In Canada, 12.2% of businesses were using AI in 2025, and their apparent productivity advantage reached 16.8%—a figure that vanishes once you account for investments and performance accumulated before adoption. Statistics Canada has just documented an association between AI adoption and preexisting complementary capacities. For many SMEs lacking sufficient complementary capacities, the gain remains uncertain.

The essentials

  • The apparent 16.8% productivity advantage of Canadian businesses using AI in 2025 disappears statistically after controlling for preexisting investments and performance (Statistics Canada).
  • AI is associated with higher productivity especially when businesses have complementary capacities.
  • Only one in eight Canadian businesses used AI in 2025; poorly digitized SMEs remain structurally outside the system.
  • Absent targeted public policy, AI productivity gains could concentrate further in large organizations with complementary capacities.
  • Avenues exist—shared platforms, conditional support, targeted training—but no program has yet scaled them.

The 16.8% figure examined

The announcement is seductive. Businesses that adopt AI are 16.8% more productive than those that don’t. The conclusion seems to write itself: adopt AI, gain productivity.

Statistics Canada did the rigorous statistical work that such a claim requires. Its researchers controlled for prior investments in technology, preexisting performance levels, company size, and sector. The 16.8% differential then ceases to be statistically significant. It disappears.

This result deserves closer attention. It does not say that AI is ineffective. It says that businesses equipped with multiple digital and complementary innovation capacities are more likely to adopt AI. Part of the productivity advantage preexisted AI adoption. The technology grafted itself onto an organization already prepared to exploit it.

The mechanism is known in economics as selection bias: early adopters of a general technology are never a representative sample of businesses. They are the best ones. Measuring their performance and attributing it to the technology is like measuring the effect of good health on recovery.

AI as a general technology: why complements are decisive

Electricity, the steam engine, computing: all major general technologies have followed the same pattern. Their economic impact is not immediate. It occurs when organizations learn to integrate them, when processes are redesigned to take advantage of them, when employee skills adapt.

Robert Solow formulated what is now called the productivity paradox: computers were everywhere in the 1980s, except in productivity statistics. Gains came a decade later, when businesses had finally reorganized their operations around the tool.

Generative AI will probably follow this same trajectory, with one crucial difference: the necessary complements—structured data, analytical skills, documented processes—are unequally distributed between large enterprises and SMEs. AI skills are already concentrating in a few large cities and in companies that could afford to recruit them before the wave hit.

A forty-person SME in manufacturing or personal services often has neither the data, nor the processes, nor the internal skills for a generic AI tool to produce a measurable effect. It can buy a subscription to a language model. Without external help, it cannot reconfigure its organization to derive real gains.

Adoption that remains concentrated

The rate of 12.2% of Canadian businesses using AI in 2025 is both high and revealing. High because diffusion has been rapid since 2022. Revealing because nine out of ten businesses remain outside it.

The distribution of this rate matters as much as its average value. Businesses with 100 or more employees report using AI more than several SME categories. They have teams to evaluate tools, budgets to integrate them, lawyers to manage data contracts. They generally have integrated information systems, structured databases, and documented processes.

The concentration of gains that Statistics Canada documents is not a transition accident. It reflects a structural logic: a technology whose return depends on costly complements tends to widen the gap between businesses with those complements and others. Employment growth already concentrates new jobs in low-wage sectors; if AI amplifies productivity gaps between businesses on top of that, the double pressure on the SME economic fabric will be hard to absorb.

Self-regulation or public intervention: a documented trade-off

Gilles Babinet, a digital entrepreneur, argues in an interview with Stratégies that AI reinforces existing hierarchies, erodes free will, and weakens collective capacity for self-determination. He argues for public authorities to set rules without waiting for a bubble to burst, against the myth of self-regulation.

Statistics Canada’s data support precisely a correlation between complementary capacities and AI adoption, without demonstrating future concentration of gains. The data show that businesses already equipped with digital capacities are more likely to adopt AI.

The opposing position deserves honest exposition. Economists close to liberal schools of thought, Tyler Cowen or Philippe Aghion on Schumpeterian growth, emphasize that general technologies naturally diffuse toward secondary users once adoption costs fall. AI in 2025 costs a fraction of what it cost in 2020. Nothing guarantees that the SME of 2030 will be as deprived as today.

