Short answer: A fundable IRAP AI Assist project isn’t “we want to use AI.” It’s a specific AI capability inside your core product or service, with a real technical unknown, measurable targets, usable data, Canadian-based work, a route to customers and a safety plan, split into phases someone can judge. Start by talking to an IRAP advisor, not by filling in a form.
Two Canadian software companies call NRC IRAP in the same week. (Both are hypothetical.)
The first says: “We want to add AI to our platform to improve customer service.”
The second says: “Our support team spends most of its day answering contract questions from carriers. We want to adapt a language model to answer them inside our product, with at least 90% accuracy on an audited test set, in under three seconds, without exposing customer data. We have 40,000 resolved tickets, two engineers and three customers who’ve agreed to pilot.”
Only one of those gives an advisor something to assess. That gap, between an idea and a scope, is what this guide is about. We’ll cover what AI Assist is and isn’t, how IRAP says it judges projects, the seven parts of a scope that holds up, how to phase the work, and what to bring to your first call.
What IRAP AI Assist actually is
In April 2024, the federal government announced $100 million for the NRC IRAP AI Assist Program to help small and medium-sized businesses scale up and build and deploy new AI solutions (Prime Minister’s Office). NRC describes it as help for innovative SMEs navigating the challenges of developing and adapting generative AI and deep learning, with support for awareness, planning and execution. It backs scientific research, product development, testing and validation so companies can build those technologies into their core products and services, and its advisor network helps firms reach the knowledge, computing resources and AI tools they need (NRC).
IRAP isn’t new to AI. Since 2017 it has contributed $705.8 million to AI-related firms, supporting 1,111 firms and 3,837 projects. AI Assist adds a dedicated stream on top. The 2026 national strategy, AI for All, sets a goal of lifting business AI adoption from 12% to 60% by 2034 and promises direct support for SMEs moving from experimentation to integration (ISED). Integration, not experimentation, is the policy backdrop.
Two different things often get mixed up under the AI Assist name:
- The SME route. You work with IRAP and an ITA on advice and, where it fits, funding for AI research and product development inside your business. This is the part you apply for.
- The advisory engagements. NRC ran a separate call for proposals so that not-for-profits and post-secondary institutions could deliver AI advice to SMEs referred by ITAs. Each engagement is up to 20 hours over a maximum of four months, one per SME per fiscal year, for a flat $500 fee. The work covers framing how AI could be used (tools, technical and business requirements, cost components) and a dataset quality review, and ends in an action plan. Services run until March 31, 2029 (NRC call for proposals). That page doesn’t list the selected organizations, so ask your ITA whether an engagement is available to you.
Those two engagement services are, in effect, a scoping workshop. Use them as the skeleton for your own scope.
How IRAP decides: five questions, translated for AI
NRC says that when it considers a project for funding it assesses five things (NRC). Here they are, in AI terms:
- The technical aspects of your innovation: what exactly are you building, and what don’t you yet know how to do?
- Your business, management and financial capacity: can this team and balance sheet carry the project, including your share of costs?
- The likelihood of achieving results: is the target realistic given your data, time and talent?
- Your commercialization plans: who pays, and how does this feature change your revenue?
- Market potential and benefits to Canada: what does Canada get, whether jobs, IP or capability?
NRC also says IRAP does not fund day-to-day operating costs, non-technical or purely commercial activities, work done outside Canada, or research with limited potential for commercialization. In practice, that means paying for a chatbot subscription to speed up your accounting doesn’t fit, while building an AI feature that customers will buy might.
The minimums: your business must be incorporated, for-profit and operating in Canada, with up to 500 full-time-equivalent employees, developing and commercializing innovative technology-driven products or services, and ready to create economic benefits in Canada. Expect to provide your CRA business number, a business plan, recent financial statements, ownership details and résumés for your management and technical team. IRAP doesn’t support ULCs or LLCs, sole proprietorships, partnerships or cooperatives.
Why AI projects collapse on paper first
RAND interviewed 65 data scientists and engineers with at least five years of experience and found five leading root causes of AI project failure. It also reports that by some estimates more than 80% of AI projects fail, twice the rate of IT projects without AI (RAND). Each cause is something a good scope can head off:
- The problem is misunderstood or miscommunicated. Scope part 1 forces a plain statement.
- The data isn’t there. Scope part 4 makes you inventory it.
- The team chases the latest technology instead of solving a user’s problem. Scope parts 1 and 6 keep the customer in view.
- Infrastructure for data and deployment is thin. Scope part 5 covers compute and tooling.
- The problem is too hard for AI. Scope parts 2 and 3 test feasibility with numbers before you commit.
The seven-part scope
Write each part in plain language, about 100 to 150 words. Together they make a one-page answer to IRAP’s five questions.
1. The product problem
Name the customer, the job they’re trying to get done, and where the AI sits inside your core product or service. Quantify the pain in hours, errors, dollars or churn.
Weak: “Improve customer service with AI.” Stronger: “Carriers wait six hours for contract answers, and contract questions are 38% of tickets.” (Illustrative numbers.)
Test: could a stranger tell from this paragraph who benefits and what changes?
2. The technical unknown
This is the heart of the scope. If the work is a standard API call that any competent developer could finish in a week, there’s little for a research-and-development program to fund. Describe what isn’t yet known: whether a model can reach your accuracy on your domain’s data, how to keep answers grounded in your documents, how to meet latency and cost limits, how to evaluate outputs, how to deploy privately. Say what you’ll try and what you’ll do if it fails.
