We Should Not Care So Much About AI Adoption Rates

Canada’s new AI strategy sets a target made for the headlines: to raise businesses’ adoption of artificial intelligence from roughly 12% today to 60% by 2034. [1] However they are arrived at, these numbers are meant to signal ambition and imply that a wave of productivity and employment gains will follow once enough firms adopt the technology.

It is also the wrong thing to measure.

Adoption is an input. It tells us that something was purchased and installed. But it tells us nothing about whether the thing worked. A hospital that buys a triage tool has adopted AI. But so has the hospital that bought a triage tool, discovered it performed poorly on its actual patient population, and quietly stopped using it. Yet both scenarios count toward that 60% goal.

Counting AI users is no more meaningful than counting computers and calling it productivity.

This is not the first time we have chosen the wrong success indicator. Through the 1980s and 1990s, firms poured money into information technology while measured productivity growth stayed stubbornly flat. Economists came to call this puzzle “the productivity paradox”. [2, 3] The machines were everywhere, but the gains were not… at least not for a long time, and not homogeneously. The lesson was not that computers were useless. Rather, it was that buying a technology and benefiting from that same technology are separate achievements, sometimes separated by a decade of painful organizational rework.

The same gap is opening now, and faster. Recent surveys of firms deploying generative AI have found that a large majority of pilot projects fail to deliver measurable returns. [4, 5] Some of that is ordinary trial and error. But when a government sets a national adoption target, it changes what everyone optimizes for. Firms report deployment. Vendors sell licenses that satisfy a procurement checkbox. Nobody has much incentive to report that the pilot didn’t work. Certainly not the minister whose target it was.

So, what should we measure instead?

Anyone who’s ever developed a logic model knows the answer. You define success as having achieved the thing you actually wanted to achieve. Did the technology produce measurable productivity gains? Did public services become faster, more accurate, or more affordable? 

If we’re talking about AI use in health care settings, we should ask: did health outcomes improve? Real health outcomes, the outcomes that doctors and patients care about, like lowered costs, shorter wait lists, and better clinical outcomes. Did administrative burden fall, or did it simply migrate from one group of workers to another? Did the jobs that emerged pay better and offer more autonomy than the ones that disappeared? Did jobs emerge at all, or were they, in fact, culled?

The 60% adoption rate target is compatible with enormous gains concentrated in a handful of large firms in two or three cities, while smaller employers, rural regions, and lower-wage workers absorb the disruption without the windfall. [6] Imagine a hospital AI tool that improves care in downtown Toronto but performs poorly in northern communities because their populations were underrepresented in the training data. The national adoption rate would be the same in both cases, masking this heterogeneity in performance. That is why adoption alone is a poor guide to policy.

Perhaps we need an independent Canadian AI Outcomes Observatory. Such a body would act at arm’s length from the departments that spend the money and would have a mandate to evaluate whether publicly funded AI investments deliver on the promises.

An Observatory would require that federally funded AI projects specify in advance what outcome they expect to move and how it will be measured. This is the discipline we already impose, imperfectly, on clinical trials and on major infrastructure. The Observatory would publish results on a fixed schedule regardless of whether they flatter the program. Of course, it would need to disaggregate everything by region, sector, and worker demographics, among other variables. This is because national averages hide precisely the distributional damage we should be watching for. It would track failures as carefully as successes and maintain a public registry of both. 

Such a body would have the independence to say, out loud, that a flagship initiative did not work.

The objection is predictable, of course: it is bureaucracy layered atop innovation, a brake on a country that often already moves too slowly. But it is ignorance, not evaluation, that is the true brake. Not knowing whether initiatives are truly delivering is what keeps governments funding the same disconnected pilots year after year, and what allows a vendor whose product failed in one province to sell it, unblemished and unchanged, in the next. 

Canada’s problem has never been an absence of good research or good ideas. It has been an inability to tell which of them were actually paying off, and to concentrate resources there.

If we are to have an ambitious AI program, we must set rational targets for outcomes, measure them independently, and publish the results transparently. That is a harder promise for governments to make, because it has accountability built in.

References

[1] “Canada’s national Artificial Intelligence strategy: AI for All,” (in eng), Jun 03 2026.

[2] E. Brynjolfsson, “The productivity paradox of information technology,” Commun. ACM, vol. 36, no. 12, pp. 66–77, 1993, doi: 10.1145/163298.163309.

[3] E. Brynjolfsson and L. M. Hitt, “Beyond Computation: Information Technology, Organizational Transformation and Business Performance,” Journal of Economic Perspectives, vol. 14, no. 4, pp. 23–48, 2000, doi: 10.1257/jep.14.4.23.

[4] “The state of Generative AI in the enterprise, Q4 report: generating a new future,” (in eng), Jan 2025.

[5] A. Singla, A. Sukharevsky, L. Yee, M. Chui, and B. Hall, “The state of AI: how organizations are rewiring to capture value,” (in eng), pp. 1-24, Mar 2025.

[6] “Fostering an inclusive digital transformation as AI spreads among firms,” (in eng), pp. 1-8, Oct 25 2024, doi: 10.1787/5876200c-en.