Does AI Development Qualify for SR&ED Tax Credits? A 2026 Guide

The Scientific Research and Experimental Development (SR&ED) program is Canada's largest tax incentive for research and development, providing over $3 billion in tax credits annually. Yet many companies building AI products, training machine learning models, and pushing the boundaries of data science never file a claim — either because they do not know their work qualifies or because they are unsure how to document it.

If your team is solving genuine technical problems in AI — not just plugging in an API — you may be leaving significant money on the table. This guide explains exactly which AI activities qualify for SR&ED, how much you can claim, and how to document your work to survive a CRA review.

The CRA's Three-Question Test

The Canada Revenue Agency (CRA) evaluates every SR&ED claim against three criteria. All three must be satisfied for work to qualify.

1. Was There Technological Uncertainty?

Could a competent professional in the field have predicted the outcome in advance? If the answer is no, you have technological uncertainty. In AI, this often looks like:

• •Not knowing whether a particular model architecture will converge on your dataset
• •Uncertainty about whether transfer learning from a pre-trained model will achieve the required accuracy for a novel domain
• •Challenges in processing data at a scale or speed that no existing solution handles

2. Was There Systematic Investigation?

Did you follow a structured process of hypothesis, experiment, analysis, and conclusion? In AI development, this maps naturally to the experimental workflow most ML teams already follow:

• •Formulating a hypothesis (e.g., "a transformer-based model will outperform our current LSTM for this time-series task")
• •Running controlled experiments with tracked hyperparameters and metrics
• •Analyzing results and iterating based on findings

3. Was There Technological Advancement?

Did the work produce new knowledge or capability, even if the project ultimately failed? The advancement does not need to be publishable in a journal. It just needs to go beyond what was previously known or available to your team. For example:

• •Discovering that a particular data augmentation strategy significantly improves model robustness for your specific use case
• •Developing a novel preprocessing pipeline that handles a class of noisy data that existing tools cannot process
• •Even negative results count — learning that an approach does not work is still technological advancement

AI Activities That Qualify for SR&ED

The following types of AI work typically meet the CRA's three-question test:

• •Training custom ML models to solve novel problems — for example, building a computer vision system to detect defects in a manufacturing process where no off-the-shelf model exists
• •Developing new data pipelines with technical uncertainty — designing systems to ingest, clean, and transform data at scales or in formats that existing tools cannot handle
• •Creating novel neural network architectures — modifying or designing model architectures to solve problems where standard approaches fall short
• •Building ML infrastructure solving unprecedented scaling challenges — designing training pipelines, inference systems, or data platforms that handle requirements beyond what current solutions support

AI Activities That Do NOT Qualify

Not all AI work is SR&ED-eligible. The CRA specifically excludes routine technical work, even if it involves sophisticated technology:

• •Using off-the-shelf AI APIs without modification — integrating ChatGPT, Google Cloud Vision, or AWS Comprehend into your product as-is, without solving novel technical problems
• •Routine data analysis with established techniques — applying standard statistical methods or well-documented ML recipes to clean datasets
• •Implementing well-known ML algorithms with no uncertainty — following a tutorial or applying a standard approach to a problem where the outcome is predictable
• •Market research or A/B testing — these are business experiments, not scientific or technological experiments in the CRA's definition

How Much Can You Claim?

SR&ED refunds can be substantial. Here is a worked example for a small AI company with three developers spending most of their time on qualifying R&D work:

This is a simplified illustration. Actual refunds depend on CCPC status, taxable income, province of operation, and the proportion of time spent on qualifying activities. Consult a qualified SR&ED advisor for an accurate estimate.

Key rates to know:

• •35% refundable federal ITC — for Canadian-controlled private corporations (CCPCs) on the first $4.5 million of qualifying expenditures
• •15% non-refundable federal ITC — for expenditures above $3M or for non-CCPC corporations
• •Provincial credits vary — Ontario offers 8% (ORDTC), Quebec offers up to 30%, and other provinces have their own rates

Documentation Requirements for AI SR&ED

Documentation is where most AI SR&ED claims succeed or fail. The CRA expects contemporaneous evidence — records created during the R&D work, not written after the fact. Here is what you need:

• •Project descriptions — a narrative for each project explaining the technological uncertainty, the approach taken, and the results achieved or lessons learned
• •Experiment records linked to git commits — your version control history is gold. Link specific experiments to commits, branches, and pull requests that show the systematic investigation
• •Time tracking by project — each developer's time must be allocated between qualifying SR&ED work and non-qualifying work (e.g., bug fixes, feature work, meetings)
• •Technical artifacts — model performance logs (loss curves, accuracy metrics), architecture diagrams, experiment tracking outputs (from tools like MLflow or Weights & Biases), and technical design documents

Common Mistakes in AI SR&ED Claims

• •Claiming routine coding as SR&ED — building a REST API, setting up a database, or writing frontend code does not qualify, even if it supports an AI product. Only the work that directly addresses technological uncertainty counts.
• •Poor documentation — filing a claim based on vague recollections months after the work was done. The CRA looks for evidence created during the project, not reconstructed afterwards.
• •Not separating qualifying from non-qualifying work — if a developer spends 60% of their time on SR&ED-eligible work and 40% on routine development, you can only claim the 60%. Mixing the two inflates your claim and invites an audit.
• •Overreaching on technological advancement claims — claiming you advanced the state of the art in deep learning when you actually fine-tuned a pre-existing model with minor modifications. Be honest and precise about what was novel in your specific context.

Filing Your SR&ED Claim

Here is how the filing process works:

• •Form T661 — this is the primary SR&ED claim form. It includes your project descriptions, expenditure calculations, and the technical narrative for each project.
• •Self-file or use a consultant — you can file SR&ED yourself, but many companies use specialized SR&ED consultants. Most work on a contingency basis, charging 15–25% of the refund received. This means zero upfront cost.
• •Filing deadline — you must file your SR&ED claim within 18 months after the end of the fiscal year in which the expenses were incurred. Missing this deadline means forfeiting the entire claim for that year.

After filing, the CRA typically processes claims within 60 days for refundable credits (CCPCs) and 120 days for non-refundable credits. However, claims selected for review can take 6 to 12 months. Strong documentation is the best way to speed up the process.

Related Resources

• •SR&ED Grant Details
• •IRAP Grant Details
• •CDAP vs. IRAP: Which Grant Fits Your Business?
• •Official CRA SR&ED Program Page