# How AI Automates SR&ED Claims: From Git Commits to CRA-Ready T661 Forms

Source: https://grantops.ai/en/blog/ai-automates-sred-claims/
Last updated: 2026-10-14

## The SR&ED Problem: Brilliant R&D, Broken Claims Process

Every year, Canadian companies leave an estimated **$1 billion or more** in [SR&ED tax credits](/en/sred/) unclaimed. Not because they are ineligible, but because the process of documenting and filing a claim is so painful that companies either skip it entirely, underreport their eligible work, or pay consultants 15-25% of their refund to handle it.

The traditional SR&ED claims process was designed in an era of paper lab notebooks and waterfall development. It assumes R&D teams will meticulously document every experiment, every hypothesis, and every result in neat narrative form. Modern software companies do not work that way. They push code in rapid iterations, track work in Jira or Linear, communicate in Slack, and rarely write anything that looks like a "technical narrative" for CRA.

The result is a massive disconnect: **the R&D evidence exists** — it lives in Git repositories, issue trackers, CI/CD pipelines, and design documents — but it is not in the format CRA wants to see on a T661 form.

AI changes everything. By connecting directly to the tools where R&D actually happens, AI can extract, analyze, and transform that raw evidence into CRA-compliant documentation automatically. This is not a future vision — [it is what GrantOps does today](/en/automate-sred/).

---

## The Traditional SR&ED Process: Why It Is Broken

Before understanding how AI fixes the problem, you need to understand how painful the traditional process is. Here is what a typical SR&ED claim looks like without automation.

### Step 1: The Year-End Scramble (2-4 Weeks)

Your fiscal year ends. Someone — usually the CFO, CTO, or an outside consultant — starts gathering information about what R&D happened over the past 12 months. This involves:

- Interviewing developers and engineers about what they worked on
- Asking them to recall which projects involved "technological uncertainty" (often 6-12 months after the fact)
- Reviewing Jira boards, commit histories, and old design docs
- Trying to reconstruct a timeline of R&D activities from fragmented records

Most developers dread this process. They are being asked to remember technical details from months ago and explain them in language that satisfies CRA's criteria. The information loss is enormous.

### Step 2: Writing Technical Narratives (2-6 Weeks)

For each eligible project, someone must write a detailed technical narrative covering:

- The **technological uncertainty** — what specific technical problem could not be solved with existing knowledge?
- The **hypothesis** — what approach did the team formulate to address the uncertainty?
- The **systematic investigation** — what experiments, prototyping, or testing was performed?
- The **technological advancement** — what new knowledge or capability resulted?
- The **documentation** — what contemporaneous records support the claim?

These narratives must be technically precise enough to satisfy a CRA Research and Technology Advisor (RTA) on review, but also clearly articulate *why* the work qualifies as R&D rather than routine engineering.

A typical company with 3-5 SR&ED projects spends 40-100+ hours writing and refining these narratives. If you are using a consultant, they are billing those hours at $200-400/hour, or taking 15-25% of your total refund as a contingency fee.

### Step 3: Financial Calculations (1-2 Weeks)

You must calculate eligible expenditures across multiple categories:

- **Salaries** — What percentage of each employee's time was spent on eligible R&D? This requires timesheets or reasonable estimates.
- **Overhead** — Calculated using CRA's proxy method (55% of salary costs) or the traditional method (actual overhead allocation).
- **Materials** — Supplies consumed during R&D, net of any residual value.
- **Contractors** — 80% of arm's-length contractor costs for R&D work.
- **Equipment** — Limited SR&ED-eligible equipment lease costs.

### Step 4: T661 Form Preparation and Filing (1-2 Weeks)

All of this feeds into the [T661 form](/en/blog/t661-form-guide/), which must be filed alongside your T2 corporate tax return. The form itself is intricate, with sections for project descriptions, expenditure calculations, and supporting schedules.

### Step 5: Wait and Hope (2-18 Months)

After filing, you wait for CRA to process the claim. If everything goes smoothly, you receive your refund in 60-120 days. If CRA selects you for review, which happens to roughly 20-30% of claims, you enter a review process that can take an additional 6-12 months.

