How to use AI when you actually run a business.

Most advice about using AI is written for people with nothing on the line. Try everything, break things, see what sticks. That's fine if you're a student. It's terrible advice if a bad output goes out on your letterhead.

Here's the version for someone with employees, customers and a quoting deadline.

Where the Canadian market actually is

Worth knowing before you decide you're behind.

Canadian businesses using AI to produce goods or deliver services: 6.1% in Q2 2024 → 12.2% in Q2 2025 → 19.2% in Q2 2026. Tripled in two years (Statistics Canada).

Now the number nobody quotes: in the same survey, 40.0% of businesses said AI simply isn't relevant to what they make or do. That's twice as many as the ones using it. Cybersecurity and privacy worries came in at 13.4%, cost at 10.6% (Statistics Canada, Q2 2026).

The spread by industry is enormous:

Sector	Using AI, Q2 2026
Information and cultural industries	42.3%
Finance and insurance	40.4%
Professional, scientific and technical services	32.4%
Construction	9.2%
Wholesale trade	7.9%
Agriculture, forestry, fishing and hunting	4.5%

And geography matters more than most people expect: 21.0% urban versus 9.9% rural. If you're running a business outside a major centre, the number of your peers doing this is less than half the headline.

One more, because it cuts against the usual story: businesses with 100+ employees sit at 27.8%, while businesses with 1 to 4 employees sit at 19.9%: barely below the national average. The very smallest firms are not the laggards here. Mid-sized ones are.

Two numbers that disagree, and why you should care

You'll see a much rosier figure quoted around. BDC surveyed 1,500 Canadian business owners in February 2026 and reported that 30% of SMEs use generative AI, and that those who do are 24% more productive (BDC).

Both numbers can be true: StatCan is counting AI used in producing goods or delivering services, BDC is counting any generative AI use, which includes someone drafting an email. They're measuring different things.

The 24% productivity figure deserves more caution. It's self-reported by business owners, in a survey run by a bank whose stated mission includes getting SMEs to adopt AI faster. That doesn't make it wrong. It does mean it isn't the same class of evidence as the adoption counts, and you shouldn't build a budget on it.

We'd rather tell you that than quote the impressive number and hope you don't check.

The one distinction that matters

Everything else follows from this.

AI is good at work where you can recognize a good answer faster than you can produce one. Drafting a reply you'd need twenty minutes to write and ten seconds to approve. Summarizing a long document you'd otherwise skim. Turning rough notes into something a customer can read. Pulling structure out of a mess. Rewriting the same thing for a different audience.

AI is bad at work where being wrong is expensive and hard to spot. Anything with a number you won't verify. Anything where you'd have to already know the answer to catch the error. Anything that depends on facts about your business it has no way of knowing.

Before you point it at a task, ask one question: if this comes back wrong, will I notice? If the answer is no, that's not the task to start with.

Notice, too, what Canadian businesses actually use it for: data analytics (36.6%), text analytics (34.5%), and chatbots (28.2%). The top two are both "make sense of information we already have." That's the pattern.

Pick one tool and stay there

The comparison shopping is a trap. The leading tools are close enough that the difference between them is smaller than the difference between using one well and using four badly.

Pick one. ChatGPT, Claude or Copilot: if your business already runs on Microsoft 365, Copilot is the path of least friction. Pay for it. The free tiers use older models and, more importantly, handle your data differently: which matters the moment someone pastes anything about a real customer.

Then use it every day for two weeks on real work. You'll learn more from that than from any comparison chart.

How to ask for something you can use

Most disappointing output is a badly-formed request. Four parts fix nearly all of it:

1. Who it's for. "Write to a customer who's been with us six years and is annoyed about a delay" beats "write an email."
2. What you're actually trying to make happen. Not the document: the outcome. "I want him to accept the new date without asking for a discount."
3. The facts it can't guess. Your prices, your timelines, your constraints, the history. It knows nothing about your business unless you tell it.
4. What good looks like. Length, tone, format. Better still: paste something you wrote before that worked, and say "like this."

Then, the part people skip, tell it what's wrong and ask again. The first output is a starting point, not a deliverable. "Too formal, cut it in half, and don't apologize twice" gets you further than rewriting the request from scratch.

Three things it will get wrong

It's confidently wrong. There is no difference in tone between an answer it's sure of and an answer it invented. It will produce a plausible statistic, a real-sounding legal citation, a supplier's phone number: none of which exist. Anything factual that leaves your building gets checked by a person.

It doesn't know your business. No access to your prices, your job history, your margins, or what you told that customer last spring. It fills the gap with something reasonable and generic, which in your context is usually wrong.

It doesn't remember. Unless you're using a feature built for it, each conversation starts cold. Yesterday's context is gone. This is why the "standing instructions" settings in every major tool are worth the ten minutes.

What never goes in

Set this rule before you need it, because your team is already making the call without you.

The legal picture in Canada is worth understanding, because it isn't what most people assume. There is no federal AI statute in force. The Artificial Intelligence and Data Act died with Bill C-27 when Parliament was prorogued, and it has not been replaced (Schwartz Reisman Institute, University of Toronto).

What does apply is privacy law, and it applies fully. Federally that's PIPEDA; in Alberta and British Columbia, provincial PIPA; in Quebec, Law 25. Health information has its own rules in every province.

In December 2023, the Privacy Commissioner of Canada and 12 provincial and territorial privacy authorities jointly issued nine principles for generative AI, updated May 2025 (OPC). Three of them turn into practical rules for a small business:

• Limiting collection and use: where personal information goes into a prompt, do it only where you're authorized to.
• Accuracy: take reasonable steps to make sure outputs are accurate enough for what you're using them for.
• Openness: tell people when a generative AI tool forms part of a decision that affects them.

That last one catches businesses off guard. If AI is screening résumés or scoring applications, the obligation to be transparent about it is already live: no new AI law required.

The practical version: if you wouldn't post it publicly, don't paste it into a tool you haven't checked the terms on.

Where to start this week

Not with a project. With one task.

Pick something that happens every week, takes 20 to 60 minutes, produces text, and where you'd immediately spot a bad result. Quote follow-ups. Job summaries. The same six customer questions you answer over and over. Turning site notes into a report.

Do that one task with AI for two weeks. Time it before and after. Then decide whether to widen it.

That's the whole method, and it's deliberately unambitious. The businesses that get burned start with an ambitious build before they know what the tool is good for in their shop. The ones that get results start with a task nobody would put in a press release.

The part worth saying plainly

The most common mistake we see isn't picking the wrong tool. It's pointing a good tool at a broken process.

If your quoting takes four days because the information lives in three places and one person's head, AI won't fix that: it'll produce a bad quote faster. Sort out how the work flows first. Then automate the part that's genuinely repetitive.

And if you're in one of those 4.5% or 9.2% sectors, take some comfort: your competitors aren't quietly running circles around you. Nobody in your trade has this figured out yet. That's an opening, not a deficit, but only if you start with a real problem instead of a tool you saw in an ad.

Sometimes the right answer costs nothing and involves no software at all. That's not a reason to skip the exercise. It's usually the best return in the building.

---

Modus is an operations and AI consulting firm based in Sherbrooke, Quebec, serving founder-led businesses across Canada. We diagnose how a business runs, fix the most expensive friction, and use AI only where the return is real.