The sentence under the number reads: businesses that used AI to produce goods or deliver services over the 12 months before they answered. That sentence, from the national survey analysis, is the whole story behind Canada’s headline adoption figure. In the second quarter of 2026, 19.2% of businesses in Canada met that test. A firm that ran 1 pilot in March and shelved it in April meets it. A firm whose staff open a chatbot at lunch and bill nothing through it does not. The number is honest. The use a board deck makes of it is where the trouble starts.
The direct answer: Canada’s 19.2% figure counts a business as an AI user if it used AI to produce goods or deliver services at any point in the 12 months before it answered the survey, which was collected between April 1 and May 6, 2026. It measures contact with AI in production, once, inside a moving 12-month window. It does not measure depth, spend, results or current use. Cite it for what it counts, and it serves a board well. Cite it as proof of a transformed sector, and the definition will not carry the claim.
- What the 19.2% figure counts, word by word
- The base trap inside the application table
- The window trap, and the 40.0% line nobody quotes
- A 4-line test for any AI figure that reaches your desk
What the 19.2% figure counts, word by word
Start with the verb phrase. Produce goods or deliver services is a production test. It excludes the owner who drafts email with a chatbot at night, the manager who tested a tool in a sandbox, and the firm that bought seats nobody opened. It includes the firm that used AI on 1 paying job in a year. Both edges matter. The figure understates how many Canadians touch AI at work, and it overstates how many firms run AI as part of the operation. A single number holds both truths because the definition was built to count a behaviour, and a behaviour count is what it is.
Then read the trend line with the same care. The figure stood at 6.1% in the second quarter of 2024 and 12.2% in the second quarter of 2025 before it reached 19.2%. The agency describes that run as a tripling in 2 years. The tripling is real under a steady definition, and a steady definition is the 1 thing a trend needs. A board can say adoption tripled. A board cannot say the average firm tripled its AI work, because the survey never measured any firm’s amount of work. It counted firms on 1 side of a line or the other, 3 years in a row.
The split under the headline carries the decision value. Information and cultural industries reported 42.3%, finance and insurance 40.4%, and professional, scientific and technical services 32.4%. Agriculture, forestry, fishing and hunting reported 4.5%, and construction reported 9.2%. Urban businesses reported 21.0% against 9.9% for rural businesses. A national figure that sits 23.1 points below 1 sector and 14.7 points above another is a map, and a map is useful only at the level where your firm actually sits. Budget 2026’s AI chapter isn’t for your firm. Two tax lines are made the same point about a federal document: the national text matters less than the 2 lines that reach your own operation.
The base trap inside the application table
The same release carries a second set of figures, and this is where decks go wrong. Data analytics leads the application list at 36.6%, then text analytics at 34.5%, virtual agents or chat bots at 28.2% and large language models at 24.8%. Those shares are shares of the firms that reported using AI. They are not shares of Canadian businesses. Read 36.6% as a share of all firms and you have overstated the data analytics market by a factor of about 5, because the base quietly changed from 19.2% of firms to the whole economy.
This trap is a reading failure. The table names its base in its notes, in plain language, for any reader who goes to the table instead of the headline. The failure happens in the retelling, when a slide lifts the bigger number because the bigger number argues better. The best AI in the world scored 12% on real professional work ran into the mirror image of this problem: a low figure that meant more than it looked, once the task set behind it was read. High or low, the number is the last thing to read. The base is the first.
The window trap, and the 40.0% line nobody quotes
The survey was in collection from April 1 to May 6, 2026. Each firm’s 12-month window depends on the day it answered, so the windows in the same release run from April 1, 2025 at the earliest start to May 6, 2026 at the latest end. A firm that adopted AI in June 2026 is a non-user in this release. A firm that stopped in June 2025 can still be a user. Neither firm is misreported. The instrument measures a trailing year, and a trailing year is always partly history by the day it prints.
Then there is the line that almost never reaches a slide. In the same release, 40.0% of businesses said AI use is not relevant to the business. Set that beside the 19.2% and the country reads differently: about 2 firms in 5 have looked at the question and answered that it does not apply to them, while about 1 in 5 used AI in production in the past year. A vendor who quotes only the growth line is selling against a market where the largest single answer is a considered no. A board that quotes only the 40.0% misses that the no is shrinking as the yes triples. Both lines are in the same document, 1 page apart, and a serious read holds both at once.
The barrier figures reward the same treatment. Cybersecurity or privacy concerns led the barriers at 13.4% of businesses, and cost followed at 10.6%. Those are shares of all businesses, a third base in 1 release. Anthropic lost $80.7B on paper. Its real operating loss was $693M. Buyers should read the second number taught the identical lesson from a filing: the figure that decides the argument is usually the second 1, the 1 with the definition attached. Half of rehired displaced workers earn less than before. Your next hire may be one of them showed what a labour figure does to a hiring plan once its base is read. The habit transfers. The documents change.
A 4-line test for any AI figure that reaches your desk
Keep 4 questions beside any statistic a vendor, a deck or a headline hands you. What behaviour does the figure count, in the survey’s own words? What is the base: all firms, firms that use AI, or firms that answered 1 question? What window does it cover, and when did that window close? What is the largest answer in the same document that argues the other way? A figure that survives those 4 questions earns its place in a board paper. A figure that fails 1 of them goes back to the person who brought it, with the question attached.
The recap is short. Canada’s adoption figure counts production use inside a trailing 12-month window: 19.2% in the second quarter of 2026, up from 6.1% 2 years earlier, with 40.0% of firms calling AI not relevant to the business. The application shares sit on a smaller base, and the barrier shares sit on the full one. The definitions are printed in the release for anyone who reads past the headline.
Your next board deck is already being drafted, and an AI figure will land in it before the month turns. Send the 4 questions with the draft request, and have the definition line pasted under every figure the deck carries. A number with its definition under it can still be argued with. A number without one can only be repeated, and repetition is how a soft figure hardens into a decision nobody checked.