Blog

Responsible AI for SMEs: where to start, governance and human oversight

For organizations · By KTMA SOLUTIONS · Published · 12 min read

Your employees may already be using artificial intelligence (AI) tools to write, summarize or analyze, with or without your approval. For a small or medium-sized business, responsible AI for SMEs does not start with a large program: it starts with knowing what is being used, with which data, and who answers for the results. This article offers a concrete starting point: modest use cases, a way to protect sensitive data, the privacy obligations to know about in Quebec and Canada, the role of people in decisions, and a cautious reading of adoption statistics.

Where Canadian businesses stand: what Statistics Canada says

On June 11, 2026, Statistics Canada published Analysis in Brief No. 11-621-M on business use of AI in the second quarter of 2026. Here is what the publication reports:

  • 19.2% of businesses used AI to produce goods or deliver services over the previous 12 months, compared with 12.2% in the second quarter of 2025 and 6.1% in the second quarter of 2024. By this measure, the proportion of businesses that used AI to produce goods or deliver services a little more than tripled between the second quarter of 2024 and that of 2026.
  • 40.0% of businesses consider AI not relevant to them. That proportion is 41.4% among businesses with 1 to 4 employees and 21.3% among those with 100 employees or more.
  • Adoption is 19.9% among businesses with 1 to 4 employees and 27.8% among those with 100 employees or more.
  • The most common uses among businesses that use AI are data analysis (36.6%), text analysis (34.5%) and virtual agents or chatbots (28.2%).
  • The top reported obstacle is cybersecurity or privacy (13.4% of businesses reported that it limited their use of AI), ahead of cost (10.6%). These percentages describe the obstacles reported in the survey; consult the publication's tables for the exact reference population before comparing them with other figures.

Two takeaways for an SME: adoption is growing quickly, but a significant share of businesses believe they do not need it; and privacy tops the list of obstacles, which justifies placing it at the centre of your approach.

Why adoption statistics differ from one source to another

If you have read elsewhere that "most employees use AI" or that "one business in two has adopted it," that is not necessarily contradictory. What matters is what is being measured. In the Statistics Canada publication:

  • the unit is the business, not the employee or the individual;
  • the question concerns the use of AI to produce goods or deliver services over the previous 12 months;
  • the data come from Statistics Canada's Canadian Survey on Business Conditions, conducted from April 1 to May 6, 2026, using a stratified random sample (the exact sample sizes appear in the methodology section of the official June 11, 2026 publication). The results describe the use of AI by businesses, not the proportion of employees who use it.

A survey that counts individuals, that includes personal use of a chatbot, or that covers larger businesses will produce different results. Before quoting or comparing a figure, therefore, check the population (businesses or individuals), the definition of "use," the reference period, the sample size and the date. We did not read other surveys for this article; this advice on how to read statistics is general.

Start small: modest, bounded use cases

A reasonable first step is to choose one or two low-risk uses where an error is easy to spot:

  • rewording or proofreading a text that contains no personal information;
  • summarizing non-sensitive internal documents;
  • sorting or tagging internal requests to help a person handle them;
  • preparing drafts that someone always reviews.

Conversely, avoid starting with uses where AI directly affects people (hiring, credit, access to a service) or handles sensitive data. For each use case, write down: the objective, the data used, who validates the result, and what happens if the tool gets it wrong. This is our recommended method; it is consistent with the principle of necessity and proportionality in the joint statement from Canadian privacy authorities (December 7, 2023), which asks that the use of generative AI and personal information be justified in relation to the intended objective, and that anonymized or synthetic data be favoured where possible.

Take stock of your tools, including "shadow AI"

"Shadow AI" refers to AI tools used by staff without the organization's knowledge. The expression is used in professional vocabulary and has no official definition. We do not cite any numerical estimate of its extent here, for lack of a sufficiently reliable and directly relevant source.

Instead, run a no-blame inventory:

  1. Ask teams which tools they use (chatbots, browser extensions, AI features added to existing software).
  2. Note for each one: vendor, personal or business account, types of data entered, where the data is processed.
  3. Classify: allowed, allowed with conditions, prohibited.
  4. Publish a short, understandable policy with a contact person.

The goal is not to punish but to replace hidden use with managed use. An overly restrictive policy risks discouraging people from reporting existing uses: favour clear rules that are proportionate to the risks and come with approved alternatives.

Sensitive data: the basic rule

The joint statement asks that the collection, use and disclosure of personal information be limited to what is necessary, and that safeguards proportionate to its sensitivity be put in place. The Office of the Privacy Commissioner of Canada recommends that businesses limit the disclosure of personal, sensitive or confidential information, and build in privacy by design.

In practical terms, define what must never be entered into an unapproved tool: client or employee information, health records, financial data, trade secrets, confidential contracts. Also specify what is allowed with an approved tool and under what conditions.

Local or cloud-based AI: a choice based on the data

Both options have their place; the sources consulted do not recommend one over the other. The criteria to compare are common-sense benchmarks, not a requirement drawn from a source:

  • Data sensitivity: the more sensitive the data, the more control you need over where it is processed and stored.
  • Location: in Quebec, a privacy impact assessment is required in particular before personal information is disclosed outside Quebec or before an entity located outside Quebec is entrusted with collecting, using, disclosing or keeping it on the business's behalf (see below).
  • Vendor terms: retention of the data entered, reuse, subcontractors, logging.
  • Ability to run it: a local model requires skills and equipment; a cloud service shifts part of the risk to the vendor without making it disappear.

