Statistics Canada published in June 2026 a national profile of workers who use artificial intelligence, based on the new Canadian Survey of Working Conditions. Its results draw a map of access inequalities rather than a homogeneous transformation of work. The tool is currently amplifying existing hierarchies. The real dividing line is less about job destruction than about the capacity, which varies depending on individuals and companies, to benefit from the technology.
The Essentials
- Professional AI use varies sharply depending on profession, qualification level, and sociodemographic characteristics, according to Statistics Canada’s survey (June 2026).
- The most qualified and best-paid workers are the early adopters: AI reinforces an existing advantage rather than creating a new one.
- Daron Acemoglu has framed it this way for several years: a technology’s destiny depends on the institutions that govern its deployment and the actors who capture its gains.
- The trajectory is not fixed: training policies, corporate reorganization, and tools accessible to SMEs could broaden diffusion by 2035.
- The signal to watch is the adoption gap between degree holders and non-degree holders, and between large companies and SMEs.
Degree Holders First, Others Later
The Canadian survey does not measure AI’s economic performance. It measures who uses it. And the answer is clear: workers who report regular professional use of artificial intelligence are concentrated in high-qualification professions, in sectors with high cognitive intensity, and in upper income brackets.
This result may seem unremarkable. New technologies often arrive from the top. But what makes this moment distinctive is the implicit promise that accompanied generative AI since 2022: that of a universal tool, accessible to an intern as well as a director, capable of compensating for differences in training or experience. The Canadian survey puts this promise in tension with observed facts. The tool may be universal in its formal access.
Its productive use remains concentrated.
Several mechanisms explain this gap. The first is cognitive: formulating an effective query to a language model requires knowing what you’re looking for, understanding the field sufficiently to evaluate the response, and correcting errors when they occur. These capacities are unequally distributed, and they are built largely through higher education. The second mechanism is organizational: large companies, which proportionally employ more degree holders, have the means to train their teams, acquire licenses for high-performance tools, and integrate these tools into existing workflows. A ten-person SME generally has neither the time nor the resources to do the same.
Acemoglu’s Anticipation
Daron Acemoglu, whose work on technology and institutions has shaped part of the economic debate of the past two decades, advances a thesis that applies directly here: technology is not neutral. Its impact on the distribution of gains depends on the institutions governing its deployment, the choices made by companies about automation versus augmentation, and the balance of forces that determines who captures the value created.
In Power and Progress, cowritten with Simon Johnson, Acemoglu distinguishes two technological trajectories. The first automates existing tasks and reduces the need for workers to accomplish them. The second increases worker productivity by giving them new tools, creates new tasks, and expands the spectrum of what they can accomplish. Both trajectories coexist in the history of technological progress, but the second is the one that distributes gains most widely. For Acemoglu, the issue is understanding which incentives and institutions direct companies toward one trajectory rather than the other.
Canadian data fit within this framework. If professional AI is first concentrated among degree holders and large structures, it is also because companies have been little incentivized to extend it to other worker categories. The tools exist. Training is rare. Work reorganizations that would make AI useful to a technician or a service agent have not been undertaken at scale.
It is a choice, often made by default.
The Optimism of Technological Diffusion Has Its Own Arguments
Acemoglu’s reading on technological capture is not the only one available. A different tradition, that of progress studies, reminds us that most technologies that initially appeared reserved for an elite eventually diffused widely, by lowering costs, simplifying interfaces, and generating the training that was initially lacking.
This reading is not naive. It rests on the long history of electricity, personal computing, and the internet. Each of these waves experienced a phase of initial concentration, followed by diffusion that ultimately benefited worker categories well beyond the early adopters. Economists like Tyler Cowen and Noah Smith emphasize that productivity gains linked to AI could, in the long term, increase real incomes even for workers who don’t directly use the tool, through general economic growth and the lowering of the cost of goods and services.
Canadian data capture a moment, not a trajectory, and do not definitively settle between the two readings. They nonetheless pose a precise question: at what pace does diffusion occur, what accelerates it, and who remains left behind and for how long.
As shown by cases of AI in Asian public administrations, formal access to a tool and its transformation into collective productive capacity are two very different things. The gap between the two does not close spontaneously.
Professional Use Is Also Measured by the Bureau of Labor Statistics
Canadian data fit within a larger picture. The U.S. Bureau of Labor Statistics published in May 2026, in its Monthly Labor Review, an article titled AI and the Rise of Software Investment, with the BLS devoting a dedicated thematic page to the question, titled Productivity and Artificial Intelligence. The results of this analysis are consistent with what we observe in Canada: professional adoption of AI follows the usual lines of economic fracture. Sectors with high knowledge intensity—finance, consulting, technology, legal services—have integrated tools earlier and more deeply. Sectors with high intensity of low-skilled labor—retail, food service, construction, home care—remain largely outside.
