In brief
23claims the report does not make, when just asked
Given a short request, a language model fills every unstated detail with its most probable guess. Asked to turn a report into social posts, the model guessed the voice, the length and the conclusions; the figures were mostly right, but the answer made 23 claims the report does not.
36 of 36figures exact, when instructed
The durable part of prompt engineering is specification: the purpose and audience, the material, the task, its constraints, the output format, and a definition of done. With all six, the same model used 36 figures, every one exact, and made no claim the report does not support.
6parts in an instruction, the unit every later post builds on
Every later post in this series builds on that unit. A skill packages an instruction, an agent follows one, and a goal is an instruction with a test attached.
Just asking
Most interactions with a language model start with a typed request in a chat window. In professional settings, these requests most often involve writing: OpenAI’s study of ChatGPT usage found that “writing dominates work-related tasks” 1. One such task is turning a single source into posts for several channels. To see what a one-line request produces, I gave an AI assistant a report and the request in Figure 1:
Turn this report into posts for LinkedIn, my newsletter and a short video.
LinkedIn post
AI use among Canadian businesses doubled in one year. Jobs mostly stayed the same.
New Statistics Canada data (Q2 2025):
📈 12.2% of businesses used AI to produce goods or deliver services, up from 6.1% a year earlier.
But adoption is uneven:
→ Information & cultural industries: 35.6%
→ Professional, scientific & technical services: 31.7%
→ Finance & insurance: 30.6%
→ Accommodation & food services: just 1.5%
The most common uses are practical rather than flashy. Text analytics (35.7%) leads, followed by data analytics (26.4%) and chatbots (24.8%).
On jobs:
✅ 89.4% of AI-using businesses reported no change in headcount.
🔽 Only 4.3% added staff because of AI, down from 8.8%.
📋 Most said AI reduced employee tasks only to a "small extent" (47.2%).
The real change is in how work gets done. 40.1% of AI users built new workflows, and 38.9% trained existing staff.
What's next? 17.9% of businesses plan to adopt AI software in the next 12 months, up from 11.5%.
And the reality check: 41.2% of businesses say AI investment isn't relevant to them at all.
AI in Canada is growing fast, but from a small base, and it's reshaping workflows more than workforces.
Where does your industry sit on this curve? 👇
#ArtificialIntelligence #Canada #FutureOfWork #DigitalTransformation #SmallBusiness
Then a 551-word newsletter and a video script. In the whole answer: 23 claims the report does not make, 6 emojis and 8 hashtags.
Statistics Canada’s report analyzed how Canadian businesses used artificial intelligence in the second quarter of 2025 2. On 11 October 2026, I sent the request to Claude Opus 5.5 at medium effort, the model’s default level, and it returned a LinkedIn post, a newsletter, and a video script in 29 seconds. Each piece was well written, and across the three, 68 of 83 figures were reported exactly, with 12 more correctly rounded or reworded.
Where my request lacked specifics, the model relied on the conventions of the genre. The LinkedIn post began, “AI use among Canadian businesses doubled in one year. Jobs mostly stayed the same,” and included five emojis and five hashtags. The newsletter ran to 551 words. Asked for a short video, it wrote a script for one, without saying that it had changed the request. Across its answer, the model drew 23 conclusions the report does not make, such as “So AI is not wiping out jobs at the business level” and “it’s reshaping workflows more than workforces,” and it ended the script with “Is your business keeping up?” Two figures described the wrong group; for instance, the 89.4% of businesses using AI that reported no change in employment became “89% of businesses.” None of this was careless: the model followed the usual practice for such content because I had not specified my requirements.
The gap between what I meant and what I wrote was filled with assumptions. Prompt engineering narrows that gap before the model has to guess.
What prompt engineering is
The term originated with the first large language models, when early guidance emphasized wording, such as specifying a role, adding key phrases, or rephrasing to improve results. Some adjustments proved effective. For example, Kojima et al. 3 found that adding “Let’s think step by step” to a question increased a model’s accuracy on the GSM8K arithmetic benchmark 4 from 10.4% to 40.7%. These phrases worked by influencing the model’s reasoning process before it answered. Current models often perform this reasoning without being asked.
Clear specification remains essential. One useful test is to imagine giving the request to a capable colleague unfamiliar with my work. They would likely ask about the audience, desired length, appropriate tone, and whether they could include their own opinions. While a model might ask similar questions, it often proceeds based on assumptions, as this one did. Effective instructions address these questions in advance. Anthropic’s documentation refers to this test as its golden rule 5:
Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.
