AI in your SME: which questions should you be asking?

About fifty statements to take stock of how you use AI, spot what deserves your attention and know where to start. Each statement explains why it matters and suggests a first step.

15 to 20 minutes · 5 context questions · 7 themes

This self-assessment is an awareness and prioritization tool. It helps you identify the practices, risks and obligations to examine in your use of AI. It is not exhaustive and is neither a legal audit, legal advice nor a certification of compliance.

Rosecape does not guarantee that this self-assessment, its results or the implementation of its recommendations ensure compliance with Law 25 or any other law. For an analysis of your situation, consult a legal advisor. The Law 25 content is based on the obligations published by the Commission d'accès à l'information and has not yet been reviewed by a privacy professional.

Your answers are saved only in your browser. They are sent to Rosecape only if you choose to send us your results at the end.

How to answer

Pick the answer that best describes what actually happens today, not what is planned.

No
We do not do this today.
Informal
We do it sometimes, without a clear process.
In place
We have a known practice that is generally followed.
Formalized
The practice is documented, has an identified owner and is reviewed periodically.
I don't know
We do not know whether this practice is in place.
Not applicable
This question does not apply to our activities.

Before you start

Your context

These five questions are not scored. They hide the statements that do not apply to you and adjust how much weight each recommendation gets.

Do you use personal information in your AI tools or systems?

For example names, email addresses, or client or employee records pasted into an AI assistant or processed by a tool.

Do you use sensitive personal information or highly confidential information with AI?

Health, finances, insurance, identity, human resources, sensitive client files.

Do your AI systems make or recommend decisions that can significantly affect a person?

Hiring, insurance, eligibility, pricing, credit, processing priority, access to a service.

Can any AI systems or agents act directly in your business tools?

Updating a CRM, sending an email, creating a ticket, changing data, starting a process, completing a transaction.

Can your data be processed, backed up or transferred outside Québec?

Think about primary hosting, but also your vendors' backups, logs and technical support.

Theme 1 / 7

AI use and governance

Do you know which tools are used, by whom, and under what rules?

We keep a register of the AI systems and tools used in the organization.
  • AI governance
  • Good practice
Why does this matter?
Why
Employees may use ChatGPT, Claude, Copilot or other tools without the organization knowing. Without a register, it is hard to know which data is going to which vendors.
First step
Ask each team which AI tools they use and why, then record the answers in a shared spreadsheet.
Examples
For each tool, record the vendor, internal owner, team, purpose, data categories used, integrations, risk level, known hosting region, planned human oversight and next review date.
Each significant AI system or use case has an identified owner.
  • AI governance
  • Good practice
Why does this matter?
Why
When nobody owns a tool, nobody checks its access, changes to its terms, or the data being sent to it.
First step
Write down the name of one owner for each tool in the register.
We classify our AI use cases by risk level.
  • AI governance
  • Good practice
Why does this matter?
Why
Drafting a generic post and assessing an insurance claim do not call for the same controls. Classifying your uses lets you focus effort where the risk actually is.
First step
Give each use case in the register a level: low, moderate, high or critical.
Examples
A simple grid suggested by Rosecape (not an official standard): low, generic drafting with no confidential data; moderate, internal analysis with business information; high, personal information, client content or significant recommendations; critical, significant decisions, highly sensitive data, autonomous actions or broad system access.
Our employees know which AI tools are approved for work.
  • AI governance
  • Security
Why does this matter?
Why
Without a clear list, everyone picks their own tools, often free versions whose terms allow the vendor to reuse data.
First step
Publish a short list of approved tools and permitted uses somewhere everyone knows to look.
We have set which categories of data can be used in which AI tools.
  • AI governance
  • Confidentiality
Why does this matter?
Why
A tool may be fine for public content and inappropriate for client files. Simple rules spare everyone from deciding case by case.
First step
Cross three or four data categories (public, internal, confidential, personal information) with your approved tools.
Examples
Public content: any approved tool. Internal documents: business-plan tools only. Personal or sensitive information: only tools whose terms and processing region have been checked.
Our AI use framework is reviewed periodically and whenever our uses or tools change significantly.
  • AI governance
  • Good practice
Why does this matter?
Why
AI tools and their terms change quickly. A framework written a year ago may no longer reflect how you actually use AI.
First step
Set a regular review (every six months, for example) plus a review whenever a significant new tool or use is adopted.

