Most of what it costs to put AI to work in an organization is not the software price. In Statistics Canada's survey for the third quarter of 2025, businesses that planned to adopt AI said they expected to train staff (49.8 percent), develop new workflows (41.9 percent), and buy cloud services or storage (27.0 percent). Among businesses that did not plan to use AI, only 4.7 percent gave "too expensive" as a reason.
Key takeaways
- Training, new workflows, and data changes were named more often than hardware purchases by Canadian businesses planning to adopt AI.
- Cost was a minor stated reason for staying out: 4.7 percent of non-adopters cited it, compared with 78.1 percent who said AI was not relevant to their goods or services.
- Stanford's 2025 AI Index reports that running a system at GPT-3.5 level became more than 280 times cheaper between November 2022 and October 2024, yet unit price is only one line of a budget.
- A 2015 Google paper on machine learning systems argues that quick results hide large, ongoing maintenance costs.
- A worked example below, built on invented numbers, shows how human review time can outweigh usage fees.
What Canadian businesses expect to pay for
Statistics Canada released "Analysis on expected use of artificial intelligence by businesses in Canada, third quarter of 2025" on September 11, 2025. It reports that 14.5 percent of businesses planned to use AI over the following 12 months. The survey ran from July 2 to August 6, 2025. For that group, the agency listed what they expected to do once systems were in place.
| Expected action after adoption | Share of businesses planning to use AI |
|---|---|
| Provide staff with AI training | 49.8 percent |
| Develop new workflows | 41.9 percent |
| Purchase cloud services or storage | 27.0 percent |
| Change data collection or data management practices | 25.6 percent |
| Use vendors or consultants to install or integrate AI | 16.0 percent |
| Purchase computing power or specialized equipment | 14.2 percent |
| Hire staff already trained in AI | 12.6 percent |
People and process sit at the top of the list. Equipment and outside integration help come further down. The survey does not put dollar amounts on any of these items, so it shows where effort is expected, not how much it costs.
Why cost is a small reason for staying out
The same Statistics Canada analysis asked the 66.7 percent of businesses with no AI plans for their reasons. The leading answer was that AI is not relevant to the business's goods or services (78.1 percent, up from 74.2 percent a year earlier). Lack of knowledge about what AI can do came next at 11.3 percent, then privacy or security concerns at 8.1 percent. "Too expensive" fell from 6.1 percent to 4.7 percent, while lack of a skilled workforce rose from 2.3 percent to 4.6 percent.
These shares apply only to non-adopters. They say little about firms that tried AI and found the total bill larger than planned, a group the survey reports only through the 1.5 percent of non-planners whose previous or current use did not meet expectations. Adoption itself had doubled: 12.2 percent of businesses reported using AI in the second quarter of 2025, against 6.1 percent a year earlier.
Cheaper to run is not cheap to own
The 2025 AI Index Report from the Stanford Institute for Human-Centered AI states that the cost of running a system at GPT-3.5 level dropped over 280-fold between November 2022 and October 2024. It also reports that hardware costs fall about 30 percent a year and energy efficiency improves about 40 percent a year. Falling unit prices lower one line of a budget, and more usage can offset the saving.
Lines that do not fall with unit prices are the subject of a 2015 paper by D. Sculley and colleagues at Google, "Hidden Technical Debt in Machine Learning Systems." Its authors argue that quick results should not be treated as free, because real-world machine learning systems often carry large ongoing maintenance costs. They name risk factors such as entanglement between components, hidden feedback loops, undeclared consumers of a model's output, data dependencies, and changes in the external world. Each one adds work after launch.
A worked example with invented numbers
The figures in this section are an illustrative example. None comes from a vendor price list or a survey. Assume a Canadian organization uses a hosted language model at a hypothetical price of CAD 3 per million input tokens and CAD 12 per million output tokens. It handles 5,000 requests a day, each using 1,500 input tokens and producing 400 output tokens.
- Input: 7.5 million tokens a day, or CAD 22.50.
- Output: 2 million tokens a day, or CAD 24.00.
- Usage total: CAD 46.50 a day, or CAD 1,395 over 30 days.
- Training: 40 employees for three hours at CAD 45 an hour is a one-time CAD 5,400.
- Review: staff check 2 percent of outputs (100 a day) at two minutes each, about 3.3 hours a day, or roughly CAD 150 a day and CAD 4,500 over 30 days.
Under these assumptions the monthly review time costs more than three times the usage fees, and the training cost arrives before the first request. Change the review rate or the wage and the ratio changes, which is the point of building the estimate line by line.
Questions to settle before budgeting
- Which tasks will the tool take over, and who checks its work? Review time belongs in the estimate.
- Is the data ready? Cleaning, labeling, and access controls were not free before AI arrived and they are not free now.
- Which systems must connect to the tool, and who maintains those connections after the vendor project ends?
- How many staff need training, and how often will the training be repeated as the tool changes?
- What happens to the bill if usage doubles, and what is the exit plan if the vendor changes its terms?
- Are privacy, security, and compliance reviews scheduled and costed?
This article gives general information on cost categories. It is not financial advice, and it does not recommend a product or vendor.
FAQ
Is AI getting cheaper?
Per unit of use, yes. Stanford's 2025 AI Index reports a drop of over 280-fold in the cost of GPT-3.5-level inference between November 2022 and October 2024. Total spending depends on volume, integration, and oversight as well.
What share of Canadian businesses use AI?
Statistics Canada reported that 12.2 percent of businesses had used AI to produce goods or deliver services in the 12 months before the second quarter of 2025, up from 6.1 percent a year earlier.
Do most businesses avoid AI because of cost?
Not according to the third-quarter 2025 survey. Among businesses with no plans to use AI, 4.7 percent said it was too expensive, and 78.1 percent said it was not relevant to their goods or services.
Why do AI systems cost money after launch?
Models depend on data and on surrounding software that change over time. The 2015 technical debt paper describes ongoing maintenance as a typical cost of machine learning systems.
Related reading
- How AI can help businesses cut costs
- Unexpected benefits and cost savings from AI
- Benefits and drawbacks of implementing AI in schools
- How AI is changing modern society
More in AI Basics.
Last verified: October 2026
Sources
- Analysis on expected use of artificial intelligence by businesses in Canada, third quarter of 2025, Statistics Canada, September 11, 2025
- Analysis on artificial intelligence use by businesses in Canada, second quarter of 2025, Statistics Canada, June 16, 2025
- The AI Index Report 2025, Stanford Institute for Human-Centered Artificial Intelligence, 2025
- Sculley D et al., Hidden Technical Debt in Machine Learning Systems, Advances in Neural Information Processing Systems 28, 2015