Key takeaway: AI SR&ED claims don't get flagged, reduced, or denied because a tool touched them. 90% of claims filed in FY2025-26 were accepted as filed and only 4% were denied, and that split is decided by whether the evidence behind a claim is real, not by who or what drafted the narrative. The CRA's eligibility test hasn't changed for a single word of Bill C-15: your work still has to clear the same two-part bar it always has, and your claim still has to survive a reviewer asking you to prove it. Whether a person or a model drafted the narrative decides nothing on its own. Who stands behind it when the CRA asks questions decides almost everything.
Every founder using AI to help with SR&ED eventually asks some version of the same question: is this going to get my claim flagged? Wrong question. AI SR&ED claims don't fail because a model wrote the words — they fail for the same reason claims always have, AI or not: nobody can back up what's on the form. Here's what the CRA actually evaluates, why the "did AI write this" framing misses it, and what a claim needs to survive a review regardless of what tool touched it.
What actually makes AI SR&ED claims risky
An AI-assisted SR&ED claim is risky for exactly one reason: if the narrative doesn't trace back to real evidence, generated by nothing more sophisticated than a founder guessing what a good SR&ED sentence sounds like. That's true whether the founder wrote it themselves at midnight before the deadline or an AI tool drafted it from a two-line prompt. The CRA doesn't have a rule against automation. It has a standing requirement that every claim be complete, accurate, and supported by evidence, stated the same way regardless of who or what drafted the document.
Put a fabricated technical narrative next to a well-documented one, and the fabricated one is the risk — not because of its authorship, but because a reviewer asking follow-up questions will find nothing behind it. A narrative with no dead ends, no failed approaches, no specific technical uncertainty reads as invented whether a person wrote it from memory eighteen months later or a language model wrote it from a vague prompt with no real project data behind it. The tool isn't the variable. The evidence is.
That reframing matters because it points you at the actual decision you need to make. The question isn't "should I let AI touch my SR&ED claim." It's "does whatever produced this narrative have real project evidence to draw from, and is someone qualified checking the result before it's filed." Answer those two questions honestly and the authorship question mostly answers itself.
The two-part test your SR&ED claim has to pass, however it's written
Your SR&ED claim has to clear the same two-part eligibility test no matter how the narrative got onto the page: a "why" and a "how." Per the CRA's eligibility guidance, your work has to be undertaken for the advancement of scientific knowledge or technological advancement, in the face of real uncertainty about how to achieve it — that's the "why." And it has to be carried out as a systematic investigation, by experiment or analysis — that's the "how." If your team's approach fails and you learn something from it, the project can still qualify; the test is about the investigation, not the outcome.
Nothing about that test changed with Bill C-15, and nothing about it mentions how you drafted the claim. A CRA Research and Technology Advisor reading your narrative isn't scoring it on prose style or checking a "written by a human" box. They're checking whether the story matches a real technological uncertainty your team actually worked through, using one of the eight recognized support-work categories — engineering, design, computer programming, testing, and the rest.
That's why the "capture vs. generation" framing you'll see elsewhere asks a narrower question than the one that decides outcomes. Is the AI just capturing what happened, or generating the argument for eligibility? Capture and generation are both just steps in producing a document. A narrative can be entirely "captured" from real commits and still be assembled by a person too rushed to connect it to the actual technical uncertainty. A narrative can be substantially "generated" by a model and still be accurate, specific, and defensible, if it was built from the actual project record and checked by someone who understands the eligibility test. The axis that predicts what happens in a review isn't capture-versus-generation. It's evidence-versus-invention, checked by someone qualified before it's filed.
Where AI genuinely helps, and where the real risk lives
AI genuinely helps with the mechanical, high-volume parts of a SR&ED claim: surfacing eligible work from your dev tools, organizing technical details as they happen, and drafting a first-pass narrative from real project data instead of a blank page. None of that is where the risk lives. A model that reads commit history, ticket threads, and pull requests to identify where your team hit a real technical wall is doing pattern-matching on real evidence, which is a genuinely useful, mechanical task. It is not the same task as judging whether that evidence clears the eligibility bar — that call still needs a qualified expert, which is what the next paragraph is about.