Anthropic, Prompting best practices
From prompt to context
A model knows nothing of my earlier posts, my readers, or past conversations unless that information is placed in its context window. Beyond what it learned in training, it works only from what is in that window: the instruction, any attached documents, and the conversation so far. Anthropic’s engineers refer to this broader field as context engineering, defined as “the set of strategies for curating and maintaining the optimal set of tokens” a model uses, and describe it as “the natural progression of prompt engineering” 6.
This example highlights two key points. First, the source must be included in the context and clearly identified as the only source; otherwise, the model may use prior knowledge. Second, adding more material does not always improve results. The same article describes “context rot,” in which increasing the amount of context reduces the model’s ability to recall information accurately. It recommends using “the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome” 6. In this case, that set consists of the report’s text and tables, about 2,400 words, rather than the entire web page with menus and links.
The parts of a good instruction
Reliable instructions tend to share a few parts, none of which depends on special phrasing.
- 1
Purpose and audience
What the output is for and who will read it.
The readers are managers and specialists evaluating AI adoption: busy, skeptical of hype, and often reading on their phones. These factors guide the tone and what to leave out.
- 2
Material
What the answer must be based on.
The report, identified in the instruction as the only source.
- 3
Task
What to produce, and how much of it.
A LinkedIn post of at most 1,300 characters, a newsletter section of 250 to 350 words with a headline, and a script for a 60-second video of at most 150 spoken words.
- 4
Constraints
What must hold for the result to be usable.
Every figure appears exactly as the report gives it, with what it measures and the period it covers; no claim goes beyond the report; and the language is plain, with no emojis or exclamation marks, at most two hashtags, and Statistics Canada named in every piece.
- 5
Output format
What form the result takes.
Three sections headed LINKEDIN, NEWSLETTER and VIDEO, then a list of every figure used with the sentence of the report it comes from, and nothing else.
- 6
Definition of done
How the model, and I, can tell the work is finished.
Each piece is within its length limit and every figure has been checked against the report, with anything that fails fixed before the answer is returned.
The definition of done is often overlooked. Without it, the model stops when the text appears complete. With it, both the model and I have a clear standard for evaluation. This concept will be revisited later in the series, where it forms the basis for agents that pursue goals instead of following step-by-step instructions. Figure 2 highlights the six parts of a complete instruction.
Examples
Demonstrating is often more effective than describing. Brown et al. 7 showed that a large model can perform new tasks from a few demonstrations “specified purely via text interaction with the model,” without additional training, a method now called few-shot prompting. A well-chosen example communicates format, length, tone, and detail more precisely than a descriptive paragraph. Current vendor guides support this: Google’s Gemini documentation recommends always including few-shot examples 8, and Anthropic suggests providing three to five 5.
Examples also present a risk: the model may treat them as templates. For instance, if every example post begins with a question, most new posts are likely to follow suit. The recommended solution is variety. OpenAI’s guide advises providing “a diverse range of possible inputs with the desired outputs” 9, while Anthropic’s engineers recommend curating “a set of diverse, canonical examples” 6 instead of an exhaustive list of edge cases.
The instruction in this post uses no examples; it describes the desired voice instead, which was sufficient for a measured tone. Examples are most valuable when a standard is easier to demonstrate than to explain. Two or three of an author’s previous posts often convey personal voice more effectively than any description.
Asking for reasoning
Wei et al. 10 showed that including worked reasoning steps in the examples, which they called chain-of-thought prompting, improved performance on arithmetic, commonsense, and symbolic reasoning tasks. Kojima et al. 3 then obtained a similar effect without any examples, using the single sentence “Let’s think step by step.”
Many current models are designed to reason before answering, and some allow users to control the amount of reasoning, a setting called effort or a thinking budget depending on the vendor. This choice involves time and financial costs, which the next post in this series examines.
Vendor guidance has evolved alongside these models. OpenAI’s guide notes that reasoning models perform better with high-level guidance than with detailed step-by-step instructions, comparing such a model to a senior colleague: “You can give them a goal to achieve and trust them to work out the details” 9. However, a goal remains a specification. The less an instruction defines the method, the more it must clarify the purpose, the material, and what counts as done.