This sectoral divergence has concrete consequences for productivity. Companies that adopt AI in already-productive sectors widen their advantage over those that don’t. The productivity gap between sectors, which is already measured in multiples, risks widening if diffusion remains concentrated. Workers who remain cut off from these gains do not necessarily lose their jobs immediately. They lose ground in terms of relative income, career advancement, and ability to negotiate their working conditions.
This last point deserves to be named. AI, in sectors that adopt it, reinforces the negotiating position of workers who master it. It makes their work harder to substitute, increases their measurable added value, and creates conditions for income progression. For those without access to it, the symmetrical effect applies: their relative position weakens, without any job necessarily disappearing in the short term.
The Next Thirty Years Don’t Write Themselves
The trajectory of professional AI is open. Two plausible scenarios structure the debate.
In the first, tools continue to simplify, their costs drop, and public training programs allow worker categories currently excluded to gain access to productive use. SMEs, which employ the majority of workers in most advanced economies, find solutions suited to their size and budget. Companies that reorganize work around AI discover that trained technicians or field agents can benefit from diagnostic, planning, or communication tools, and that these gains translate into productivity and income. In this scenario, the initial advantage of degree holders persists, but the gap narrows over time.
In the second, AI remains a tool of large companies and already well-positioned workers. Continuing education budgets remain insufficient. Tools accessible to SMEs remain limited compared to those available to large structures. The productivity gap between sectors and between worker categories widens. AI gains concentrate in a narrow segment of the economy, and political debate eventually crystallizes around redistribution of these gains rather than their expansion.
The signals that make it possible to distinguish these trajectories are identifiable right now. The evolution of the usage gap between degree holders and non-degree holders in annual surveys is the first. The adoption rate in SMEs versus large enterprises is the second. If these two indicators converge in the coming years, if the gap narrows rather than widens, the diffusion scenario becomes plausible. If they diverge, concentration becomes the dominant trajectory.
What makes the diffusion scenario possible is not market dynamics alone. It requires deliberate policy choices. Public AI training programs do exist in several countries—Canada, Denmark, and Singapore have launched initiatives in this direction—but remain limited in volume and scope. Reaching workers who need it most requires designing them differently than as additional university-level training. Pedagogical tools and on-the-job learning, integrated into collective agreements or human resources policies, would probably be more effective than additional certifications reserved for those who already have time and training.
For SMEs, the question is different. Access to tools is less a matter of cost than integration. Accounting, inventory management, or customer service software that integrates AI functions at a level usable without advanced technical skills already exist in some sectors. Their generalization depends partly on software vendors, but also on sectoral policies and professional organizations that can accelerate adoption.
Training as Leverage, Not Supplement
Companies that have succeeded in broadening AI use beyond their most qualified teams share a characteristic: they have invested in work reorganization, not just in tool access. Training a technician on ChatGPT without changing how their tasks are structured produces little effect. On the other hand, rethinking a workflow so that an AI tool generates a first draft, data analysis, or action plan that the technician refines and validates changes the nature of their work and their measurable added value.
This reorganization costs managerial time and requires a form of experimentation. Large companies can afford it more easily than SMEs. But it is not impossible at small scale: sectoral associations, chambers of commerce, or vocational training organizations can play a dissemination role by documenting and sharing what works in comparable companies.
Acemoglu posed the question before the arrival of generative AI, and Canadian data bring it back into focus: who decides on the deployment of a technology, and in whose interest. The answers depend on power relations within companies, public policies, and organizational choices. Statistics Canada’s data provide a starting point for orienting these choices. The trajectory that follows depends on those who decide to act on these signals.
Sources
- Statistics Canada, Workplace artificial intelligence use, June 17, 2026, https://www150.statcan.gc.ca/n1/en/catalogue/75-006-X202600100007
- Statistics Canada, Profile of workers using AI at work, June 2026, https://www150.statcan.gc.ca/n1/daily-quotidien/260617/dq260617b-eng.htm
- Statistics Canada, AI business analysis, Q2 2026, June 2026, https://www150.statcan.gc.ca/n1/pub/11-621-m/11-621-m2026010-eng.htm
- U.S. Bureau of Labor Statistics, AI and the Rise of Software Investment, Monthly Labor Review, May 2026, https://www.bls.gov/opub/mlr/2026/article/ai-and-the-rise-of-software-investment.htm
- Daron Acemoglu & Simon Johnson, Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity, PublicAffairs / Hachette, 2023, https://www.hachettebookgroup.com/titles/daron-acemoglu/power-and-progress/9781541702547/?lens=publicaffairs
- Daron Acemoglu, Nobel Lecture, American Economic Review, June 2025, https://www.aeaweb.org/articles?id=10.1257/aer.115.6.1709