Theme 2 / 7

Personal information and Law 25

Protecting personal information: have you put the main Law 25 habits in place?

A person in charge of the protection of personal information is clearly identified in our organization.
  • Law 25
Why does this matter?
Why
Required since September 2022 for every business that holds personal information. By default, the law assigns this role to the person with the highest authority (often the president): without a written designation, that person holds the responsibility, possibly without knowing it. The role can be delegated in writing, in whole or in part.
First step
Confirm who holds this role today and put the designation in writing, even if the president keeps the role.
The title and contact information of our person in charge of personal information are easy to find on our website.
  • Law 25
Why does this matter?
Why
The law requires this information to be published on the business's website or, failing that, made available by other means, so that people know whom to contact.
First step
Add the job title and a dedicated contact address (for example privacy@yourcompany.ca) to the footer or the privacy policy.
We have established, and we apply, policies and practices that govern how we handle personal information.
  • Law 25
Why does this matter?
Why
Since September 2023, every business must establish and apply policies and practices covering the life cycle of personal information, from collection to destruction, the roles and responsibilities of its staff and the handling of complaints.
First step
Write a short policy covering the life cycle of your data, your staff's roles and how complaints are handled.
Examples
A policy can fit in a few pages: which information, why, for how long, who can access it, how to handle a complaint.
Our website presents detailed information about our personal information governance policies and practices, in clear and simple terms.
  • Law 25
  • Transparency
Why does this matter?
Why
Since September 2023, this information must be published on the business's website. A business can have its policies and still not have published the required information.
First step
Add a page to your website that summarizes your policies and practices in plain language.
We know which categories of personal information we hold, why we use them, where they are and who can access them.
  • Law 25
  • Good practice
Why does this matter?
Why
You cannot protect what you do not know you have. This map is also the starting point for deciding what may be used with an AI tool.
First step
List the information you hold on clients, employees, prospects, suppliers and partners, with where it lives and who can access it.
Before collecting or using personal information in an AI system, we check that it is actually necessary for the purpose.
  • Law 25
Why does this matter?
Why
The law limits collection to the information needed for the stated purpose. With AI, it is tempting to send a whole file when a few fields would do.
First step
For each AI use involving personal information, write down the purpose and the fields actually needed; remove or mask the rest.
When we collect personal information, we clearly explain why it is collected and how it will be used.
  • Law 25
  • Transparency
Why does this matter?
Why
At the time of collection, each person must be told the purposes, their rights of access, correction and withdrawal of consent and, where applicable, the third parties their information may be shared with or the possibility that it may be communicated outside Québec.
First step
Reread your forms and privacy policy: do they mention the AI tools or vendors that process this information?
When consent is required, we have a clear way to obtain it and to record what the person agreed to.
  • Law 25
Why does this matter?
Why
Consent must be manifest, free, informed and given for specific purposes. It depends on context and does not replace the necessity requirement: unnecessary information does not become acceptable because someone consented.
First step
Find the uses where you rely on consent (for example a secondary use of client data) and check how it is requested and recorded.
When consent is required for sensitive information (health, financial or biometric information, for example), we obtain it expressly, through a separate request.
  • Law 25
Why does this matter?
Why
Sensitive information requires express consent, requested separately. With AI, also think of sensitive files someone might want to analyze for purposes other than the ones originally intended.
First step
Identify where you handle sensitive information and add a separate consent step.
If a decision about a person is made exclusively through automated processing, we know which transparency and review obligations apply.
  • Law 25
Why does this matter?
Why
When a decision relies solely on automated processing, the person must be informed. On request, they can learn the information and main factors used, have their information corrected, and submit observations to someone able to review the decision.
First step
Find the decisions where no person steps in before the final outcome, and check how you inform the person concerned.
Examples
In insurance: a tool that approves or rejects a claim with no human review.
We know whether our AI systems are used to profile, rank or assess people, and we have identified the obligations that may apply.
  • Law 25
Why does this matter?
Why
Analyzing a person's behaviour, preferences, financial situation, performance or eligibility is profiling. Among other things, the law requires informing people when technologies that can profile them are used.
First step
For each AI use, ask: is it used to assess or rank people?
Examples
In HR: a tool that ranks applicants or summarizes employee files for a performance review.
When personal information is used to make a decision about a person, we have a way to check that it is accurate and up to date.
  • Law 25
Why does this matter?
Why
A decision based on outdated information, or information the AI extracted incorrectly, can cause real harm. The law requires information used to make a decision to be accurate and up to date.
First step
For significant decisions, identify the source of each piece of information used and who confirms it is accurate.
We know how long to keep personal information and how to destroy it when it is no longer needed.
  • Law 25
Why does this matter?
Why
Once the purpose has been achieved, the information must be securely destroyed. Anonymization is possible in some cases, but it is subject to specific criteria. Keeping information indefinitely increases your exposure in every incident. Remember the copies kept in your AI tools' history, too.
First step
Set a retention period for your main categories of information and check the history-deletion options in your AI tools.
We have a process for responding to access and correction requests and other rights that apply to personal information.
  • Law 25
Why does this matter?
Why
A person can ask to see or correct the information you hold about them, and the law sets response deadlines. If that information also sits in AI tools, you need to know how to find it.
First step
Name who receives these requests, note the applicable response deadline and the list of systems to check.
We could give a person who asks for it their computerized personal information in a structured, commonly used format.
  • Law 25
Why does this matter?
Why
Since September 2024, a person can require their computerized information in a structured, commonly used format. Data scattered across several systems makes this tedious.
First step
Check whether you can export a person's information from each of your systems.
When we build products or services that have privacy settings, we build in privacy from the design stage and by default.
  • Law 25
Why does this matter?
Why
If you offer the public a technology product or service (app, client portal, platform), its settings must provide the highest level of confidentiality by default, without any action by the user. Features that identify, locate or profile a person stay turned off until the person turns them on.
First step
Check the default settings of your apps or portals: sharing, history, profile visibility.