Test: what would a smart competitor still be unable to copy and paste?
3. Targets and the test that proves them
Pick three to five numbers: quality (accuracy, precision and recall, or human-rated usefulness), speed, cost per query, safety (rate of harmful or leaked outputs) and reliability. Say how you’ll measure them: test-set size, who labels it, and the baseline you’re beating, whether that’s your current process or an off-the-shelf model. One disclosed AI Assist project is described as determining its target architecture and quantitative technical objectives, which tells you how central targets are.
Test: if the project ended tomorrow, could you say whether it succeeded?
4. Data, rights and governance
State what you have (volume, format, quality, labels), where it came from, whether you have the right to use it for model development under your customer contracts and privacy law, what’s missing and how you’ll get it. The AI Assist advisory engagements include a dataset review covering governance, transparency, confidentiality, security, bias, consent and privacy-law adherence. Do that review yourself before you call.
Test: could you hand your data-use policy to a regulator without wincing?
5. Team, compute and Canadian work
List who does the technical work (employees versus subcontractors), their experience, and the compute and tools you need. Because IRAP doesn’t fund work done outside Canada, keep the build work and subcontractors in Canada and raise any exception with your ITA early.
Test: can you name a person responsible for every task?
6. The commercial path and benefits to Canada
Say who will pay, how the feature changes pricing, retention or new markets, which customers will pilot it, and what Canada gains in jobs, skills or IP. NRC assesses commercialization plans and market potential but doesn’t fund purely commercial activity, so keep the scope itself technical and explain the market alongside it.
Test: is there a named customer who wants this?
7. Safety and responsible AI
AI Assist is about developing these technologies safely and ethically. Map your risks (privacy, bias, hallucination, security, misuse), your controls (human review, logging, guardrails, red-teaming) and who monitors the system after launch. Canada’s voluntary code of conduct for advanced generative AI is a free scaffold, built on six principles: accountability, safety, fairness and equity, transparency, human oversight and monitoring, and validity and robustness (OECD.AI).
Test: what happens when the model is wrong, and who finds out?
Phase it
Big-bang scopes look risky. Phased ones look manageable. A scan of AI Assist records in the government’s grants and contributions database shows a pattern, though it’s a small, non-random sample:
- Phase 0, qualify and plan. One record is a $75,000 agreement (July 2025 to January 2026) titled as an AI qualification and planning project for a construction-software company building an AI onboarding engine (record). Another funds an innovation plan to explore AI in a firm’s core product.
- Phase 1, prove it. Records describe pilot development, testing and optimization of a generative AI model for a health app (record) and pilot testing of AI strategies to fix a target architecture (record).
- Phase 2, build and deploy. Another record covers developing and deploying an AI customer-service agent inside a mobile app (record).
Each phase should end in a decision: continue, change course or stop. NRC also notes the relationship typically starts with advice and referrals before any funding is provided, so Phase 0 can begin as conversation, not paperwork.
Red flags that make advisors hesitate
- The project is “adopt a tool” with no technical unknown.
- The AI is for internal operations, not your product or service.
- There’s no data, or no right to use it.
- No numbers define success.
- The plan relies on work done outside Canada.
- The proposal is a strategy deck with no build.
- The timeline assumes a model will work first try.
- The “scope” is mostly marketing and sales.
What to prepare before you call
A senior executive calls 1-877-994-4727. A contact centre decides whether you’re ready to meet a client engagement advisor, who may refer you to an ITA.
Have ready:
- Your CRA business number, business plan, recent financial statements, ownership structure and team résumés (NRC lists these).
- The seven-part scope on one or two pages.
- A short description of your data and any customer commitments.
- A phased budget that keeps technical work in Canada.
- Questions for your ITA: which phase fits our stage, and is an AI Assist advisory engagement available?
What it costs, and what the public record doesn’t say
IRAP support is delivered as non-repayable contributions that reimburse eligible costs, and NRC’s public pages don’t publish cost-share percentages. Advisory firms commonly report up to 80% of eligible technical salaries and up to 50% of eligible subcontractor costs (Leyton). Treat those as typical, not promised: your contribution agreement sets the real rates. Advisors also generally warn that costs incurred before approval aren’t eligible, so don’t start spending before you’ve agreed the project. You’ll need to show you can fund your share, and IRAP money is government assistance that interacts with SR&ED claims, so loop in your tax advisor early.
Guides describe continuous intake with no fixed deadline, and some advisors suggest talking to an ITA before April 1 to line up with the federal fiscal year. Confirm timing with your ITA.
If IRAP isn’t the right fit
NRC advisors can point you to other programs, and ISED’s Business Benefits Finder lists others. The government also launched a $200 million Regional Artificial Intelligence Initiative to speed AI adoption by SMEs through regional development agencies (ISED). IRAP also funds organizations such as the Vector Institute to run programs like a 2025–26 cohort that seconds machine-learning associates to up to 13 companies for generative and agentic AI work (record). Ask your ITA about programs like these.
The bottom line
IRAP doesn’t fund enthusiasm for AI. It funds specific, technical, commercially plausible work that its advisors believe you can pull off. Your job is to make that easy to see: one problem, one unknown, a few numbers, honest data, a Canadian team and a customer who’s waiting.
Spend an afternoon writing the seven-part scope before you call. If you can’t fill in a part, that’s your first piece of work, and a good conversation starter for your advisor.