The argument is serious. But Canadian data introduce a caveat: lowering the cost of the tool does not necessarily offset the absence of organizational complements. Structured data, documented processes, and training are especially necessary for in-depth organizational integration and more complex AI uses, not for every use of a language model. These complements require distinct investments. Spontaneous diffusion of AI toward SMEs is probable over the very long term; it remains uncertain in the short and medium term—precisely the period when productivity gaps could widen.

Two trajectories for the next decade

The question posed by Statistics Canada’s study goes beyond Canada. It concerns any country where AI adoption follows the same logic of selection by preexisting digital capital. And it has a temporal dimension: observed productivity gaps could worsen if the highest-performing businesses accumulate resources faster for the next waves of technology. More productive businesses generally have more resources to invest in capacities that will facilitate future adoption.

Two distinct trajectories emerge over a ten-year horizon.

In the first, federal policies targeting AI adoption by SMEs exist, but their effectiveness at correcting gaps is not yet established. SMEs with fewer digital capacities can access off-the-shelf tools, but available data show that complementary capacities are strongly associated with AI adoption. The productivity gap between large and small businesses could widen. The economic fabric slowly recomposes around a smaller number of more powerful actors. Benefits can concentrate in certain capital returns and certain businesses, but the effect on worker incomes is ambiguous and depends on adoption mechanisms, complementarities, and displacement.

In the second trajectory, public policies explicitly target poorly digitized businesses. Support covers audits of digital capacities, data structuring, team training, and process redesign—not limited to financing software licenses. Shared platforms allow SMEs to share data infrastructures they could not build alone. Public aid is conditional on a prior diagnosis of the business’s capacities, to avoid financing adoptions with no real effect.

This second trajectory is not theoretical. Precedents exist in other technological transitions: some SME digitalization support programs have shown the importance of combining financing, technical support, and training, rather than limiting themselves to investment tax credits alone. Direct transposition to AI is straightforward in its logic, even if specific needs—data quality, algorithmic governance—add new requirements.

Two signals will show which trajectory is being taken. First, the evolution of AI adoption rates in businesses with fewer than fifty employees, compared to that of large enterprises: if the gap widens year after year, spontaneous diffusion is not operating. Second, the productivity gain measured in SMEs that benefited from a public support program, after controlling for preexisting investments—the same methodological control applied by Statistics Canada. Without this rigorous measurement, the association between AI adoption and productivity can reflect a selection bias comparable to that observed in overall data.

Levers for public authorities

Avenues for progress abound. They share a common feature: they act on complements, not on the technology itself.

The first is sharing data infrastructure. An SME in agribusiness or transport cannot build alone the structured database that would allow an AI model to produce demand forecasts or detect anomalies in the supply chain. A sectoral grouping or a public platform can build it for them, provided that data governance guarantees the sovereignty of each participant.

The second is conditional aid. Financing the purchase of an AI service subscription without evaluating the organization’s capacities risks supporting adoptions that bring no significant gain. A prior audit—available data, documented processes, internal skills—allows resources to be directed toward businesses where support will produce real marginal gain, not toward those that would have adopted the tool anyway.

The third measure is targeted training. The challenge is to train operational managers capable of identifying relevant use cases, formulating requests adapted to tools, and interpreting their outputs—not training AI developers in SMEs. This translator profile between technology and business is what small structures need most. Public procurement already finances foreign industry in many technological sectors; directing part of these flows toward strengthening local SME capacities would be a direct lever.

The open question is one of political ambition. Support programs for SME digitalization exist in Canada as in most OECD countries. Their scaling toward specific AI issues—data quality, algorithmic governance, training in use cases—remains to be built. Statistics Canada’s study at least provides a solid empirical basis to justify the effort: without complementary investments, SMEs risk limiting the gains from deep AI integration, but they can nonetheless obtain modest benefits from standard tools.


Sources

  1. Statistics Canada, “Artificial Intelligence and the Productivity of Canadian Businesses,” 2026
  2. Statistics Canada, Daily Release, July 30, 2026
  3. Gilles Babinet, “We Are Already Becoming Machines,” Stratégies