**Total time invested in a traditional SR&ED claim: 80-200+ hours over 8-14 weeks.**

**Total cost if using a consultant: $20,000-$75,000+ or 15-25% of your refund.**

---

## How AI Transforms SR&ED: The Technical Architecture

AI-powered SR&ED automation works by connecting to the systems where your R&D evidence already lives, analyzing that data against CRA's eligibility criteria, and generating audit-ready documentation. Here is how each layer works.

### Layer 1: Data Ingestion — Connecting to Your R&D Tools

The first step is connecting to the tools your engineering team already uses. [GrantOps integrates with](/en/automate-sred/):

- **Git repositories** (GitHub, GitLab, Bitbucket) — Every commit, branch, merge, and pull request
- **Issue trackers** (Jira, Linear, Asana, GitHub Issues) — Every ticket, epic, story, and bug
- **CI/CD pipelines** (GitHub Actions, CircleCI, Jenkins) — Build results, test outcomes, deployment records
- **Communication tools** (Slack, Microsoft Teams) — Technical discussions, design decisions, problem-solving conversations
- **Documentation platforms** (Confluence, Notion, Google Docs) — Design documents, architecture decisions, technical specs

This is a read-only integration. The AI accesses your existing data — it does not modify anything in your tools.

### Layer 2: R&D Activity Identification — What the AI Looks For

Once connected, the AI analyzes your development activity to identify work that meets CRA's five SR&ED eligibility criteria. Here is what the AI examines in each data source and how it maps to CRA requirements.

#### In Git Commits and Pull Requests

The AI analyzes commit messages, code diffs, branch naming patterns, and PR descriptions to identify:

- **Experimental branches** — Branches named with patterns like `experiment/`, `spike/`, `research/`, `poc/`, or branches with unusually high commit-to-merge ratios indicating iterative experimentation
- **Failed experiments and reverts** — Commits that are later reverted, branches that are abandoned, or PRs that are closed without merging. These are strong indicators of systematic investigation
- **Algorithm development** — Changes to files in ML model directories, algorithm implementations, data pipeline transformations, or mathematical computation modules
- **Performance optimization beyond known methods** — Commits addressing performance that go beyond standard optimization (e.g., novel caching strategies, custom data structures, new algorithmic approaches)
- **Prototype and proof-of-concept work** — Rapid iteration patterns where multiple approaches are tried and discarded
- **Novel integrations** — Code that bridges systems or technologies in ways that required solving unknown technical challenges

#### In Issue Trackers

The AI analyzes ticket descriptions, comments, labels, and workflows to identify:

- **Technological uncertainty language** — Tickets describing problems as "unknown," "experimental," "needs research," "not sure if possible," or "investigating feasibility"
- **Research spikes** — Tickets explicitly tagged as research, spikes, or investigations
- **Multi-iteration resolution** — Tickets that were reopened multiple times, had approaches changed, or required extended investigation before resolution
- **Technical blocker patterns** — Tickets blocked by unresolved technical questions rather than resource constraints

#### In CI/CD Pipelines

- **Repeated build failures during experimental work** — Systematic testing of hypotheses
- **Performance benchmark runs** — Evidence of measuring and comparing different approaches
- **A/B test configurations** — Controlled experiments comparing implementations

#### In Communication Tools

- **Technical design debates** — Discussions weighing different approaches to solving a problem
- **"We don't know if this will work" language** — Direct evidence of technological uncertainty
- **Cross-functional R&D coordination** — Engineers, data scientists, and domain experts collaborating on technical challenges

### Layer 3: Mapping to CRA's Five Criteria

The AI does not just find R&D activity — it maps each identified project to CRA's five eligibility criteria for SR&ED. This is the critical step that determines whether work qualifies.

