Ease AIThe Scientific Research and Experimental Development programme changed materially for taxation years beginning on or after 16 December 2024. The refundable expenditure limit rose from $3 million to $6 million. Phase-out thresholds moved from $10 million–$50 million to $15 million–$75 million. Capital expenditures were reinstated, including development-related software and cloud infrastructure.
Those changes are real and they matter. But the threshold question has not changed at all.
What SR&ED actually funds
SR&ED funds experimental development. That phrase has a specific meaning. The work must address technological uncertainty that cannot be resolved through standard practice. The outcome must be unknown at the start. You must be attempting something for which existing methods are insufficient.
Uncertainty about whether a client will like the interface is not technological uncertainty. Uncertainty about whether fine-tuning will converge on your dataset is.
The Canada Revenue Agency does not care that AI is new to your organisation. It cares whether the work itself advances the state of capability in a way that required experimentation.
When AI work qualifies
You are building a retrieval system that needs to handle 40,000-page regulatory documents with nested cross-references, and you are testing whether hierarchical embeddings with citation weighting will maintain accuracy at scale. That is experimental development.
You are designing a multi-agent system where one model critiques another's output in a feedback loop, and you do not know whether the critique agent will destabilise or whether token cost will make it commercially viable. That is experimental development.
You are training a model to detect a specific defect in manufacturing imagery, and existing pre-trained models fail on your materials, so you are experimenting with synthetic data augmentation techniques no one has validated for this substrate. That is experimental development.
In each case, the outcome is uncertain. Standard practice does not solve it. You are testing hypotheses.
When AI work does not qualify
You are configuring a chatbot platform to answer questions from a knowledge base. The platform exists. The method is documented. You are following the vendor's implementation guide. That is not experimental development.
You are automating a workflow by connecting an AI model to your CRM through published APIs. You are choosing which fields to pass and testing whether the prompts return useful output. That is configuration work, not experimental development.
You are building a website with an AI search feature using a hosted vector database and a managed embedding API. You are not developing new retrieval methods. You are applying existing ones. That is implementation.
The fact that your client has never used AI before does not make your work experimental. The fact that you are solving a real business problem does not make it experimental. The fact that it took skill and judgement does not make it experimental.
If you can describe the work as "we used X to do Y", and X is a product or method someone else built, you are not doing experimental development. You are doing valuable work. You are solving a problem. You are not eligible for SR&ED.
The capital expenditure change matters more than you think
Before the 2026 expansion, capital costs were largely excluded. If you bought compute or cloud infrastructure for development, you could not claim it.
Now you can. Development-related software and cloud infrastructure are back in. If you are running hundreds of training runs on rented GPUs to test whether a model architecture will work, those costs are claimable—provided the work itself qualifies.
This changes the economics for anyone building models or running large-scale experiments. It does not change the economics for anyone configuring SaaS tools.
Why this matters to small implementers
Most AI implementation work billed between $3,500 and $20,000 does not qualify. The timeline is too short. The work is too applied. The method is too well understood.
That is fine. The work still has value. The client still benefits. You should still do it. But your accountant should not be filing SR&ED claims for it, and you should not be pricing it as though she will.
If a consultant tells you a standard chatbot build will generate SR&ED credits, ask what technological uncertainty it addresses. If the answer is vague or centres on business uncertainty, the claim will not survive scrutiny.
When you should talk to your accountant
You are building something no one has built before. You do not know if it will work. You are testing methods that are not standard practice. You are generating data about what does and does not work, and that data has value beyond this one project.
You are spending material amounts on compute, on cloud infrastructure, on software licensing for development tools. You are tracking your time and you can separate experimental work from implementation work.
You have a taxation year beginning on or after 16 December 2024, so the expanded limits apply to you.
In those cases, talk to your accountant. Bring documentation. Be prepared to explain what you tried, what you learned, and why standard methods were insufficient.
What we do not claim
We build AI chatbots, workflow automation, and custom AI applications for small and mid-sized businesses. Most of that work is implementation. We configure platforms. We connect APIs. We test prompts and refine outputs.
That work does not qualify for SR&ED, and we do not tell clients it does.
If you are trying to understand what is claimable and what is not, your accountant is the right person to ask. If you want to understand what a typical AI build costs and whether other Ontario funding programmes might apply, we can help with that.
Contact us at info@easeaiworks.com or visit our pricing page to see what implementation work actually costs.