Why do so few Canadian companies report a return on AI?
Because most of them are counting use rather than production. KPMG Canada found that 93 percent of business leaders say their organization uses AI, and 2 percent report a return on it. Statistics Canada, asking which businesses use AI to produce goods or deliver services, counts 19.2 percent. The gap between those two questions is the answer.
Two Canadian numbers, both published, both defensible, four times apart. The measurement problem has to be settled before the returns problem can be discussed at all.
Which number is right, 93 percent or 19.2 percent?
Both are, because they answer different questions.
KPMG Canada surveyed 753 business leaders between 15 August and 3 September 2025 and asked whether their organization uses AI. Ninety-three percent said yes, up from 61 percent a year earlier. Thirty-one percent called it fully integrated across operations. Two percent reported a return on the investment. Stephanie Terrill, the firm’s Canadian Managing Partner of Digital and Transformation, said when the survey landed that only a small sliver of Canadian businesses are generating growth from their AI investments today.
Statistics Canada asked a narrower question of a far larger sample. Did this business use artificial intelligence to produce goods or deliver services in the last twelve months. In the second quarter of 2026, 19.2 percent said yes, tripled from 6.1 percent two years earlier.
| KPMG Canada | Statistics Canada | |
|---|---|---|
| Question asked | Does your organization use AI | Did the business use AI to produce goods or deliver services |
| Who answered | 753 business leaders, self-reported | Businesses, in the national business conditions survey |
| What it counts | AI present anywhere, a subscription included | AI inside the work that earns revenue |
| Result | 93 percent, with 2 percent reporting a return | 19.2 percent |
Ninety-three percent measures permission. Nineteen-point-two percent measures production. Only one of those can produce a return, which accounts for most of the distance between the two figures in the headline.
What happens to the productivity gain once you control for everything else?
Most of it goes away. Jiang Li of Innovation, Science and Economic Development Canada and Huju Liu of Statistics Canada linked the Survey of Digital Technology and Internet Use to administrative business microdata and published the result in Canadian Public Policy in 2026. Firms using AI looked 16.8 percent more productive than firms that did not. Once the authors controlled for how productive those firms already were and for their other digital investments, the premium dropped to 5.1 percent and stopped being statistically significant.
The same study shows what predicted adoption in the first place. Firms already running data analytics were 15.0 percentage points more likely to adopt AI. Advanced robotics added 8.1 points, formal ICT training 3.0, cloud computing 2.8, and research and development spending 2.0.
Read the two findings together. The firms that gained were the firms that had already built the capability AI depends on. Adoption did not create the advantage, it compounded one that was there. The survey waves behind the estimate are from 2019 and 2021, so it predates the current generation of systems, and it remains the most careful measurement Canada has of what AI does to output rather than to sentiment.
Why 2 percent is a measurement failure before it is a technology failure
Some of those companies are getting a return and cannot show it. In the same KPMG work, 57 percent said capturing value from AI is a challenge, up from 40 percent a year earlier, and only 38 percent had a clear plan for extracting it. Among the firms that did report returns, 31 percent could not put a number on them.
A company that cannot name the unit, the baseline and the owner has no way to report a return, whether or not one exists. Four things make the claim provable.
- A unit of work the business already recognizes. An invoice, a claim, a quote, a shift.
- A baseline taken before the system went in, covering time, error rate and cost per unit. Taken afterwards it is an estimate wearing the clothes of a measurement, and the error budget has nothing to sit against.
- A named owner who reports the number on a schedule, rather than a slide assembled the week somebody asks.
- The full running cost in the denominator, which is the work cost per decision exists to do.
Count production, not permission
Decide what counts as use inside your own company, and set the bar at a system doing work in a live process rather than a licence somebody expensed. Run the readiness assessment against the process you want measured, take the baseline before anything changes, then report per unit at ninety days. ORKA AI builds against that number with operators in Calgary and across Alberta, because it is the only one a board can act on.
Ninety-three percent of leaders can say yes to a survey. Two percent can show the arithmetic. What separates those two answers is work inside the business, and no vendor can do it on your behalf.
Sources
- KPMG Canada, Canadian businesses adopting AI, but few are seeing ROI, November 2025. Survey of 753 business leaders, 15 August to 3 September 2025.
- Jiang Li and Huju Liu, Artificial intelligence adoption and productivity in Canadian firms, Canadian Public Policy, 2026, summarized by Statistics Canada.
- Statistics Canada, Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026.
Software is only the surface. Infrastructure is the rest.
