Measured cost savings from artificial intelligence are real in narrow, well-defined jobs and hard to find in company-wide figures. Statistics Canada found in 2026 that the productivity edge of Canadian AI-using firms was not statistically significant once other capabilities were accounted for, and a large executive survey reported a small average effect on productivity over three years. The best-documented savings come from tightly bounded tasks such as data center cooling, where the developer itself reported results.
Key takeaways
- Among Canadian firms, AI adopters showed 16.8 percent higher labor productivity than non-adopters, but the gap fell to 5.1 percent and lost statistical significance after controls.
- In Statistics Canada's 2024 survey, 84.9 percent of AI-using businesses reported no change in employment, so savings did not usually come from cutting staff.
- An executive survey in the NBER working paper "Firm Data on AI" reports that AI raised annual labor productivity growth by 0.2 percent over three years.
- Google's DeepMind reported cutting data center cooling energy by about 30 percent on average in a system with human override, a developer-reported figure.
- A savings claim should state the baseline, the period, the cost of the AI itself, and who measured it.
What Canadian firm data show
Statistics Canada published "Artificial intelligence adoption and productivity in Canadian firms" on April 22, 2026, written by Jiang Li and Huju Liu. It uses the Survey of Digital Technology and Internet Use linked to administrative records. On its face, AI adopters have a labor productivity level 16.8 percent higher than non-adopters. Controlling for initial productivity cuts the gap to 10.2 percent. Adding complementary capabilities such as data analytics, cloud computing, and training cuts it to 5.1 percent, which is not statistically significant. The authors write that AI adoption does not appear to have a significant relationship with short-term productivity growth, and they note that benefits may take time and depend on organizational change.
The reading is that firms good at digital work tend to adopt AI, so a simple comparison of adopters and non-adopters mixes the effect of the tool with the effect of the firm.
What businesses say changed
Statistics Canada's June 20, 2024 analysis of survey results from April and May 2024 asked businesses about the previous 12 months. Of the 6.1 percent of businesses that used AI in their output, 84.9 percent reported no change in employment after adoption, and 6.3 percent reported a decrease. On tasks, 44.1 percent said AI reduced work previously done by employees to a small extent, and 39.2 percent said the reduction was moderate or large. The survey did not ask for dollar savings. Those answers describe work shifting, not money saved.
What executives report across countries
The NBER working paper Firm Data on AI, first issued in February 2026 and revised in October 2026, surveys senior executives across seven countries including Canada and the United States. The current version reports that 78 percent of firms use AI and that US firms expect to spend about USD 2,000 per employee on it in 2026. Executives report essentially no net effect on employment over the last three years. They say AI raised annual labor productivity growth by 0.2 percent over that period and expect 0.6 percent per year over the next three. Because the paper was revised, figures in earlier press coverage may differ from the current version, and all of them are executive self-reports.
A case where savings were measured: data center cooling
In July 2016, DeepMind reported that a machine learning system had cut the energy used for cooling a Google data center by 40 percent, which it translated into a 15 percent reduction in overall power usage effectiveness overhead. The test compared a day with the recommendations on against a day with them off at one facility. In August 2018, DeepMind and Google described a version that controlled cooling directly and reported savings of about 30 percent on average after nine months, measured against the earlier baseline.
Three features make this case checkable. The target is a single measurable quantity, energy per unit of cooling. The baseline is the plant's own prior operation. Operators could override the system at any time, and a local controller checked each instruction against safety limits. The figures still come from the developer, and the system relied on thousands of sensors, so the result may not transfer to settings with sparser data.
A short test for any savings claim
| Question | Why it matters |
|---|---|
| What was the baseline? | Savings against a poor starting process look larger than savings against a good one. |
| Does the figure include the AI's own costs? | Licenses, cloud use, integration, and review time can offset the gain. |
| Was the comparison concurrent? | A before-and-after comparison can capture seasonal or market changes. |
| Who measured it? | Vendor and developer reports are less independent than audited or survey data. |
| Did the work move or disappear? | Tasks shifted to remaining staff are not savings. |
FAQ
Does AI usually save money by replacing workers?
The Canadian survey data point the other way. In 2024, 84.9 percent of AI-using businesses reported unchanged employment, and the NBER executive survey likewise reports essentially no net employment effect over three years.
Why do task-level studies and company-level data disagree?
A gain on one task may be absorbed by review, rework, or unchanged processes around it. Company-level statistics also mix firms that adopted for different reasons, so averages can hide both gains and failures.
Is the 30 percent cooling figure typical?
It is a developer-reported result for one type of facility with unusually rich sensor data. It should not be read as a general expectation for other industries.
What costs reduce AI savings?
Staff training, new workflows, data changes, and review time. In a September 2025 Statistics Canada analysis, 49.8 percent of businesses planning to use AI said they would train staff after implementation.
Related reading
- The wider cost of artificial intelligence
- AI cost estimation for businesses
- Benefits and savings from implementing AI
- Emerging technologies and industry benefits
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Li and Liu, Artificial intelligence adoption and productivity in Canadian firms, Economic and Social Reports, Statistics Canada, April 22, 2026
- Statistics Canada, analysis of 2024 survey results on business AI use, series 11-621-M, June 20, 2024
- Yotzov et al., Firm Data on AI, NBER Working Paper 34836, February 2026 (revised October 2026)
- DeepMind AI reduces Google data center cooling bill by 40%, DeepMind, July 20, 2016
- Gamble and Gao, Safety-first AI for autonomous data center cooling and industrial control, DeepMind, August 17, 2018
- Statistics Canada, analysis of 2025 survey results on planned business AI use, series 11-621-M, September 2025