Rapport technologique : méthode permanente

This page holds the parts of the Technology report's method that stay the same from issue to issue. Each issue has its own dated page with that month's datasets, figures and sources; the issue page wins wherever the two differ.

This season the report asks one question: will artificial intelligence (AI) finally move Canada's productivity numbers? Every issue takes one dataset, one correlation and one implication for the reader, and lands just before a decision the reader is making.

The Canadian Survey on Business Conditions (CSBC) is Statistics Canada's quarterly survey of business conditions. Its AI question asks whether a business used AI to produce goods or deliver services in the previous 12 months, and a separate question asks whether it plans to in the next 12 months. We use the published shares as they are, with no adjustment, and compare like with like: second quarter against second quarter.

The survey reports a share of businesses, not how intensively each one uses AI, so a firm with one licence counts the same as a firm that has rebuilt a workflow around it. It does not publish law, accounting, consulting or engineering separately; the finest cut is professional, scientific and technical services. Each issue's method gives Statistics Canada's quality grade for the rates it relies on.

The Job Vacancy and Wage Survey (JVWS) is Statistics Canada's quarterly survey of unfilled positions. We use vacancies, the vacancy rate (vacancies as a share of occupied and vacant positions), payroll employment and the average offered hourly wage, unadjusted, as published. Its finest industry cut for this audience is the three-digit professional services sub-sector, and its occupation table has no industry split, so occupation counts such as data scientist vacancies can only be read for the whole economy.

Labour productivity is real output per hour worked by industry, from Statistics Canada table 36-10-0480. We set each industry's productivity growth against its AI use rate where the two surveys' industries match, and say how many matched. A check that rests on one year of growth is reported as a supporting line, never as a finding.

We keep a list of Canadian law, accounting, consulting and engineering firms: the largest by headcount plus the regional firms that readers outside Toronto and Montreal know first, each tagged national or regional, with the page each headcount came from.

Postings come from each firm's careers page, from monthly Internet Archive snapshots of that page over the previous 12 months, from Workday job feeds and from Job Bank Canada, the Government of Canada job board. We remove duplicates by firm, title and link. A firm whose careers site returns nothing is reported as missing, never as zero.

Coverage bar. A firm counts as covered when at least 6 of the 12 months have archived postings and it has at least 20 unique postings. Firms below the bar are not used for any rate we publish.

Classification. A posting is AI related when its title matches one of: artificial intelligence, machine learning, generative AI, AI, large language model, data scientist, data science, natural language, prompt engineer, ML (machine learning) engineer, deep learning. Titles that match only "automation", "innovation" or "legal technology" are counted separately.

Tests that wait for enough firms. The firm-level tests run only once at least ten firms meet the coverage bar: AI share against headcount (Spearman rank correlation), against practice mix, billing model and tier (Kruskal-Wallis test and share of variance between groups), and a check for two clusters of firms (bimodality coefficient, where above 0.555 suggests two modes, and a two-component Gaussian mixture against one by the Bayesian information criterion, where a gap above 10 in favour of two is strong evidence). Until then the firm data is reported as supporting evidence only.

Where we compute a correlation or a change ourselves, the interval is a 95 percent percentile bootstrap interval from 2,000 resamples, with a fixed seed so the published file reproduces it exactly.

The field test runs a real professional task with a named, general-purpose tool bought by the token, timed and costed, so the result reflects what any firm can buy and price for itself. The Cost of Intelligence tracker records the list price per million input and output tokens, in United States dollars as published, for the models our readers buy. Once the field test runs, the tracker adds the cost of that task for each model: the task's token count multiplied by the list price.

Each issue makes dated, public claims with a confidence and a resolution date. They go into the predictions ledger when the issue is published, and a later issue grades each one when it comes due.

Statistics Canada data is used under the Statistics Canada Open Licence. Careers page facts and Job Bank listings are public. Vendor prices are as published on each vendor's pricing page on the retrieval date.