Generative AI writes follow-up emails, summarizes documents and answers specific questions in clean prose. For a Quebec small business, it is the most accessible form of artificial intelligence: no IT project, just a browser. It is also the most misunderstood, because nothing tells a beginner what can safely be delegated to it and what should never be shared with it. This guide covers the basics, with verifiable Canadian figures and simple rules.
1. Why everyone talks about it — and why so many businesses still do nothing
In the second quarter of 2025, 12.2% of Canadian businesses reported having used AI to produce goods or deliver services over the previous twelve months, up from 6.1% a year earlier: the number doubled in a single year, according to Statistics Canada's analytical article based on the Canadian Survey on Business Conditions, conducted April 1 to May 5, 2025 among businesses with employees.
Adoption remains low and highly concentrated: 35.6% in information and cultural industries, 31.7% in professional, scientific and technical services, 30.6% in finance and insurance — but only 1.5% in accommodation and food services. And 41.2% of businesses consider AI investment not relevant to their operations.
This gap usually comes from the same place: nobody connected the technology to a specific task in the business. "AI" is not relevant; summarizing 40 information requests every Monday is. A useful fact for skeptics: 89.4% of businesses using AI report no change in headcount attributable to it. The subject is not replacing your team; it is recovering time.
2. What it does well: three concrete uses
Writing and transforming text. Follow-up emails, answers to frequently asked questions, product descriptions, meeting summaries, first drafts of simple documents. Anything that starts with an idea and ends with reviewed text.
Reading and sorting. According to Statistics Canada, the most common application among businesses using AI is text analytics (35.7% of users): customer reviews, emails, survey responses. Data analytics (26.4%) and virtual agents or chatbots (24.8%) follow. Two of the three most popular applications consist of understanding what arrives, not producing more output.
Answering common questions around the clock. A chatbot connected to your website can handle "what are your hours?" while you work. That is a real use case, but it is also the first level where the stakes harden: the tool speaks in your name. The rest of this guide returns to that.
3. A fictional example: the Monday morning inbox
Consider an entirely fictional case, with no claimed client result. A 12-employee cleaning company receives half a day's worth of emails every Monday morning: estimate requests, schedule changes, billing questions.
The owner adopts a generative AI assistant in the browser, with one rule: nothing outgoing ships without review. The tool summarizes the 32 messages received, prepares a reply draft for each and suggests an estimate template. The owner reviews each draft, corrects details, sends personally.
First incident: the tool mentioned a "satisfaction or money-back guarantee" the company has never offered. Review caught it before sending. That, in miniature, is everything this article defends: the technology prepares, the human approves, and the process only holds if the approval is real.
4. What it does not do: four limits to know before paying
It is not a source of truth. Generative models produce plausible content, not guaranteed content. OWASP lists misinformation produced by a model (LLM09) among the major risks of generative AI applications. Verify every figure, date, name and technical clause before use.
It does not know your business. Without context, it writes generic. Worse: if you paste in confidential details to "feed" it, you may circulate data beyond the purposes intended. The next section covers that risk.
It does not decide. Granting a discount, answering a dispute, a decision affecting an employee remain human decisions. A prepared draft is not a decision made.
It does not act on its own. Generative AI produces content; it is not the agent that booked a visit or sent a follow-up. Those actions require stricter safeguards — we covered the difference between a virtual assistant and an AI agent in earlier articles, and prompt injection among them.
5. Ground rules for risk-free first use
Stay in the browser at first. Copy and paste, with no connection to your invoicing or accounting systems. Automating actions comes after weeks of demonstrated quality — as in our invoice follow-up example.
No identifiable personal data in mass-market tools. Canada's privacy authorities, including Quebec's Commission d'accès à l'information, published principles for generative AI in December 2023: establish legal authority to collect personal information, be open and transparent, and limit disclosure of personal or confidential information. Until you are comfortable with those questions, work without client names, addresses or invoice numbers.
Systematic review of everything outgoing. Every piece of content destined for a customer, supplier or authority is reviewed by a human. No exceptions at first.
Transparency and legal obligations. Innovation, Science and Economic Development Canada published a voluntary code of conduct for generative AI in September 2023: human oversight, transparency, validity and robustness. That code is not a law: it changes none of your existing obligations (PIPEDA federally, Quebec Law 25 provincially).
Prompt injection. As soon as a tool receives external text (client emails, web forms), its behaviour can be manipulated by the incoming content. Text from a client never constitutes, by itself, an administrative authorization.
6. What it costs, from beginner to managed project
Three tiers, from free to structured: first, free consumer services to explore and validate a use case; then professional subscriptions in the tens of Canadian dollars per month once volume becomes regular; finally a turnkey project if you want AI connected to your processes. As an example of verified prices (October 3, 2026, CAD, taxes extra) on pro-ai-agent.com/en/tarifs: a one-time setup of $1,500 with $200 per month maintenance for the Essential plan, $2,500 and $400 per month for the Professional plan, which notably includes automated quotes and invoices. Those amounts cover setup and support; they are neither a quote nor a promise of returns.
The open-source approach on a dedicated server, which we use at PRO-AI-AGENT, eliminates per-use API costs as stated on the pricing page — at the price of a setup investment. Two costs are systematically forgotten: human review time and data governance (who can ask for what). A "cheap" subscription demanding tight review can cost more attention than it saves.
FAQ
Do I need technical staff to use generative AI?
Not to start. A browser and an account are enough to experiment. Technical skills become necessary as soon as you connect the tool to your systems or automate actions.
Does my data train the models?
It depends on the tool and your plan. Read the privacy policy before pasting anything sensitive. As a simple precaution: no identifiable personal information in mass-market tools; dedicated hosting is an option once data becomes genuinely confidential.
What is the difference between generative AI and an AI agent?
Generative AI produces content from a prompt; an AI agent performs actions (sends, bookings, registry updates) within a defined scope. If your pain is writing, start generative; if it is operational and repetitive, the agent is the right subject — provided safeguards are installed first.
What does Canadian regulation say?
A voluntary ISED code of conduct has covered advanced generative AI since September 2023, and privacy authorities published principles for its use in December 2023. That is voluntary guidance: your legal obligations, like Quebec Law 25, apply independently. This article is not legal advice.
How soon will I see a return on investment?
No promised timeline; it depends on the task chosen and your volume. Measurable from the first pilot: time spent on the task, share of usable drafts without correction, errors caught in review. Assess on one month of real data.
Let's define your first use case, in two weeks
Pick one recurring, writing-heavy task, define two or three metrics before starting (time spent, share of usable drafts, corrections), then run a two-week pilot. Expand only if quality is demonstrated. PRO-AI-AGENT supports Quebec small businesses on these projects, from choosing the first use case to integration on a dedicated server.
Describe the first use case you have in mind and we will tell you whether it is ready for generative AI — or worth waiting for better conditions.