Instructions can also focus reasoning on verifiable elements. For example, asking the model to check every figure against the report before answering is more useful than a general request to think carefully, as it specifies what to verify.
The same request, instructed
Figure 2 shows the instruction I prepared for the same task, with each part labeled. The report appears first, followed by the instruction, as Anthropic’s documentation recommends for long inputs 5.
<report> ... the Statistics Canada report, about 2,400 words with its tables ... </report>MaterialI write about technology and work for professionals in Canada, on LinkedIn, in a monthly newsletter, and in short videos. My readers are managers and specialists deciding whether and how to use AI. They are busy, skeptical of hype, and many read on their phones.Purpose and audience
Above is a Statistics Canada report on how Canadian businesses use artificial intelligence. It is the only source for this task.Material
From it, write three pieces: - a LinkedIn post of at most 1,300 characters; - a newsletter section of 250 to 350 words, with a headline; - a script for a 60-second video, at most 150 spoken words.Task
Requirements: - Every figure must appear in the report exactly as written there, with what it measures and the period it covers. - Make no claim the report does not support: no causes, forecasts or advice it does not state. - Use plain, measured language: no emojis, no exclamation marks, and at most two hashtags, only in the LinkedIn post. - Name Statistics Canada as the source in each piece.Constraints
Return three sections headed LINKEDIN, NEWSLETTER and VIDEO, then a section headed FIGURES that lists every figure you used, each with the sentence of the report it comes from. Return nothing else.Output format
Before you return, check each piece against its length limit and every figure against the report, and fix anything that fails either check.Definition of done
The instruction is 229 words, while the request is 14. Each added detail clarifies expectations, including the audience, valid sources, length limits, permissible claims, the required format, and what counts as done. The language is direct, like a brief for a capable colleague.
When sent to the same model, the instruction took 90 seconds and produced the four requested sections without extraneous content. The LinkedIn post was 1,221 characters, and the newsletter section was 316 words. All 36 figures were exact, each with the correct measure and period, and all 17 sentences in the list of figures were quoted verbatim from the report. There were no emojis, only two hashtags, and no unsupported claims. Where the first answer claimed AI was “rewiring how they work,” this one reported that “the most reported change was developing new workflows (40.1%).” Figure 3 sets the two answers side by side, sentence by sentence.
Just asked
One line, with the report attached
Instructed
The same request, in a 229-word instruction
How to read it. Each answer reads left to right, one segment per sentence, as long as the sentence, so a longer tape is a longer answer. The color is what the check found, and orange is what to avoid.
- Carries figures, all exact
- A claim the report does not make
- A wrong or unsourced figure
- Other text
- An emoji
Same model, same report: 23 claims the report does not make when just asked, and none when instructed.
Show the comparison as a table
| Measure | Just asked | Instructed |
|---|---|---|
| LinkedIn post | 1,338 characters, 5 emojis, 5 hashtags | 1,221 characters (limit 1,300), no emojis, 2 hashtags |
| Newsletter | 551 words | 316 words (asked for 250 to 350) |
| Video script | 122 words; about 135 read aloud | 144 words (limit 150); about 167 read aloud |
| Figures used | 83 | 36 |
| Exact as the report gives them | 68, and 12 correctly rounded | 36 |
| Wrong measure, or not in the report | 3 | 0 |
| Claims the report does not make | 23 | 0 |
| Time to answer | 29 seconds | 90 seconds |
The instructed answer had some flaws worth noting. It referred to the release as “new Statistics Canada data,” though the report was sixteen months old and the model had the current date. The video script met the 150-word limit in writing, but read aloud, with “12.2%” spoken as “twelve point two percent,” it reached about 167 words and ran longer than a minute. The definition of done counted words, though duration was what mattered. A definition of done must measure what matters, and checking it by hand is slow; both issues return in the posts on agents and goals.
Three lessons emerge. First, the instruction made the answer usable as written: within its limits, in the requested form, and verifiable line by line. Second, the risk in the first answer was not its figures, which were mostly correct even without the instruction, but the claims it added; those were what needed constraining. Third, a written instruction cannot verify its own result. An agent that counts spoken seconds and compares each sentence with the source can, and that is the subject of the post on agents.
Writing the instruction took much longer than writing the request. This effort is justified when the result is important, but it becomes burdensome if the task recurs weekly. Each new report, talk, or article requires the same approach, and retyping the instruction each time would be tedious and increase the risk of errors.
