Almost always in tokens for billing — a different measure entirely from the share of Canadian businesses using AI, which Statistics Canada tracks separately.
Short answer
Almost always in tokens — small chunks of text, roughly a few characters or part of a word — because that's the unit the underlying model actually processes and the one vendors bill against. A completely different, and easily confused, sense of “AI usage” is the share of businesses using AI at all, which Statistics Canada tracks by survey, not by token.
A token is the small chunk of text — often a word or part of a word — that a model actually reads and generates one at a time; it's the unit behind both rate limits and billing. Anthropic’s own developer documentation, for example, describes a dedicated endpoint that lets a developer “determine the number of tokens in a message” before sending it, specifically so they can manage cost and rate limits, and it flags plainly: “The token count is an estimate.” (Anthropic, Token counting) The same documentation notes that a newer tokenizer from that vendor produces roughly 30 percent more tokens than an older one for identical text — a vendor-specific figure about that vendor's own product, not a constant that applies to every AI tool.
A completely separate use of the word shows up in national statistics. Statistics Canada's business-conditions survey asks whether a business used AI over the preceding 12 months and reports the share that did, verbatim: “In the second quarter of 2026, 19.2% of businesses reported using AI to produce goods or deliver services over the 12 months preceding the survey. This proportion has tripled since the second quarter of 2024 (6.1%).” (Statistics Canada, AI use by businesses, Q2 2026)
A token count inside a billing dashboard and a national adoption percentage answer completely different questions — one measures how much of a specific tool a specific account consumed, the other measures how many businesses use any AI tool at all. Confusing the two produces claims that sound precise and mean nothing. See also how a document's length affects what a model can actually process at once and the same fixed training data behind why a model doesn't know the date.
See what actually drives a custom AI build's cost, beyond a per-token rate card.