Theme 3 / 7

Data, vendors and sovereignty

Do you know where your data goes and what your vendors can do with it?

We know when a privacy impact assessment (PIA) is required and we have a way to decide whether an AI project needs one.
  • Law 25
Why does this matter?
Why
Since September 2023, a privacy impact assessment must precede any project to acquire, develop or overhaul a system that involves personal information. It is also required before communicating personal information outside Québec. Not every AI project needs one.
First step
Add one question to your project approval: does this project involve personal information or send it outside Québec?
Before communicating personal information outside Québec, including by entrusting it to a cloud tool hosted elsewhere, we assess the risks.
  • Law 25
  • Sovereignty
Why does this matter?
Why
Communicating personal information outside Québec requires a prior assessment. A US cloud tool (email, CRM, AI) that hosts your client or employee data is a communication outside Québec.
First step
Take stock of your tools: where are your client and employee data actually hosted?
For our main AI tools, we know the regions where our data may be processed, backed up or transferred.
  • Law 25
  • Sovereignty
Why does this matter?
Why
Data can be hosted in Canada and still be backed up, logged or accessed by support staff elsewhere. Knowing these locations is necessary to assess communications outside Québec.
First step
For your three main AI tools, check your plan's terms, the processing regions and the list of subprocessors.
Examples
Check: primary hosting, backups, logs, telemetry, support access and subprocessors.
We know which laws and jurisdictions may apply to the vendors that process or host our data.
  • Sovereignty
  • Good practice
Why does this matter?
Why
A vendor may be subject to the laws of its home country even when your data is hosted elsewhere. The US CLOUD Act is a frequently cited example.
First step
For each significant vendor, note the country of its head office and its hosting regions.
For our sensitive or strategic information, we have determined whether Canadian hosting is preferable or required in our context.
  • Sovereignty
  • Good practice
Why does this matter?
Why
Law 25 does not require data to be hosted in Canada, but it does govern communicating it outside Québec. For some sensitive or strategic data, Canadian hosting can simplify the assessment or meet your clients' expectations.
First step
Identify the data categories whose location matters to you or your clients, then decide whether Canadian hosting is desirable.
We know whether our data, prompts and files can be reused by the vendor, including to train or improve its models.
  • Confidentiality
  • Good practice
Why does this matter?
Why
Depending on the plan, a vendor may keep your prompts and files or use them to improve its models. Free versions and business plans often have very different terms.
First step
Check each tool's terms and settings to see whether your data can be used for training, and turn that option off where possible.
Examples
To answer “Formalized”: the business-plan terms have been reviewed, the settings are configured and the contractual commitments have been checked.
We assess the security and privacy posture of our AI vendors before approving them.
  • Security
  • Good practice
Why does this matter?
Why
Approving an AI tool means handing data to a third party. A short check avoids discovering unacceptable terms after the fact.
First step
Prepare a one-page checklist and use it before approving each new tool.
Examples
Cover: data use, subprocessors, retention and deletion, processing location, breach notification, encryption, authentication and access control, business versus consumer terms.
When we entrust personal information to a vendor, our agreements govern how it is used, protected and deleted.
  • Law 25
  • IP / contracts
Why does this matter?
Why
When a vendor processes personal information on your behalf, the law calls for a written agreement that covers its protection, limits its use to the mandate and provides for its destruction when the contract ends.
First step
For each vendor that receives personal information, check that the contract or data processing agreement covers these points.

