help us determine whether this scoring model agrees with expert graders."
Salaries, the proxy method, and where capital equipment fits
Almost all of an edtech SR&ED claim runs through salaries, with the prescribed proxy amount adding a 55%-of-salary overhead allowance instead of requiring receipts for every server and subscription. That's simpler for edtech than for a hardware-heavy sector, since most platforms don't own significant equipment.
The exception is accessibility and assistive-technology hardware — a custom sensor rig for an alternative-input device, or specialized recording equipment for building a speech dataset. Eligible capital property acquired after December 15, 2024 is claimable again, at 40% refundable on the 35% rate, for equipment bought specifically to run experimental work rather than to operate an already-validated product commercially.
What edtech companies get wrong about documentation
The most common failure isn't claiming a curriculum study as SR&ED outright — most teams sense that a reading-comprehension trial isn't computer science. It's under-documenting the technical work that did qualify, because the loudest artifacts of an edtech company's year are pedagogical: lesson plans, pilot reports, teacher feedback forms. The commit history telling you the grading model went through four scoring representations before one held up rarely gets narrated anywhere a reviewer would look.
A ticket that says "grading model updated after pilot feedback" doesn't say what the team hypothesized was wrong with the previous scoring approach, which alternatives they tried, or why the fix wasn't obvious from published grading-model literature. The CRA's own position is that a project doesn't need to succeed to qualify — an adaptive-sequencing approach the team built, tested against real student data, and abandoned for a better one is evidence, as long as someone recorded the hypothesis and the failure, not just that a pilot ran.
There's a second, quieter version of the same problem: engineers on an edtech team often don't think of their own work as R&D, because the company's public story is about students and teachers, not algorithms. A developer who spent three weeks fighting a scoring model's false-positive rate on edge-case student answers may never mention it in a standup written for a room full of curriculum people. That work existed. It just never made it into a form a claim can point to a year and a half later, which is a documentation gap, not an eligibility one.
Who should think twice before claiming
Skip this, or claim narrowly, if most of the year's work was curriculum design, instructional content creation, efficacy studies, or pilots run to prove the product to a buyer rather than to resolve an open technical question. A semester-long study showing students learned more with your app is a strong case study. On its own, it isn't R&D.
The mirror mistake shows up just as often. Edtech founders who assume "we're basically a content company" write off engineering work that clears the bar easily, because the team building it never framed it as research to begin with. If a real computational or algorithmic problem sits under the platform, it's worth a proper look regardless of how education-flavored the rest of the business feels.
What SR&ED for edtech startups is actually worth
The rates don't change because the product ships to classrooms instead of enterprises. A Canadian-controlled private corporation earns the enhanced 35% refundable rate on up to $6 million of qualifying expenditures a year, for tax years beginning after December 15, 2024, with the basic 15% rate applying above that. At the full limit that's up to $2.1 million a year at the enhanced rate. Current expenditures like salaries at the 35% rate are 100% refundable for most CCPCs, though excluded corporations receive 40% instead.
Provincial credits stack on top, but not additively. A company also claiming a provincial R&D credit sees its federal base reduced under the s.127(18) grind rule — in a simplified illustrative example assuming a full grind on the same base, an Ontario edtech company's combined rate lands closer to 40.2%, not a naive 43%, though the exact number moves with expenditure mix and proxy treatment. Equivalent non-additive math applies everywhere else. The filing clock is the same as any other sector: a corporation's SR&ED reporting deadline is 18 months after its fiscal year end, with no extension process available regardless of how the school calendar or a fundraise cycle happens to line up.
That's also where continuous documentation earns its keep, and it doesn't care whether the codebase grades essays or processes payments: evidence captured from the tools a team already uses — commits, tickets, model experiment logs — reviewed by a qualified independent SR&ED expert before anything is filed. Automation on our side handles the collecting; a named human still stands behind what gets submitted. It doesn't decide whether a grading model or a sequencing engine clears the bar — that judgment stays with the two-part test — but it means the reasoning behind an abandoned model version is still there next year, instead of living only in a Slack thread nobody can find.
Frequently asked questions
Is pedagogical or curriculum research eligible for SR&ED? No. The CRA excludes social science or humanities research from SR&ED, and studying which instructional approach improves learning outcomes falls in that category, however rigorous the methodology.
Can an adaptive learning algorithm qualify for SR&ED? Yes, if personalizing content from a student's sparse or noisy interaction data requires solving a problem existing techniques don't reliably handle. An algorithm licensed or adapted from published, working methods is applying existing technology, not advancing it.
Does running a pilot program with a school or school board count as R&D? Only if the pilot is designed to test a genuine technical hypothesis. A pilot run mainly to generate efficacy data for sales or renewal conversations is market research, which the CRA excludes regardless of how much engineering sits underneath the platform.
Does outsourcing a pilot or efficacy study to a research partner change the claim? Usually the underlying study still isn't SR&ED, so the 80% contractor rule is moot. It only applies when the contracted work genuinely resolves a technical uncertainty — for example, validating a scoring model's output against expert human graders.
Can automated grading or assessment software qualify? It can, when grading open-ended, unstructured responses requires a model that handles genuine ambiguity rather than matching against a fixed answer key. Grading multiple-choice or exact-match responses is applying well-established techniques, not resolving a technical unknown.
Does capital equipment like assistive-technology hardware qualify? It can, if acquired after December 15, 2024 for experimental work rather than to operate an already-validated product, per the CRA's restored capital-expenditure eligibility. Capital expenditures at the 35% rate are 40% refundable.
Education doesn't get its own SR&ED rulebook. It gets the same two-part test as everything else, applied honestly to the fraction of an edtech company's year that's genuinely computer science — and not to the fraction that's research about how people learn.
See what your edtech engineering work could be worth — estimate your refund or check your eligibility.