Skip to content
Souverain Technologie
FR Contact us
Menu
Back to articles

Saving time with AI in a Quebec SME: where to start

By Vincent Savard Management and SMEs

In the second quarter of 2025, 12.7% of Quebec businesses said they had used artificial intelligence to produce goods or deliver services in the previous twelve months, according to the Institut de la statistique du Québec (publication in French, November 28, 2025). A year earlier, the figure was 9.4%. Among businesses with 1 to 4 employees, the share was 12.2%.

So nearly nine in ten Quebec businesses were not using it in their operations. This article is for the people who run them: where AI saves time today, how to check it in your own business without sinking a quarter into it, what makes these projects fail, and what happens to your documents when you hand them to an AI service.

What businesses already use it for

Among Quebec businesses that used AI, text analytics was the most frequently reported use (56.7%), according to the same ISQ publication. And 37.7% of these businesses said AI had reduced, to a moderate or large extent, tasks previously done by their employees. Across Canada, Statistics Canada (June 16, 2025) reports that 89.4% of businesses using AI had not changed their number of employees because of it. In most cases, AI changed the tasks, not the headcount.

In an SME, the tasks that lend themselves to it have a family resemblance. They come back every day or every week, they follow a predictable format, and a person can check the result in a few seconds. Some typical examples:

  • Sorting incoming email and drafting replies to recurring questions (delivery times, business hours, order status), which the employee reads over before sending.
  • Reading a supplier invoice, pulling out the date, amount, taxes and purchase order number, then preparing the entry in the accounting software.
  • Producing the weekly summary of sales or service calls from data already in your systems.
  • Finding information in the company’s documents: the warranty granted to a customer two years ago, the latest version of a procedure, the price quoted for a similar project.
  • Writing a first draft of a quote, a job posting or a letter, which you then correct.

In each case, AI suggests and a person decides. The gain shows up mostly in what people do with the time freed up. In a typical SME, the accountant who no longer retypes 200 invoices a month can call the customers who are behind on payments. The owner who no longer answers the same question ten times a day has an hour to prepare the quote that matters. You get the repetitive work off your plate so your attention goes where the business gains something.

Start small: one process, one measurement, a few weeks

You do not need to write a company-wide AI strategy to get started. One specific task is enough.

Choose a single process. A good candidate takes time every week, annoys the people who do it and, ideally, contains no sensitive personal information. Entering supplier invoices can be a good first choice. Employee files are not.

Measure before you change anything. For two weeks, record the time spent on the task, the number of cases handled and the errors corrected after the fact. Without that baseline, you will only have impressions at the end of the pilot.

Then run a four- to six-week pilot with the person who does the task today. They know where the exceptions are, and they are the one who will have to live with the tool. Repeat the same measurement as at the start.

Then decide: continue, adjust or stop. Stopping is a valid result. A pilot that shows a task is not worth the effort has cost you only a few weeks.

Why these projects fail

Four situations can derail a pilot, and none of them is mainly technical.

The data is disorganized. If invoices arrive in three different mailboxes and the procedure exists in five versions on the server, the AI will work with that mess. It can only search what it is given, and it will find the wrong version as easily as the right one. Putting the documents in order is often the real first step of the project.

Nobody is responsible for it. A pilot assigned to “everyone” fizzles out in the first busy week. You need one named person, with time set aside in their schedule, to track the measurement and make the call when the tool gets it wrong.

The tool was chosen before the problem. A business subscribes to a service because it gets good press, then looks for something to use it for. In the reverse order, the task dictates the tool, and sometimes the answer is a simple automation with no AI at all.

Nobody checks. A language model produces text that looks right, even when it is not. Where an error is costly (an amount on an invoice, a commitment to a customer), a person has to review the output. Training matters too: according to Statistics Canada, 38.9% of businesses that used AI in the second quarter of 2025 had trained their employees to use it.

Where your documents go

When you upload a document to an online AI service, it is processed on the provider’s servers. Depending on the provider, those servers may be elsewhere in Canada, in the United States or in another country.

If the document contains personal information (a customer’s name and address, an employee file, a complaint), Law 25 applies. The Commission d’accès à l’information (page in French) explains that before communicating personal information outside Quebec, or entrusting a person or body outside Quebec with using or keeping it on your behalf, a business must conduct a privacy impact assessment (PIA). The communication is allowed only if the assessment establishes that the information would receive adequate protection, and it must be the subject of a written agreement. The business remains responsible for the information it has entrusted to others. The obligation comes from section 17 of the Act respecting the protection of personal information in the private sector, and it also covers a communication to another Canadian province.

Read the service’s terms as well: how long your documents are kept and whether they can be used to train the provider’s models. For businesses bound by professional secrecy, we covered the reach of US law in an article on the CLOUD Act and Law 25. This article is for information and does not replace legal advice. For your own situation, consult the person in charge of the protection of personal information in your business, or a lawyer.

In practice, a first pilot on documents with no personal information spares you most of these questions while you learn.

The self-hosted option

Another route is to keep the processing in-house or with a hosting provider in Quebec.

For automation, n8n, for example, installs on your own server and connects your email, your accounting software and your shared folders. Its source code is public, and its licence (the Sustainable Use License) allows use for a business’s internal purposes. For the AI itself, open models can run on a server hosted in Quebec or on a computer in your office: the documents do not leave a machine you control.

This route has a cost. You need someone to install and maintain the server, hardware powerful enough to run the model, and the acceptance that a locally run model will sometimes perform less well than the largest online services. That is exactly what the pilot lets you check: the before-and-after measurement will tell you whether a local model is good enough for your task.

What to do this month

For one week, ask each person on the team to write down the tasks they repeat and how much time those tasks take. Pick one, preferably with no personal information. Measure it for two weeks, name a person in charge and launch a one-month pilot.

If you would like help choosing the process, measuring it and running the pilot, Souverain Technologie works with Quebec SMEs on automation and AI adoption, using tools hosted in Quebec when possible. Write to us.

Sources