Canadian businesses doubled their AI adoption rate in a single year — from 6.1% to 12.2% between Q2 2024 and Q2 2025, according to Statistics Canada's Canadian Survey on Business Conditions. That sounds like momentum. It is also still less than half the OECD average, and the number that actually matters is buried three paragraphs down in most coverage: 42% of companies scrapped the majority of their AI initiatives in 2024, up from 17% the year before.
That is not a technology failure curve. That is a strategy failure curve. And it is about to get more expensive.
The Abandonment Rate Nobody Is Centering
S&P Global's finding — 42% of AI projects abandoned in 2024, cited in TechAhead's 2025 enterprise AI research — maps almost perfectly onto organizations that made the make-vs-buy call without understanding what they were actually deciding. The question sounds like a procurement question. It is actually a talent question, a maintenance liability question, and increasingly, a data-sovereignty question with a Canadian-specific cost structure attached.
Start with the buy side, because that is where most Canadian enterprises are defaulting. The Zylo 2026 SaaS Management Index puts the average annual organizational spend on AI-native apps at $1.2 million — up 108% year-over-year. That is the baseline before any customization, before compliance work, before integration with legacy systems that were not designed to talk to a large language model. A vendor relationship that looks like a manageable $100,000 pilot commitment can become a $1.2 million line item within eighteen months if the use case actually works. That is not a bug in the vendor's pricing model. It is the feature. Consumption-based billing is structurally designed to scale with your success, and enterprises that did not read that clause carefully are going to have an uncomfortable budget conversation before 2026 is out.
The build math is not obviously better. A mid-complexity custom AI project runs $40,000 to $250,000 upfront, according to CloudZero and Riseup Labs enterprise cost guides published in 2026. Then — and this is the part that gets glossed over in every CTO deck — ongoing maintenance typically costs 15 to 30 percent of the original build cost every single year. AI models drift on production data. They degrade. They require retraining cycles. KPMG Canada found in 2025 that only 24% of Canadian employees have received any AI education or training. You cannot execute a build strategy with a workforce that has not been trained to maintain what you build.
BC's Power Clause and the Infrastructure Cost Nobody Quoted
For any enterprise considering on-premises AI infrastructure in British Columbia, there is a variable that most national coverage skips entirely: BC's Bill 31 — 2025, the Energy Statutes Amendment Act, which requires AI and data centre operators to bid for provincial power. That is a material cost input, not a footnote. It is also a directional signal. On-premises builds that might have been viable in Metro Vancouver are now facing a power-procurement layer that adds timeline and cost uncertainty. The practical effect, according to analysis from Bennett Jones published in December 2025, is that on-premises AI builds are likely to migrate toward Alberta, fragmenting Western Canadian enterprise AI infrastructure in ways that will compound integration costs for BC-headquartered firms.
Layer on the regulatory trajectory. The proposed Artificial Intelligence and Data Act died on the order paper in January 2025, leaving Canada without overarching AI-specific legislation, per Osler's December 2025 analysis. The federal government is now regulating AI through privacy law and investment rather than a standalone statute. The proposed Consumer Privacy Protection Act has not landed yet, but the direction is clear: data sitting on American servers is going to carry increasing regulatory and reputational freight for Canadian enterprises in financial services and healthcare. That is a cost that does not show up in a vendor's per-seat pricing slide.
What MIT's Success Rate Actually Implies for a Hybrid Strategy
The most operationally useful framing comes from MIT's 2025 enterprise AI research: buying or partnering on AI succeeds roughly 67% of the time; fully internal builds succeed at approximately half that rate. Those are not odds that favor a pure build play for most Canadian mid-market firms.
The smarter operators are landing on a hybrid position — buy the commodity inference layer, build on top of proprietary data. A Microsoft Copilot deployment or a Claude API integration handles the model; the enterprise's own customer history, operational data, and domain logic become the defensible moat. That is a real strategy. It is also significantly harder to execute than either the pure-buy or pure-build narrative suggests, because it requires an honest internal audit of what data you actually have that is clean enough to train on or retrieve from.
The second-order effects of getting this wrong are worth naming directly:
- Mid-market Canadian firms that over-index on buy strategies will face renegotiation leverage loss as vendors reprice at scale — the same dynamic that played out with enterprise SaaS in 2018–2021.
- Firms that build proprietary training datasets now will command acquisition premiums; those that do not will be commoditized by the same platforms they are paying to use.
- Canadian data-sovereignty rules will quietly make U.S.-hosted AI platforms a compliance liability for regulated industries within 24 months — a cost that is currently invisible in most build-vs-buy analyses.
- BC's Bill 31 power-bidding requirements will push on-premises builds toward Alberta, creating a fragmented infrastructure picture for any enterprise operating across both provinces.
The Imagination Gap Is Real and the BDC Number Requires a Caveat
RBC's research identifies what it calls an "imagination gap" — most Canadian SME leaders cannot envision AI's practical relevance to their operations. BDC data appears to contradict this: 97% of AI-adopting SMEs report tangible benefits. Both numbers are true, and the tension between them is the whole story.
The 97% figure is a survivor-bias number. You are hearing from firms that adopted AI and stayed in the sample — not the ones that tried, failed quietly, and walked it back. A person familiar with enterprise software rollouts from the ERP era of the early 2000s would recognize the pattern immediately: large organizations bought licenses they could not integrate, built custom platforms their teams could not maintain, and ended up with shelfware. The failure modes for AI are identical. The difference is velocity and pricing model.
Statistics Canada's Q3 2025 survey found that 66.7% of Canadian businesses report no plans to adopt AI in the next 12 months. That cohort is not waiting for better compute infrastructure. They are waiting for a use case that is legible to their actual operations — which is an imagination and change-management problem, not a technology problem.
The federal government's C$925.6 million sovereign AI compute commitment in Budget 2025 is meaningful for researchers and large enterprises sophisticated enough to access it. It does nothing for the two-thirds of Canadian businesses that have not yet decided to move. The Pan-Canadian AI Strategy funding — fully dispersing through Mila, Vector, and Amii in 2026 — was designed to bridge the research-to-commercial gap, but federal research dollars flow slowly into enterprise capability at the mid-market level.
The Real Question Canadian Boards Are Approving Last
The make-vs-buy framing is ultimately a distraction if the organization has not solved its 24% AI training problem first. Most Canadian boards are approving AI budgets before they have approved AI literacy programs. That sequencing is exactly backwards.
Innovation, Science and Economic Development Canada received over 11,300 public submissions for its renewed national AI strategy, with six pillars revealed April 28, 2026 — but the full strategy has not been released as of this writing. Enterprises waiting for regulatory clarity before committing to large-scale in-house builds are waiting for a document that keeps moving. That is a reasonable position. It is also a position that compounds the adoption lag.
A senior integration consultant at a Vancouver-based enterprise software firm, who asked not to be named because their firm advises clients on both buy and build strategies, put it plainly: "The clients who fail are the ones who treat this as a vendor selection process. The clients who succeed treat it as an organizational redesign with a technology component."
According to McKinsey's Q1 2026 Global AI Survey, 65% of organizations globally now use generative AI in at least one business function — double the rate from ten months earlier. PwC Canada's Value in Motion report projects that closing Canada's AI adoption gap could lift Canadian GDP to $3.65 trillion by 2035, versus a $3.34 trillion baseline. That $310 billion delta is the actual stakes of the make-vs-buy decision at a national scale.
The enterprises that get there will not be the ones that bought the right platform or built the most sophisticated model. They will be the ones that figured out what they were actually deciding before they signed anything.