**Criterion 1: Technological Uncertainty**

The AI identifies evidence that your team faced a problem that could not be solved by applying existing knowledge or standard practice. It looks for:

- Documentation of failed approaches
- Evidence of exploring multiple solutions
- Technical discussions questioning feasibility
- References to limitations of existing tools, libraries, or methods

**Criterion 2: Technological Hypothesis**

The AI extracts evidence of a formulated approach or hypothesis. This maps to:

- Design documents proposing a technical approach
- Architecture decision records
- PR descriptions outlining the proposed solution
- Jira tickets describing an intended approach

**Criterion 3: Systematic Investigation**

The AI identifies methodical testing and experimentation:

- Iterative commit patterns showing incremental progress
- Benchmark comparisons between approaches
- Test suites for experimental features
- Evidence of controlled experiments

**Criterion 4: Technological Advancement**

The AI determines what new knowledge or capability resulted:

- Successfully merged code implementing a novel solution
- Documentation of lessons learned
- New internal libraries or tools created
- Performance improvements beyond what existing methods could achieve

**Criterion 5: Documentation**

The AI itself generates the documentation by compiling evidence from all connected sources into a coherent timeline. This is perhaps its most powerful function — CRA wants contemporaneous records, and your Git history, Jira tickets, and Slack messages are inherently contemporaneous.

### Layer 4: T661 Generation — From Evidence to CRA-Ready Forms

Once the AI has identified eligible projects and mapped them to CRA's criteria, it generates:

- **Project-level technical narratives** — Written in the format CRA expects, with clear sections for uncertainty, hypothesis, investigation, and advancement
- **Financial expenditure calculations** — Based on commit attribution (which developers contributed to which projects) and your payroll data
- **Time allocation estimates** — Derived from actual development activity (commits, PRs, ticket resolution) rather than after-the-fact guesswork
- **Supporting evidence packages** — Compiled documentation that can be provided if CRA requests a review
- **Complete T661 form data** — Ready for your accountant to file with your corporate tax return

---

## Time and Cost Comparison: Three Approaches to SR&ED

The following comparison uses a realistic scenario: a **20-person SaaS company with $2 million in annual R&D salaries**, 4 eligible SR&ED projects, and an expected refund of approximately $385,000 (at the 35% enhanced rate on eligible expenditures).

| Dimension | **Traditional Consultant** | **DIY (In-House)** | **GrantOps AI** |
|---|---|---|---|
| **Time investment (your team)** | 30-50 hours (interviews, reviews) | 100-200+ hours | 5-10 hours (setup, review, approval) |
| **Elapsed calendar time** | 8-14 weeks | 10-16 weeks | 1-2 weeks |
| **Direct cost** | $57,750 - $96,250 (15-25% of refund) | $0 (but massive opportunity cost) | Platform subscription |
| **Net refund after fees** | $288,750 - $327,250 | $385,000 (minus 100-200 hours of eng time) | ~$385,000 minus subscription |
| **Quality of documentation** | High (experienced practitioners) | Variable (depends on internal expertise) | High (AI-generated, evidence-backed) |
| **Audit readiness** | Good | Poor to moderate | Excellent (all evidence pre-compiled) |
| **Scales with company growth** | Cost increases proportionally | Time burden becomes unsustainable | Marginal effort stays flat |

### The Hidden Cost of DIY

Many companies try to handle SR&ED in-house to avoid consultant fees. The math looks good on paper — you save 15-25% of your refund. But consider the true cost:

If your senior engineers spend 150 hours on SR&ED documentation, and their loaded cost is $100/hour, that is $15,000 in direct salary cost plus the **opportunity cost** of 150 hours of engineering time not spent building product. For a startup, that opportunity cost can dwarf the consultant fee you were trying to avoid.

### The Hidden Cost of Consultants

SR&ED consultants provide genuine expertise, but the contingency fee model creates misaligned incentives:

- **They optimize for claim size, not claim quality** — Larger claims mean larger fees, but also higher audit risk
- **They have no incentive to help you improve your R&D documentation practices** — if your process improved, you might not need them next year
- **15-25% is a massive haircut** — On a $385,000 refund, you are paying $57,750 to $96,250

### The AI Advantage

AI-powered automation addresses the problems with both approaches:

- **Minimal time investment** — Your team spends days, not months
- **Evidence-based claims** — Documentation is derived from actual development data, not reconstructed memories
- **Consistent quality** — The AI applies the same rigorous mapping to CRA criteria every time
- **Audit-ready from day one** — Every claim is backed by a compiled evidence package
- **Improves over time** — The AI learns your codebase, your team's patterns, and your R&D focus areas

---

## Case Study: A 20-Person SaaS Company

Let us walk through a realistic (hypothetical) example of how AI-powered SR&ED automation works in practice.