What the rest of the series builds on
Every later step in this series is built from instructions like the one in Figure 2. Each instruction goes to a model at some effort level, a choice that applies at every step. A skill packages an instruction so that it need not be rewritten. An agent follows one while choosing its own steps and tools, and, given a test for done, it can pursue a goal until the test passes. Sub-agents each receive an instruction as their brief, and a workflow coordinates many of them within a fixed structure. A vague instruction weakens every one of these steps, and the cost of that vagueness grows with the system built on it.
The next post takes up that choice. Both runs here used one model at one effort level; the next runs the instruction in Figure 2 on several models and at each effort level, scores every answer with the same checks, and shows what each choice changes in accuracy, time, and cost.
About the experiment
Both requests were sent once to Claude Opus 5.5 on 11 October 2026, at medium effort, the model’s default 11, with no tools and no earlier conversation. Effort sets how much work the model puts into a response, including how long it reasons, so another level could change either answer. Every figure in both answers was checked against the report, every claim was read against it, and lengths were counted by script. One run of each request shows what can happen, not how often it happens; both answers are reproduced in full in the appendix.
References
11 sources: 4 papers, 4 documentation pages, 1 working paper, 1 report, 1 article
Show all 11Hide
- 1
How people use ChatGPT
A. Chatterji, T. Cunningham, D. J. Deming, Z. Hitzig, C. Ong, C. Y. Shan, and K. Wadman
National Bureau of Economic Research, Working Paper 34255, Sep. 2025
Working papernber.org - 2
Analysis on artificial intelligence use by businesses in Canada, second quarter of 2025
V. Bryan, S. Sood, and C. Johnston
Analysis in Brief, Statistics Canada, Catalogue no. 11-621-M, Jun. 16, 2025Adapted under the Statistics Canada Open Licence
Reportstatcan.gc.ca - 3
Large language models are zero-shot reasoners
T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, and Y. Iwasawa
in Advances in Neural Information Processing Systems 35 (NeurIPS 2022), 2022
- 4
Training verifiers to solve math word problems
K. Cobbe, V. Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, C. Hesse, and J. Schulman
OpenAI, 2021. Dataset: GSM8K
Papergithub.com - 5
Prompting best practices
Anthropic
Claude Developer Platform documentationaccessed Oct. 11, 2026
Documentationplatform.claude.com - 6
Effective context engineering for AI agents
P. Rajasekaran, E. Dixon, C. Ryan, and J. Hadfield
Anthropic Engineering, Sep. 29, 2025
Articleanthropic.com - 7
Language models are few-shot learners
T. Brown et al.
in Advances in Neural Information Processing Systems 33 (NeurIPS 2020), 2020
Paperpapers.neurips.cc - 8
Prompt design strategies
Google
Gemini API documentationaccessed Oct. 11, 2026
Documentationai.google.dev - 9
Prompt engineering
OpenAI
OpenAI API documentationaccessed Oct. 11, 2026
Documentationdevelopers.openai.com - 10
Chain-of-thought prompting elicits reasoning in large language models
J. Wei et al.
in Advances in Neural Information Processing Systems 35 (NeurIPS 2022), 2022
- 11
Effort
Anthropic
Claude Developer Platform documentationaccessed Oct. 12, 2026
Documentationplatform.claude.com
Appendix: the two answers in full
Both answers in full and word for word, with their Markdown formatting shown as it renders.
The one-line request1,166 words, verbatim
LinkedIn Post
AI use among Canadian businesses doubled in one year. Jobs mostly stayed the same.
New Statistics Canada data (Q2 2025):
📈 12.2% of businesses used AI to produce goods or deliver services, up from 6.1% a year earlier.
But adoption is uneven:
→ Information & cultural industries: 35.6%
→ Professional, scientific & technical services: 31.7%
→ Finance & insurance: 30.6%
→ Accommodation & food services: just 1.5%
The most common uses are practical rather than flashy. Text analytics (35.7%) leads, followed by data analytics (26.4%) and chatbots (24.8%).
On jobs:
✅ 89.4% of AI-using businesses reported no change in headcount.
🔽 Only 4.3% added staff because of AI, down from 8.8%.
📋 Most said AI reduced employee tasks only to a "small extent" (47.2%).