Theme 4 / 7

Security and access

Who can access what, and what can your AI tools do in your systems?

Accounts used to access business AI tools are protected by multi-factor authentication where available.
  • Security
Why does this matter?
Why
A compromised AI tool account exposes the conversation history and uploaded files, which often contain confidential information.
First step
Turn on multi-factor authentication for every business AI tool account, or use your company's single sign-on.
Work use of AI happens through accounts and environments approved by the organization.
  • Security
  • AI governance
Why does this matter?
Why
A personal account is outside your control: you cannot manage its settings, recover its content or close it when the employee leaves.
First step
Provide business accounts for approved tools and ask people to stop using personal accounts for work.
Access to AI tools is removed or adjusted promptly when someone changes roles or leaves the organization.
  • Security
Why does this matter?
Why
A former employee who keeps access to an AI tool also keeps access to its history and the documents uploaded to it.
First step
Add AI tools to your checklist for departures and role changes.
Users and AI systems only have access to the data and tools needed for their role or task.
  • Security
  • Good practice
Why does this matter?
Why
An AI assistant connected to your documents can show a user information they should not see. The tool's access should follow the person's and the task's.
First step
For AI tools connected to your systems, check what data they can reach and remove what is not needed.
AI agents that can act in our systems only have the permissions their task requires.
  • Security
  • AI governance
Why does this matter?
Why
An agent with access to your email, CRM, files and databases carries a very different level of risk than a chat assistant: a mistake or a malicious instruction can turn directly into an action.
First step
List what each agent can do, remove what is not needed and require human approval for sensitive actions.
Examples
Actions to consider: editing a client record, sending an email, creating a ticket, changing data, starting a process, completing a transaction.
For high-risk AI uses, we can trace the main actions taken, the data accessed and the systems changed.
  • Security
  • Good practice
Why does this matter?
Why
After a mistake or incident, you need to be able to reconstruct what happened: who asked for what, which data was accessed and what was changed.
First step
Check which logs your high-risk AI tools keep, for how long and who can read them.

Theme 5 / 7

Use-case risk and human oversight

Do you know when a human needs to look, and how to check what the AI produces?

We have defined which AI-generated content, recommendations or actions need human review before they are sent or carried out.
  • AI governance
  • Good practice
Why does this matter?
Why
Reviewing everything is unrealistic; reviewing nothing is risky. What matters is deciding in advance what needs a human eye.
First step
For each use case, state whether the output can go out as is, must be read over, or must be approved.
When AI output can have significant consequences, we have defined how its accuracy must be checked.
  • AI governance
Why does this matter?
Why
AI can produce answers that sound right but are wrong: invented citations, faulty calculations, incorrectly extracted information.
First step
For output with consequences, specify what must be checked and how: the source, a recalculation, a second read.
Examples
Check: citations, calculations, information extracted from documents, client recommendations, legal or financial statements.
For important information, our AI systems rely on identified, reliable data sources.
  • AI governance
  • Good practice
Why does this matter?
Why
An AI that draws on outdated or conflicting documents gives inconsistent answers. Important information should come from known, current sources.
First step
For each significant AI use, identify the reference source of the information: a system, a document or a responsible person.
When AI contributes to a significant decision about a person, we have determined when human review or a means of recourse is needed.
  • AI governance
Why does this matter?
Why
Even when a person signs off on the decision, an AI recommendation can carry a lot of weight. The people affected should be able to ask for it to be reconsidered.
First step
For each significant AI-assisted decision, specify who can review it and how a person can challenge it.
Examples
Hiring, insurance, eligibility, pricing, credit, processing priority, access to a service.
We have defined when the use of AI must be clearly disclosed to clients or users.
  • AI governance
  • Transparency
Why does this matter?
Why
Your clients expect to know when they are dealing with an automated system or when a deliverable was largely produced by AI. A clear rule avoids awkward situations.
First step
Decide when you tell clients that AI is being used and write a standard notice.
Examples
Situations to consider: a deliverable largely produced by AI, an automated exchange that could be mistaken for a person, a significant AI-assisted decision, client data processed by an AI tool.