**Company profile:**
- 20 employees (14 engineers, 2 data scientists, 2 product managers, 1 CTO, 1 CEO)
- $2 million annual R&D salary expenditure
- Building a B2B analytics platform with ML-powered anomaly detection
- Uses GitHub, Jira, Slack, and Confluence
- Fiscal year ending December 31

### Before GrantOps: The Old Way

In January, the CTO and CFO begin the annual SR&ED scramble. They hire an SR&ED consultant for a 20% contingency fee. Over the next 10 weeks:

- The consultant interviews 8 engineers (2 hours each = 16 hours of engineer time)
- Engineers review and revise technical narratives (4 hours each = 32 hours)
- The CFO compiles financial data and timesheets (20 hours)
- The CTO reviews all documentation (15 hours)
- Multiple rounds of revision between consultant and team (12 hours)

**Total internal time: 95 hours**
**Claim filed: Mid-March (10 weeks after year-end)**
**Refund received: June (if no review)**
**Refund amount: $385,000**
**Consultant fee: $77,000 (20%)**
**Net refund: $308,000**

### After GrantOps: The AI Way

In the first week of January, the CTO connects GrantOps to their GitHub, Jira, and Slack workspaces. The AI begins analyzing 12 months of development activity.

**Week 1:** The AI identifies 6 potential SR&ED-eligible projects based on development patterns, commit analysis, and ticket investigation trails. It presents these to the CTO with confidence scores and supporting evidence summaries.

**Week 1-2:** The CTO reviews the AI's findings, confirms 4 projects as eligible (the AI correctly flagged 2 borderline projects for human review), and approves the generated technical narratives with minor edits.

**Week 2:** GrantOps generates the complete T661 form data and financial schedules based on Git commit attribution and payroll data provided by the CFO. An audit-ready evidence package is compiled automatically.

**Total internal time: 8 hours (5 hours CTO review, 3 hours CFO financial data)**
**Claim ready: Mid-January (2 weeks after year-end)**
**Refund amount: $385,000**
**GrantOps cost: Platform subscription (fraction of consultant fee)**
**Net refund: ~$385,000 minus subscription**

The company saved 87 hours of internal time and tens of thousands in consultant fees, and had their claim ready 8 weeks faster.

---

## What About Accuracy? Can AI Really Understand CRA Requirements?

This is the most important question, and it deserves a direct answer.

### Where AI Excels

- **Pattern recognition at scale** — An AI can analyze 10,000 commits, 2,000 Jira tickets, and 50,000 Slack messages in minutes. No human can process that volume of evidence.
- **Consistency** — The AI applies the same eligibility criteria to every project, every time. Human assessors have off days.
- **Contemporaneous evidence** — Git history, Jira timestamps, and Slack messages are inherently contemporaneous. The AI's documentation is derived from real-time records, not reconstructed months later.
- **Reducing false negatives** — Companies routinely miss eligible projects because they forget about work done early in the fiscal year. The AI does not forget. For examples of what qualifies, see our [SR&ED eligible project examples](/en/blog/sred-eligible-project-examples/) covering software, manufacturing, biotech, and more.

### Where Human Review Remains Essential

- **Final eligibility judgment** — The AI identifies and recommends projects, but a knowledgeable human should make the final determination on borderline cases
- **Business context** — The AI may not understand why a particular piece of work was strategically important or how it connects to the company's broader innovation goals
- **CRA communication** — If a claim is selected for review, human expertise is needed to communicate with CRA effectively
- **Edge cases** — Unusual R&D structures (collaborative research, multi-entity projects, shared IP) require human judgment

The GrantOps approach is **AI-assisted, human-reviewed**. The AI does the heavy lifting of evidence extraction, analysis, and document generation. Your team provides the final review, context, and approval. This combines the speed and thoroughness of AI with the judgment and accountability of human oversight.

---

## Getting Started: From Zero to Automated SR&ED

If you are interested in automating your SR&ED claims, here is the practical path forward.