The risk shows up at the handoff: the moment a draft becomes the filed claim without anyone who understands SR&ED eligibility checking whether the story actually holds up. That's true of AI-assisted drafts, and it was true of human-assisted drafts before AI existed — a consultant template filled in by a junior associate who's never seen the codebase has the exact same failure mode. Automation didn't invent the risk of an unreviewed claim. It just made unreviewed claims faster and cheaper to produce, which is precisely why review matters more now, not less.
So the honest version of "is AI risky for SR&ED" is: automation without expert review is riskier than it used to be, because it's now trivially easy to generate a claim that reads fine on the surface and has nothing real underneath it. Automation with expert review in front of every claim isn't riskier at all — it's the same claim you'd get from a careful human writer, produced faster and pulled from real records instead of memory.
What a defensible AI-assisted claim actually looks like
A defensible AI-assisted SR&ED claim has three things a shaky one lacks: real project evidence behind every sentence, a narrative that names the actual technical uncertainty and the actual dead ends, and a qualified person who checked it before it went in. None of the three depends on how much of the drafting was automated.
Real project evidence means the narrative can point to something — a commit, a ticket, a design doc, a Slack thread — for every claim it makes about what the team tried and why it didn't work the first time. A narrative built this way holds up under questions because the questions have real answers behind them.
Naming the actual technical uncertainty means the claim doesn't read like a template with your company's name swapped in. "We built a scalable, high-performance system" tells a reviewer nothing. "We didn't know whether our indexing approach would hold sub-second query times past 50 million rows, tried three schema designs, and the one that worked required a rewrite of the ingestion pipeline" tells them exactly what T661 is asking for. AI can help draft that second version — but only if it's working from the real project data, not a prompt with no specifics in it.
And expert review before filing is the part that has nothing to do with the tool at all. A qualified, independent SR&ED expert reading the narrative before it's filed catches the same problems whether the first draft came from a person or a model: eligibility calls that are borderline, claims that overstate the technical uncertainty, expenditures that don't tie back cleanly to the projects described. That check is what turns a fast draft into a filed claim someone will actually defend if the CRA reviews it.
That's the model Glauq is built on. Automation does the capture, pulling technical evidence continuously from the tools your team already works in, so the narrative comes from real commits and tickets instead of a prompt written from memory. A qualified, independent SR&ED expert reviews and stands behind every claim before it's filed. The AI produces the draft faster. The named expert is who answers if the CRA has questions. Automation earns its keep by doing the capture well; it was never going to be the thing that stands behind the claim.
Questions worth asking before you trust an AI SR&ED tool
Before trusting any AI SR&ED tool, or your own AI-assisted drafting, ask where its evidence comes from, who reviews the output, and what happens if the CRA asks a follow-up question. The same questions apply whether you're evaluating a purpose-built SR&ED platform or just pasting your own summary into ChatGPT to draft a narrative — the tool changes, the evidence-and-review standard doesn't. Those three answers tell you almost everything you need to know.
- Where does the narrative's evidence come from? A tool pulling from your actual GitHub commits, Jira tickets, and technical discussions is working from real project data. A tool that turns a two-sentence prompt into a polished paragraph is working from nothing, however fluent the output sounds.
- Who reviews it before it's filed, and are they named? "Reviewed by our team" is not the same claim as "a qualified, independent SR&ED expert checked this claim and will speak to it if the CRA has questions." If nobody with SR&ED expertise is checking eligibility calls before filing, you're the one absorbing that risk, tool or no tool.
- What happens at the follow-up question? If the CRA asks for the specific technical uncertainty behind a claimed project, can you point to real evidence, or would you be reconstructing an explanation for what the AI wrote after the fact? If the answer to a review question is "let me ask the software," that's the tell that the evidence was never really there.
