Two-thirds of Canadian businesses have no plans to adopt AI in the next year, and the single most common reason is that it isn't relevant to what they do — a useful corrective before asking what the technology can and can't handle.
Key takeaways
A demo shows a system at its best, on a case chosen to show it off. A survey of actual usage shows something more honest, and Statistics Canada runs the only one that covers the Canadian economy directly. Its second-quarter-2026 release found that 19.2% of Canadian businesses had used AI to produce goods or deliver services over the preceding 12 months — a real and fast-growing figure, having tripled from 6.1% two years earlier — but also that 40.0% of businesses said AI simply “is not relevant to the business” (Statistics Canada, June 2026). A separate, forward-looking survey from the third quarter of 2025 found two-thirds (66.7%) of businesses had no plans to adopt AI over the following year, and of those not planning to adopt, 78.1% gave the reason as relevance, not failure or fear: AI “was not relevant to the goods or services they currently provide”, well ahead of a lack of knowledge (11.3%), privacy and security concerns (8.1%), or the view that the technology isn't mature enough (7.6%) (Statistics Canada, October 2025). Any answer to “what can AI do” that implies universal usefulness across every business is contradicted by the only Canadian survey there is.
Among the businesses that do use it, the pattern in what they use it for is informative. The most common applications in the same 2026 survey were data analytics (36.6%), text analytics (34.5%), virtual agents or chat bots (28.2%), and natural language processing (27.0%) — language- and pattern-shaped tasks with a wide margin for approximation, covered in more depth in how AI handles language versus numbers. Businesses have converged on these applications at real scale, which is itself evidence that the technology genuinely delivers value on tasks with this shape: drafting, summarizing, classifying, responding — work where a good-enough answer produced quickly is a real gain over doing it by hand, and where a person still reviews the result before it goes anywhere consequential.
Canada's Cyber Centre does not describe current limitations as edge cases to be patched in a future update; it names eight specific, standing risks in its own guidance on generative AI (ITSAP.00.041): misinformation and disinformation, phishing, privacy exposure through what users paste into a prompt, malicious code, buggy code introduced into development pipelines, poisoned training datasets, biased content, and loss of intellectual property. None of these is framed as rare. The same document states, as a general property of the outputs rather than an occasional glitch, that they “can be incorrect,” “might not make sense,” “might not take certain factors into account,” and “can be biased” — and instructs users to “always be aware of and validate” what they're given before treating it as accurate.
The standards-body framing sharpens exactly where this comes from: NIST's AI Risk Management Framework warns that systems “poorly generalized to data and settings beyond their training” carry increased risk and reduced trustworthiness (NIST AI RMF, a U.S. framework). Current systems are reliably good within the shape of their training data and measurably worse the further a real case drifts from it — which is a description of a boundary, not a fixed capability ceiling, but one that has to be actively managed rather than assumed away.
Canada's federal voluntary code for advanced generative systems is a useful checklist here precisely because it describes what “doing this well” is supposed to look like, even though it binds only its signatories and is not law. Its six committed outcomes are Accountability, Safety, Fairness and Equity, Transparency, Human Oversight and Monitoring, and “Validity and Robustness” — the last defined as systems that “operate as intended, are secure against cyber attacks, and their behaviour in response to the range of tasks or situations to which they are likely to be exposed is understood” (ISED voluntary code). Read as a checklist rather than a rulebook, it is a reasonable proxy for what “can do today, responsibly” requires beyond the raw model: a named owner, a pre-deployment risk assessment, attention to fairness, enough transparency for a user to make an informed decision, monitoring after launch, and behaviour that is actually understood across the situations the system will meet — not just the ones it was demoed on.
A working summary, not a scorecard
This is also, structurally, why a process that "still runs" is not the same claim as one that "still works" — the practical failure mode covered in why automations break quietly, and why the honest framing throughout this hub is capability at a named task, not general intelligence — the subject of is AI actually intelligent?
As of the most recent Statistics Canada survey, most do not — 19.2% reported using it over the preceding 12 months, up sharply from 6.1% two years earlier, while 40.0% said it wasn't relevant to their business. Adoption is real and growing quickly, but it is still a minority of Canadian businesses, concentrated in specific industries and application types.
Based on the mechanism and the Canadian sources here, unsupervised decision-making about a person, and exact computation, are the two weakest points — the first because no Canadian privacy or federal guidance treats it as acceptable without human review, the second because the underlying pattern-completion mechanism has no built-in way to guarantee an exact answer.
Some will likely narrow over time, but there is no published Canadian source that puts a date or a figure on that improvement, and this article does not invent one. The safer planning assumption is to build in review and monitoring for today's known limitations rather than to plan around a future capability that hasn't been demonstrated yet.
Matching a real task in your business to what current AI systems are actually good at — rather than what a demo suggests — is the starting point for a workable plan.