Privacy impact assessments (PIAs): when are they required?

In Quebec, a PIA (évaluation des facteurs relatifs à la vie privée, EFVP) is mandatory in the situations set out in law (the guide from the Commission d'accès à l'information du Québec, CAI, April 2024, lists several, some of which concern the public sector). For private-sector businesses, two situations are particularly relevant to AI projects:

  • you plan to disclose personal information outside Quebec, or to entrust an entity located outside Quebec with collecting, using, disclosing or retaining personal information on the business's behalf (section 17 of the Act respecting the protection of personal information in the private sector). The disclosure may take place only if the PIA concludes that the personal information would receive adequate protection and if a written agreement is entered into. The business must also check and meet the applicable information and consent obligations, taking into account the exceptions provided by law;
  • you plan to acquire, develop or overhaul an information system or electronic service delivery system involving personal information (section 3.3 of the same Act). Whether the project qualifies must be assessed based on its actual characteristics.

In other words, connecting an external AI tool to client data can trigger the obligation. The guide adds that the scope of the assessment must be proportionate to the sensitivity of the information, its purpose, its quantity, its distribution and its medium, and that the business must consult its person in charge of the protection of personal information. The guide itself is not binding in its form, since the law does not dictate how to carry out the PIA.

If your tool does not touch any personal information, these provisions do not apply to that case, but any doubt should be resolved in favour of caution: have your situation validated by the responsible person in your organization or by legal counsel. This article is not legal advice.

Automated decisions: when a human must step in

The CAI states that when a decision is based exclusively on automated processing of personal information, the business must:

  • inform the person concerned no later than when the decision is communicated;
  • on request, tell them the personal information used, the reasons and the main factors and parameters that led to the decision, and their right to have the information used corrected;
  • give them the opportunity to submit their comments to a staff member who is in a position to review the decision.

These are legal obligations, and they apply to fully automated decisions. A decision that a person genuinely reviews before it is made is not of the same nature; but beware of false oversight: a rubber-stamp approval is not a control. Our good-practice recommendation is to keep a competent human who is authorized to change the outcome for any decision that affects a person, and to record that person's intervention.

The Office of the Privacy Commissioner of Canada notably recommends transparency about the use of generative AI (for example, a notice when it was used to create content or make decisions), as well as understandable and explainable tools. These recommendations do not, on their own, constitute a general obligation applicable to every use of AI; legal obligations depend on the context and the applicable laws.

Minimal governance for an SME

The governance measures proposed here are good practices. However, designating the person in charge of the protection of personal information and publishing their contact information, PIAs in the situations set out in law, the obligations relating to decisions based exclusively on automated processing, and the obligations applicable to confidentiality incidents are governed by Quebec legislation. The specific obligations depend on the situation.

  • A person responsible for AI governance. This is separate from the following function: the person with the highest authority in the business is responsible for the protection of personal information by default; they may delegate that function, in whole or in part, in writing. The title and contact information of the person in charge must be published in accordance with the applicable CAI requirements.
  • A register of tools and use cases.
  • A one-page usage policy.
  • A review before each new use case involving personal information.
  • A procedure for reporting errors and incidents. In Quebec, the business must assess the risk of serious injury to the people concerned following a confidentiality incident; if such a risk exists, it must notify the Commission d'accès à l'information and the people concerned, in accordance with the applicable legal requirements. It must also take reasonable measures to reduce the risks and prevent further incidents, and keep a register of all confidentiality incidents.
  • Short training for staff. In its second-quarter 2026 survey, Statistics Canada also reports that 44.4% of businesses that use AI changed their training or staffing practices; this percentage applies to businesses that use AI only.

Frequently asked questions

Do I need a PIA for every AI tool? No, not automatically. According to the CAI guide, it is mandatory in situations defined by law, of which the two most common for an SME are disclosure outside Quebec and an information system project. Assess your own case.

Is an AI-generated meeting summary an automated decision? The obligations described above apply to decisions based exclusively on automated processing of personal information. A summary is not, in itself, such a decision. But it may contain personal information: privacy rules apply.

Do I have to tell clients that I use AI? For a decision based exclusively on automated processing of personal information in Quebec, the business must inform the person concerned no later than when the decision is communicated. Other information or consent obligations may apply depending on the context. The Office of the Privacy Commissioner also recommends transparency, in particular a notice when content or a decision comes from generative AI.

My business is small: does this apply to me? The Act respecting the protection of personal information in the private sector applies to personal information that a person collects, holds, uses or communicates to third parties in the course of carrying on an enterprise, subject to its scope and the exceptions provided by law. Statistics Canada also shows that 19.9% of businesses with 1 to 4 employees used AI.

Where to start?

Take stock of the tools in use, choose a first low-risk use case, and define the data-handling rules and the human validation step. Before using personal information, check the applicable obligations and consult your person in charge of the protection of personal information.

Going further with KTMA SOLUTIONS

To structure your approach, explore our artificial intelligence and IT consulting services, or tell us about your project.

KTMA Agent offers an approach to AI assistance that favours local operation and subjects sensitive decisions to explicit rules and human oversight. It does not replace the user's judgment, is not a universal or autonomous agent, and offers no guarantee of results.

Sources

Back to the blog · Blog RSS feed