Your competitors are moving to AI: what the numbers show, and where to start

On a Tuesday morning, a small business asks three firms to review some forty documents before an acquisition. The first firm sends a preliminary reading the next day, every amount traced to its page; the other two promise an answer by the end of the week. None did poor work, but the client remembers the fastest, with evidence. The scene is made up; the numbers that follow are not.

In its Canadian Survey on Business Conditions, conducted from April 1 to May 6, 2026 (9,251 responses), Statistics Canada found that 19.2% of businesses had used artificial intelligence (AI) to produce goods or deliver services over the previous 12 months. The share was 12.2% a year earlier, and 6.1% in the second quarter of 2024. In the third quarter of 2026, 25.2% of businesses planned to use it over the next 12 months, up from 14.5% a year earlier; 52.7% had no such plans.

Quebec has picked up speed. Drawing on the same survey, the Institut de la statistique du Québec (ISQ) found that 12.7% of Quebec businesses used AI in the second quarter of 2025, on par with the Canadian average (12.2%), but with slower growth than in Ontario. A year later, the survey puts Quebec at 20.1%, Ontario at 18.1% and the country as a whole at 19.2%; 27.4% of Quebec businesses plan to use AI over the next 12 months.

Among businesses that used AI, 36.6% used it to analyze data, 34.5% to analyze text, 28.2% for virtual agents or chatbots, and 24.8% for large language models. Business AI is not, first and foremost, a conversation with an assistant.

What holds the others back: 40.0% consider AI not relevant to their business, 13.4% cite cybersecurity or privacy, 26.4% in health care and social assistance.

The sector Statistics Canada calls “professional, scientific and technical services” includes, in the industry classification (NAICS), offices of lawyers and notaries, offices of accountants, and research and development in the life sciences. If you run a firm, this is your sector; a contract research organization can also be classified here, depending on its main activity.

In this sector, use barely moved in a year: 31.7% of businesses had used AI in the second quarter of 2025, and 32.4% in the second quarter of 2026. Intentions, however, surged: 26.3% of businesses planned to use it in the third quarter of 2025, and 54.8% in the third quarter of 2026, the highest share of any sector. In Quebec, the estimate is higher still (65.7%), but less precise, as the sample is smaller. And these intentions have already been exceeded two years running: 24.6% planned in 2024, 31.7% users in 2025; 26.3% planned in 2025, 32.4% in 2026. Among the sector’s businesses that plan to use AI, the applications cited most are large language models (54.9%) and data analytics (40.5%).

Workers are not waiting. In March 2026, according to questions added to the Labour Force Survey, 35.9% of Canadian workers had used generative AI tools at work over the previous 12 months; in professional, scientific and technical services, 65.6%. The two surveys do not measure the same thing: one counts workers, the other businesses. But in your sector, two workers in three already use these tools, and nothing says they all do so with a tool their employer chose.

No official statistic counts the clients a firm loses to a better-equipped competitor. What follows is therefore an argument, supported by measured facts: a firm sells qualified time, and the cost of a file is mostly made of hours.

• Cost per file. When reading documents, filing them and producing a first draft take fewer hours, the same file costs less. The competitor can lower its fees, or keep its margin and reinvest it.
• Response time. The firm that sends a first analysis the next day changes what the client considers normal. The others have not become slower; they have become the slowest in the market.
• Peak-period capacity. Tax season, year-end closing, a regulatory filing: capacity depends on headcount. In October 2026, the Bank of Canada reported that several firms are counting on automation and other efficiency gains to meet higher sales without hiring.
• Consistent quality. A tool-assisted reading, where every value points to its source and a person decides the uncertain cases, depends less on the hour and on who handles the file. Fewer discrepancies mean less rework.

These gains come from reorganized work, not from the tool alone. Among businesses that plan to use AI, 48.7% intend to train their employees and 47.5% to develop new work processes. And, among these same businesses, 68.1% expect their total employment to stay the same, 10.8% expect a decrease and 9.8% an increase: most do not plan to change their headcount.

