Learn
SR&ED for teams building with AI
Short answer: yes, R&D done with AI coding agents can qualify for SR&ED. What qualifies is the human-directed experimental work: defining technical problems, forming hypotheses, designing experiments, and judging results. What has changed is how you prove it. Commits and closed tickets no longer show that a person did the work, so AI-driven teams need different records.
Key Points
What changes when agents write the code
- 01
Eligibility hasn't changed
SR&ED rewards systematic investigation aimed at technological advancement, not typing code. Under subsection 248(1) of the Income Tax Act, programming has only ever been eligible when it directly supports that investigation.
- 02
Agents move effort up the stack
When agents handle routine implementation, your team spends more of its day forming hypotheses, designing experiments, and evaluating results. That can increase the share of time that is eligible.
- 03
Proof is the hard part
The question is rarely whether the work qualifies. It's whether you can show that a human directed it, and how long that took.
- 04
Capture intent and time
Agents now produce much of the commit history, so pair records of human decisions (specs, ADRs, overruled recommendations) with a light record of human time (story points, recorded standups, retros, or a weekly split).
Two Kinds of AI Company
Building AI vs. building with AI
Building AI and ML products
Work on new model architectures, training methods, evaluation approaches, or data techniques can qualify when the outcome wasn't knowable in advance and standard practice couldn't resolve it. Integrating an existing model or API, or routine prompt and parameter tuning, generally doesn't qualify on its own.
Using AI agents to build software
The eligibility test is the same as it was before agents. The work qualifies when your team is resolving technological uncertainty through systematic investigation. The challenge is evidence: showing that humans directed the investigation, and how much time it took.
Terms
Key concepts
The vocabulary we use when advising AI-heavy engineering teams on SR&ED.
- Decoupling problem
- The loss of any reliable link between code output and human effort once AI agents generate most of a team's commits, pull requests, and closed tickets. Commit history stops proving who did the work, or for how long.
- Intent metrics
- Evidence of human direction: design specifications, Architectural Decision Records, notes from sessions where engineers challenged or overruled an agent, and documented hypotheses. They show a human acted as the researcher directing the work.
- Vibe coding
- Mixing ideation, design, and routine implementation in one long, unstructured agent chat. It hurts agent output quality and makes it nearly impossible for a CRA reviewer to separate experimental work from routine coding.
- Agentic engineering
- A structured, spec-driven way of working with AI agents: a project charter with goals and risks, precise specifications, and those documents fed to agents as context. It improves agent output and produces an audit-ready record of the investigation.
Checklist
Records worth keeping
Most of these come out of normal engineering practice. You need some evidence of intent and some evidence of time, not all of them.
Project charter
Goals, the technical questions to answer, and foreseen risks.
Design specifications
Constraints, architecture decisions, and milestones given to agents as context.
Architectural Decision Records
Why each approach was chosen, and what was tried and rejected.
Agent review notes
Where engineers challenged, corrected, or overruled the agent.
Human-scoped story points
Sprint estimates sized around human effort, not agent output.
Recorded standups
A 60-second weekly update per developer, captured by your meeting notetaker.
Sprint retrospectives
Where estimates blew up and why. These are your technical setbacks, documented.
Weekly time split
Two minutes on Friday logging a percentage split by project.
Further Reading
Our writing on SR&ED and AI
June 2026
Your AI Agents and Your SR&ED Claim Should Grow Together
Why vibe coding dilutes your evidence, and how agentic engineering fixes it.
Series, Part 1
Stop Documenting "for SR&ED"
Healthy engineering habits already produce the records a strong claim needs.
Series, Part 2
The 'Decoupling' Problem
What CRA needs to see when agents write the code, and low-friction ways to capture it.
FAQ
Common questions from AI-driven teams
Is your agentic workflow claim-ready?
We'll look at how your team works with AI agents and tell you what your records already support, and what's missing. No cost, no obligation.