The real change is in how work gets done. 40.1% of AI users built new workflows, and 38.9% trained existing staff.
What's next? 17.9% of businesses plan to adopt AI software in the next 12 months, up from 11.5%.
And the reality check: 41.2% of businesses say AI investment isn't relevant to them at all.
AI in Canada is growing fast, but from a small base, and it's reshaping workflows more than workforces.
Where does your industry sit on this curve? 👇
#ArtificialIntelligence #Canada #FutureOfWork #DigitalTransformation #SmallBusiness
Newsletter
Subject line: AI adoption in Canada doubled. Here's what actually changed.
Preview text: Spoiler: workflows, not headcount.
Statistics Canada released new data from its Canadian Survey on Business Conditions (collected April–May 2025). It is one of the clearest pictures we have of how Canadian businesses are actually using AI, as opposed to how they talk about it.
Here are the five things worth knowing.
1. Adoption doubled, but it's still a minority
12.2% of Canadian businesses used AI to produce goods or deliver services in the past 12 months. A year earlier, the figure was 6.1%.
Doubling in a year is significant. Still, almost nine in ten businesses are not yet using AI in their operations.
2. The gap between industries is huge
| Industry | Share using AI |
|---|---|
| Information & cultural industries | 35.6% |
| Professional, scientific & technical services | 31.7% |
| Finance & insurance | 30.6% |
| Agriculture, forestry, fishing & hunting | 1.8% |
| Transportation & warehousing | 1.8% |
| Accommodation & food services | 1.5% |
Knowledge-work sectors are moving quickly. Sectors built on physical work have barely begun. The same three leaders also topped the list in 2024, so the gap is widening rather than closing.
3. The everyday tools are winning
Among AI users, the top applications were:
- Text analytics: 35.7% (up from 27.0%)
- Data analytics: 26.4%
- Virtual agents/chatbots: 24.8%
- Marketing automation: 23.1% (up from 15.2%)
Some categories declined. Image and pattern recognition fell from 21.8% to 11.4%, and natural language processing fell from 28.9% to 23.1%. Businesses appear to be focusing on tools that solve immediate, practical problems, such as making sense of emails, reviews and survey responses.
4. Jobs: mostly unchanged so far
This is the part most people want to know about.
- 89.4% of AI-using businesses reported no change in total employment (up from 84.9%).
- 6.3% reported a decrease, the same as last year.
- 4.3% reported an increase, down from 8.8%.
On tasks, 47.2% said AI reduced work previously done by employees only to a small extent, and just 5.3% said to a large extent.
So AI is not wiping out jobs at the business level. It is also creating fewer new roles than it did a year ago. Hiring AI-trained staff fell from 8.2% to 4.9%.
5. The real change is in how work is organized
What businesses changed after adopting AI:
- Developed new workflows: 40.1% (up from 35.2%)
- Trained current staff: 38.9%
- Purchased cloud services or storage: 25.7% (up from 16.1%)
- Worked with vendors or consultants: 18.2% (up from 11.3%)
Businesses are upskilling the people they already have instead of hiring new ones, and they are increasingly relying on outside help to integrate AI.
Looking ahead
17.9% of businesses plan to adopt AI software in the next 12 months, up from 11.5%. Health care and social assistance stands out: planned adoption rose from 10.8% to 25.5%.
Many businesses remain unconvinced. Only 8.3% call AI investment "very important," while 41.2% say it is "not relevant" to their operations.
My take
AI in Canada is following a familiar pattern for new technology. Knowledge-heavy industries adopt it first, the uses are practical, and the biggest change is in how work is organized rather than how many people do it. For most businesses, the question has moved past "Will AI replace my team?" to "Have we redesigned our workflows and trained our people to use it well?"
Source: Statistics Canada, "Analysis on artificial intelligence use by businesses in Canada, second quarter of 2025" (June 16, 2025). Survey of 9,103 businesses.