Theme 6 / 7

Confidentiality, intellectual property and contracts

Three separate questions: are you allowed to give this content to the AI, can the vendor reuse it, and can you use the result?

Our employees know not to give an AI tool confidential or protected content they are not authorized to share.
  • IP / contracts
  • Confidentiality
Why does this matter?
Why
A confidentiality agreement with a client or supplier may prohibit passing its documents to a third party, including an AI tool.
First step
Remind teams to check confidentiality commitments before uploading an outside document to an AI tool.
Examples
Construction plans or bids, production documents containing a supplier's intellectual property, client documents in professional services.
Before approving an AI tool, we check its terms on our data, confidentiality, intellectual property and reuse of content.
  • IP / contracts
  • Good practice
Why does this matter?
Why
The terms of use determine what the vendor may do with your data and who holds the rights to the content produced.
First step
Add a review of the terms (data, confidentiality, intellectual property, reuse) to your tool approval process.
We have determined under what conditions AI-generated content may be used in our deliverables, products or communications.
  • IP / contracts
  • Good practice
Why does this matter?
Why
Generated content can reproduce protected material, and your client contracts may include guarantees about the originality of deliverables.
First step
Specify which deliverables may include generated content and with what review, then check what your client contracts say.

Theme 7 / 7

People, training and incidents

Do your teams know what to do day to day, and what to do when something goes wrong?

Our employees know which data they may and may not send to AI tools.
  • AI governance
  • Training
Why does this matter?
Why
Rules only work if people know them. Many AI-related leaks come from a copy-and-paste done in good faith.
First step
Walk through the data and tool rules in 15 minutes at a team meeting, using examples from your day-to-day work.
Our employees know the main AI risks, including confidentiality, hallucinations, bias, fraud and intellectual property.
  • AI governance
  • Training
Why does this matter?
Why
Invented answers, bias, more convincing phishing, cloned-voice fraud: employees who know these risks spot them sooner.
First step
Plan a short yearly training on the main risks, with concrete examples.
People who use high-risk AI systems receive training suited to those uses.
  • Good practice
  • Training
Why does this matter?
Why
People who use AI for significant decisions or with sensitive data need more specific guidance than others.
First step
Identify the users of high-risk use cases and give them guidance specific to those uses.
Employees know whom to ask when they are unsure about using an AI tool or a piece of data.
  • AI governance
  • Good practice
Why does this matter?
Why
Someone who does not know whom to ask often ends up deciding alone. A clear point of contact prevents many mistakes.
First step
Name a person or a chat channel for AI questions and let everyone know.
Our incident response process also covers incidents involving AI tools or information sent to them.
  • Law 25
  • Security
Why does this matter?
Why
Sensitive data uploaded to the wrong tool, a compromised AI account or a vendor breach are all incidents. When an incident happens, you need to act fast: contain it, assess the risk of harm, notify if needed.
First step
If you do not have a process yet, write a one-page one: who does what, in what order. Add AI-related scenarios to it.
Examples
Sensitive data sent to the wrong tool, a compromised AI account, a vendor breach, an accidental disclosure, an AI answer exposing another client's information.
We keep a register of our confidentiality incidents.
  • Law 25
  • Security
Why does this matter?
Why
The confidentiality incident register is mandatory, and the Commission d'accès à l'information (CAI) can ask for a copy at any time. An incident can be a leak, unauthorized access or a loss of information, including through an AI tool.
First step
Create the register now, even if it is empty: date, nature of the incident, information involved, measures taken.
We know when we must notify the Commission d'accès à l'information and the people concerned of a confidentiality incident.
  • Law 25
  • Security
Why does this matter?
Why
Any incident that presents a risk of serious harm must be reported to the CAI and to the people concerned. You need to be able to assess that risk quickly, when the incident happens.
First step
Build the risk-of-serious-harm test into your incident process and name who decides to notify.

Answer every question to see your results. “I don’t know” is a valid answer.