### Step 1: Assess Your Current State

Before automating, understand what you are working with:

- How many developers do you have? What are their salaries?
- What tools do you use for version control, project management, and communication?
- How many distinct R&D projects did you work on in the past fiscal year?
- Have you filed SR&ED claims before? If so, what was your refund?

### Step 2: Connect Your Tools

[GrantOps connects to your existing development stack](/en/automate-sred/) in minutes through OAuth integrations. No code changes, no agents to install, no disruption to your workflow.

### Step 3: Review AI-Identified Projects

The AI analyzes your development history and presents candidate SR&ED projects with supporting evidence. You review, confirm, or reject each one.

### Step 4: Refine and Approve Documentation

The AI generates technical narratives and financial calculations. You review and edit where needed — the AI provides a strong first draft, not a final submission.

### Step 5: File with Your Accountant

Export the completed T661 data and hand it to your accountant for filing with your corporate tax return. The audit-ready evidence package is archived and available if CRA requests a review.

---

## Frequently Asked Questions

### Does GrantOps file the SR&ED claim for me?

GrantOps generates CRA-ready T661 form data, technical narratives, and financial schedules. Your accountant files the T661 alongside your T2 corporate tax return. We provide the documentation; your tax professional handles the filing.

### What if CRA audits my AI-generated claim?

Claims generated with GrantOps are backed by comprehensive evidence packages compiled from your actual development records. This contemporaneous, verifiable evidence is exactly what CRA looks for during reviews, as detailed in our [SR&ED audit preparation guide](/en/blog/sred-audit-preparation-guide/). In many cases, AI-generated documentation is more audit-ready than traditionally prepared claims because every assertion is linked to specific commits, tickets, and records.

### Do I still need an SR&ED consultant if I use GrantOps?

For most small and mid-size companies, no. GrantOps handles the evidence gathering, eligibility analysis, technical narrative generation, and financial calculations that consultants typically perform. If your company has complex R&D structures (multi-entity claims, international collaborations, appeals), you may benefit from specialist advice for those specific situations.

### How does the AI determine which developers spent time on eligible projects?

The AI analyzes Git commit attribution (who committed code to which repositories and branches), Jira ticket assignments, and PR reviews to build a time allocation model. This is combined with your payroll data to calculate eligible salary expenditures. The model is presented for your review — you can adjust allocations before finalizing.

### What if my company does not use GitHub or Jira?

GrantOps integrates with multiple platforms including GitLab, Bitbucket, Linear, Asana, and others. If your tools are not directly supported, you can provide data through CSV uploads or API integrations. The core analysis works with any structured development data.

### Can AI-generated claims be trusted by CRA?

CRA evaluates claims based on the quality of the technical narratives and supporting evidence, not on who or what prepared them. AI-generated claims that are properly reviewed and approved by knowledgeable company representatives meet the same standard as consultant-prepared claims. The key is the quality and accuracy of the content, not the tool used to create it.

### Is my source code safe?

GrantOps uses read-only access and processes code metadata (commit messages, diffs, file names, branch structures) rather than extracting or storing your full source code. All data is encrypted in transit and at rest, and you can revoke access at any time. [Review our security practices](/en/automate-sred/).

---

## The Future of SR&ED Is Automated

The SR&ED program delivers approximately $4.5 billion annually to Canadian businesses, but the claiming process has barely changed in decades. Companies still reconstruct R&D narratives from memory, pay consultants a quarter of their refund, and wait months for CRA to process their claims.

AI automation eliminates the friction. Your R&D evidence already exists in your development tools. AI extracts it, maps it to CRA's criteria, and generates audit-ready documentation — in days instead of months, at a fraction of the cost.

The companies that adopt this approach get their refunds faster, keep more of their money, and spend their engineering time building product instead of writing retrospective documentation.

[Start free — connect GitHub and get a T661 draft in days, not months](https://app.grantops.ai/auth?mode=signup).

---

Try GrantOps: https://app.grantops.ai/auth?mode=signup (free to start, no credit card). Plans: Starter $99 CAD/month, Professional $499 CAD/month, plus an 8% success fee. Pricing: https://grantops.ai/en/pricing/