Short Video Script (~60 seconds)
Format: Vertical (9:16), talking head with bold on-screen text and simple animated charts.
| Time | Voiceover | On-screen |
|---|---|---|
| 0–3s | "AI use in Canadian businesses just doubled. But that's not the interesting part." | AI USE: 6% → 12% (big animated jump) |
| 3–12s | "According to Statistics Canada, 12.2% of businesses now use AI, up from 6.1% a year earlier." | "12.2% of Canadian businesses use AI" / Source: StatCan |
| 12–22s | "But it depends heavily on your industry. Over 30% of tech, professional services, and finance firms use it. Restaurants and hotels? Just 1.5%." | Bar chart: Info & Culture 35.6% / Prof. Services 31.7% / Finance 30.6% / Food & Accommodation 1.5% |
| 22–30s | "And the top use isn't robots or deepfakes. It's text analytics: making sense of emails, reviews and surveys." | "#1 use: Text analytics (35.7%)" |
| 30–42s | "Now the big question: jobs. Almost 90% of businesses using AI said their headcount didn't change at all." | 89.4% — NO CHANGE IN JOBS |
| 42–52s | "What did change? Workflows. 40% built new ones, and nearly 39% trained their existing staff." | "40% → new workflows" / "39% → trained staff" |
| 52–60s | "So AI isn't replacing teams yet. It's rewiring how they work. Is your business keeping up? Follow for more." | "Workflows > Workforces" / Follow CTA |
Caption: AI adoption in Canada doubled in a year, but 89% of businesses say it didn't change their headcount. The real shift is in how work gets done. 📊 Source: Statistics Canada, Q2 2025. #AI #Canada #FutureOfWork
Note: This report was released in June 2025 and covers Q2 2025 data. Since it's now October 2026, Statistics Canada may have published a newer quarter. Check before posting so you don't present older numbers as the latest. I've labelled everything "Q2 2025" so the content stays accurate either way.
The instruction1,356 words, verbatim
How many Canadian businesses use AI? New Statistics Canada data gives a measured answer.
In its survey collected from April 1 to May 5, 2025, 12.2% of businesses reported using AI to produce goods or deliver services over the previous 12 months. In the second quarter of 2024, the share was 6.1%.
Use is uneven. In the 2025 survey, 35.6% of businesses in information and cultural industries reported using AI. So did 31.7% in professional, scientific and technical services and 30.6% in finance and insurance. In accommodation and food services, it was 1.5%.
On jobs: of businesses that used AI, 89.4% reported no change to their employment levels after implementation. The figure was 84.9% in the second quarter of 2024.
Among AI users, the most reported change was developing new workflows (40.1%). Another 38.9% trained current staff to use AI.
Meanwhile, 41.2% of all businesses said investment in AI was "not relevant" to their operations.
In short, use was higher than a year earlier, gaps between industries were wide, and most AI users reported no change to employment levels.
Source: Statistics Canada, Canadian Survey on Business Conditions, second quarter of 2025.
#ArtificialIntelligence #FutureOfWork
NEWSLETTER
AI use by Canadian businesses: rising, uneven, and little change to employment
Statistics Canada has released new figures on how Canadian businesses use artificial intelligence. They come from its Canadian Survey on Business Conditions, collected from April 1 to May 5, 2025.
Use is rising. In all, 12.2% of businesses reported using AI to produce goods or deliver services over the 12 months before the survey. That was up from 6.1% in the second quarter of 2024.
It is concentrated. AI use was reported by 35.6% of businesses in information and cultural industries, 31.7% in professional, scientific and technical services, and 30.6% in finance and insurance. It was 1.5% in accommodation and food services.
Text leads. Among businesses using AI, the most common application was text analytics, at 35.7%. Statistics Canada notes it may be used to process insights from sources such as customer reviews, emails or survey responses.
Employment mostly unchanged. Of businesses that used AI, 89.4% reported no change to their employment levels after implementation. The figure was 84.9% in the second quarter of 2024. The share reporting an increase in employment due to AI use fell from 8.8% to 4.3%. Businesses were also asked how far AI had reduced tasks previously done by employees, and 47.2% said to a small extent.
Workflows change. Among AI users, the most reported change was developing new workflows (40.1%), followed by training current staff to use AI (38.9%).
Many see AI investment as not relevant. Across all businesses, 8.3% said investment in AI was very important to their operations, while 41.2% said it was not relevant. Looking ahead, 17.9% reported plans to adopt AI software over the next 12 months. That was up from 11.5% in the second quarter of 2024.
For managers weighing AI, the survey describes a practice that is spreading but is still reported by a minority of businesses, with wide gaps between industries.
VIDEO
New Statistics Canada data shows how Canadian businesses are using AI.
A survey was collected from April 1 to May 5, 2025. In it, 12.2% of businesses said they had used AI to produce goods or deliver services in the previous 12 months. In the second quarter of 2024, it was 6.1%.