- Does it know the current rules? The 2026 changes raised the enhanced 35% refundable rate to apply on up to $6 million of qualified expenditures for CCPCs, restored certain capital expenditures, and opened the enhanced credit to some public corporations for the first time. A tool trained on stale program details can produce a confident, wrong claim just as easily as an outdated consultant template can.
If you're doing your own SR&ED claim by hand with no AI involved at all, none of this changes what you should be checking for. The evidence-and-review standard is the same one — it's just that AI makes it faster to skip both, which is exactly why it's worth being deliberate about.
What happens if an AI-assisted claim gets reviewed
An AI-assisted SR&ED claim gets reviewed exactly the same way any other claim does — the CRA doesn't have a separate process for claims that used automation. A reviewer reads the technical narrative, checks it against the eligibility test, and may ask for the records behind specific projects: the commits, the test results, the notes from the point where the approach changed. Per the CRA's most recent program statistics, 90% of claims filed in FY2025-26 were accepted as filed, with 6% accepted after modification and 4% denied — outcomes driven by the strength of the underlying evidence and narrative, not by whether a founder, a consultant, or a model drafted the first version of the sentences.
The uncomfortable case is the one worth thinking through honestly: a claim drafted almost entirely by AI, filed without meaningful human review, that reads well on the surface. It might get accepted as filed — plenty of claims are, on a risk-assessment basis, without a deep look. But if it's selected for review, the person defending it needs to be able to answer questions the drafting tool never had to answer: why this project, why this uncertainty, why this cost allocation. A claim with nobody behind it who can answer those questions in real time is a worse position to be in than a claim that took longer to prepare but has an expert who helped shape it and knows exactly where the evidence lives.
That's the actual stakes of the AI question, and it's not really about AI. It's the same stakes that have always applied to any claim nobody qualified checked before it was filed. Automation just changed how fast you can arrive at that position, in either direction — a bad unreviewed claim faster, or a well-evidenced, expert-checked claim faster too.
Frequently asked questions
Does using AI to write my SR&ED claim increase my audit risk? Not directly. The CRA's risk-assessment and review process doesn't check for AI involvement — it checks whether the technical narrative and the expenditures hold up to questions. What increases risk is a narrative with no real evidence behind it and nobody qualified checking it before filing, which happens with or without AI. An AI-drafted claim built from real project data and reviewed by a qualified SR&ED expert carries the same risk profile as a well-prepared claim drafted entirely by hand.
Is it legal to use AI-generated text in a SR&ED claim? Yes. The CRA's submission guidance requires the claim to be complete, accurate, and supported by evidence — it doesn't restrict who or what drafts the narrative. You, and whoever signs off on the claim, remain responsible for its accuracy regardless of drafting method.
Can an AI tool get my SR&ED claim rejected? An AI tool can't get a claim rejected on its own — a reviewer rejects a claim because the eligibility test isn't met or the evidence doesn't support what was claimed. What an AI tool can do is make it easier to produce a fluent-sounding narrative with nothing real behind it, faster than a person could write the same thin narrative by hand. That's a documentation problem, not an AI problem, and expert review before filing is what catches it either way.
What should I ask an SR&ED software vendor about their use of AI? Ask where the AI's evidence comes from (your actual dev tools, or a generic prompt), whether a named, qualified SR&ED expert reviews every claim before it's filed, and what happens if the CRA asks a follow-up question the AI wasn't equipped to answer. A vendor that can't answer all three plainly is asking you to carry review risk they should be sharing.
Is AI replacing SR&ED consultants and experts? No, and treat any pitch claiming otherwise skeptically. Automation is well-suited to the mechanical parts of a claim — surfacing eligible work, organizing evidence, drafting a first pass. It isn't suited to the part that actually protects you: a qualified, independent expert making the eligibility calls and standing behind the claim if the CRA has questions. The winning setup pairs both; neither one replaces the other.
The question was never whether AI is safe for SR&ED. It's whether the claim in front of you — however it got drafted — is built from real evidence and checked by someone who'd stand behind it if the CRA called. Answer that, and the tool stops mattering.
See what your claim could be worth — estimate your refund or check your eligibility.