Waiting has a cost that never shows up on an invoice: learning put off until later. Among businesses that used AI, 44.4% changed their training or staffing practices. Bank of Canada economists point out that, for a general-purpose technology, the effects come with a lag: the time it takes to learn and adjust. Those who start late pay for that lag when others have already paid for it.

Getting it wrong is costly too, in two ways. The first is AI for everyone: each person with their own tool, and no rules. In a Bank of Canada survey of 314 business leaders (December 2025, not representative of all businesses), more than two thirds said they personally use AI at work, but only 8% of the businesses made significant use of it in their operations. Personal use does not change a process; it sends information to tools nobody has assessed: that is shadow AI.

The second is AI for everything: handing a large language model what a formula, a rule or a document reader does better. A language model estimates; a formula computes. Yet in professional services, large language models are precisely what businesses plan to use first: all the more reason to choose the technique task by task. Choosing the technique is the subject of a separate article: AI is more than large language models.

A good first project has four traits: a frequent task, a result you can check, an error that can be caught before it reaches the client, and data you are entitled to use that way.

1. Read the documents, and have a person validate whatever is uncertain

Tax slips, exhibits in a litigation file, certificates of analysis: optical character recognition (OCR, which turns the image of a page into text), followed by extraction of the useful fields, does most of the work. Each value comes with its position on the page and a confidence score; below a threshold set in advance, a person checks it. When OCR is enough, and when a generative model is worth its cost: see our article on reading documents.

What to measure: minutes per file, the share of fields corrected at review, errors found afterwards. Prerequisites: a hundred or so real documents, the list of fields, and what Quebec’s Law 25 requires before personal information goes into an AI system: a privacy impact assessment, a written contract with the vendor and, if the information leaves Quebec, an assessment that also considers the legal framework where it is sent. First step: time twenty files this week.

2. Stop re-keying from one application to another

Copying a new client from the intake form into the practice management software, then into billing: this work often needs no AI at all. An automated flow, or a software robot when the application offers no programming interface, starts from data that has already been validated, enters it, reads back what it has just entered and flags discrepancies. What to measure: hours of re-keying per week, discrepancies found further down. Prerequisites: a stable process, a validated source, the terms of use of each software publisher. First step: for one week, everyone notes what they copy from one screen to another. On Microsoft 365, see what is possible for quotes and invoices.

3. A supervised internal assistant, with a written usage rule

A first draft of a letter, a summary of a long exchange, a clause to find in your own templates: an AI assistant chosen by the business does these jobs, within a framework. The Canadian Centre for Cyber Security recommends a plan that identifies policies on how AI should be used, oversight and review processes, and avoiding personal or sensitive corporate data in prompts.

What to measure: time spent on first drafts, the share of texts that need deep rework, incidents reported. Prerequisites: a tool that has been chosen and covered by a contract, a one-page rule, a short training session. First step: write the rule, then run a trial with five volunteers.

1. 1

   Take inventory

   For one week, list the repetitive tasks: reading, filing, copying, checking.

2. 2

   Score each task

   Volume, time, the cost of an error that slips through, data sensitivity. Start with high volume, recoverable errors and low-sensitivity data.

3. 3

   Choose the technique

   For each task, the simplest technique that does the job (see the grid below). AI only where it is needed.

4. 4

   Try, measure, extend

   A limited trial, on real cases, compared with the starting measurement. Extend what has proven itself, stop the rest.

• Ask three people on the team which AI tools they already use, and for what.
• List the week’s repetitive tasks, scored on volume, time, the cost of an error and data sensitivity.
• Pick a single project, and measure the starting point before changing anything.
• Check with the person in charge of the protection of personal information (your privacy officer) what the project requires.
• Write a one-page usage rule for AI assistants, and say which tool is approved.
• Set the date when you will compare the measurements, and the rule for extending or stopping.