Use varies widely. In information and cultural industries, 35.6% of businesses reported using it. In accommodation and food services, 1.5%.
What about jobs? Among businesses using AI, 89.4% reported no change to their employment levels after implementation.
Their most reported change was developing new workflows, at 40.1%.
And 41.2% of all businesses said investment in AI was not relevant to their operations.
In short: more use than a year earlier, wide gaps between industries, and most AI users reporting no change to employment levels.
Source: Statistics Canada, Canadian Survey on Business Conditions.
FIGURES
- April 1 to May 5, 2025 (survey collection period): "From April 1 to May 5, 2025, representatives from businesses across Canada were invited to complete an online questionnaire about business conditions and business expectations moving forward."
- 12.2% and 12 months (businesses using AI to produce goods or deliver services, second quarter of 2025): "In the second quarter of 2025, 12.2% of businesses reported having used AI to produce goods or deliver services over the 12 months preceding the survey."
- 6.1% (same measure, second quarter of 2024): "This was an increase from the 6.1% reported in the second quarter of 2024,[2] highlighting the growing presence of AI in Canadian business operations."
- 35.6%, 31.7%, 30.6% (AI use by industry, second quarter of 2025): "Among businesses that reported using AI in producing goods or delivering services over the last 12 months (12.2%), those in information and cultural industries (35.6%); professional, scientific and technical services (31.7%); and finance and insurance (30.6%) were most likely to do so."
- 1.5% (AI use in accommodation and food services, second quarter of 2025): "In the second quarter of 2025, AI use was lowest in accommodation and food services (1.5%), followed by agriculture, forestry, fishing and hunting (1.8%) and transportation and warehousing (also 1.8%)."
- 35.7% (text analytics among AI users, second quarter of 2025): "Among businesses that reported having used AI over the last 12 months (12.2%), the most adopted application was text analytics (35.7%)."
- 89.4% (AI users reporting no change to employment levels, second quarter of 2025): "In the second quarter of 2025, of the businesses who reported using AI over the last 12 months (12.2%), the vast majority (89.4%) reported no change to their employment levels after implementation."
- 84.9% (same measure, second quarter of 2024): "This was similar to results from a year earlier, when 84.9% of businesses that used AI in the second quarter of 2024 (6.1%) also reported no change to employment levels."
- 8.8% and 4.3% (AI users reporting an increase in employment due to AI use, second quarter of 2024 and 2025): "Meanwhile, the proportion of businesses that reported an increase in employment levels due to AI use declined from 8.8% in the second quarter of 2024 to 4.3% in 2025."
- 47.2% (AI users reporting tasks reduced to a small extent, second quarter of 2025): "Nearly half (47.2%) of businesses that reported using AI over the last 12 months indicated that tasks were reduced to a small extent, up from 44.1% in the second quarter of 2024."
- 40.1% (AI users that developed new workflows, second quarter of 2025): "Among businesses that used AI in producing goods or delivering services over the last 12 months (12.2%), the most reported change was developing new workflows, reported by 40.1% of businesses, compared to 35.2% in the second quarter of 2024."
- 38.9% (AI users that trained current staff to use AI, second quarter of 2025): "In comparison, training current staff to use AI was reported by 38.9% of businesses in the second quarter of 2025, a proportion that remained relatively unchanged from 2024 (38.5%)."
- 8.3% (businesses rating AI investment very important, second quarter of 2025): "In the second quarter of 2025, 8.3% of businesses in Canada reported that investment in AI was very important to their operations, while 20.1% considered it somewhat important."
- 41.2% (businesses rating AI investment not relevant, second quarter of 2025): "In contrast, many businesses in Canada do not currently consider AI investment to be required for their operations, with 41.2% reporting it to be “not relevant”."
- 12 months (planning horizon for adoption plans): "Businesses were asked whether they plan to adopt software or hardware using AI over the next 12 months."
- 17.9% (businesses planning to adopt AI software, second quarter of 2025): "In the second quarter of 2025, nearly one in five (17.9%) businesses reported plans to adopt AI software, while 6.0% reported plans to adopt AI hardware."
- 11.5% (same measure, second quarter of 2024): "This is an increase from the 11.5% of businesses that reported plans to adopt AI software and 5.0% that reported plans to adopt AI hardware in the second quarter of 2024."