Canada, the US, the UK and others
- Lesson 7
- 30 min
- Module 2: Laws and standards in plain words
Where we are
In the last two lessons, we read the EU AI Act and the main framework and standards. But Larkfield is a Canadian company, with customers in Quebec, and the rules at home are different again. In this lesson, we look at Canada and Quebec, with short notes on other places, and we set up a rule watch, like a shop that sells in many regions and keeps its tax rates up to date.
Lesson 7 of 19 · AI Governance: Policies, Risk and AccountabilityCanada, the US, the UK and others
1. What governance is for
2. Laws and standards in plain words
3. Inventory, risk tiers and impact assessment
4. Policies people can follow
5. Controls, documentation and evidence
What you will be able to do
Larkfield must follow the rules at home as well as the EU’s. You learn what Canada and Quebec ask today and what is only proposed. You also learn to watch the rules, so that you know when to ask a lawyer. Then you fill the rules for all eight systems.
Before you startWhat to know and to have ready
What you need to know
- The six kinds of rule, from A map of the landscape, and the EU AI Act’s categories and roles, from The EU AI Act.
What you need
- The map of laws and standards, the rules per system template and the eight system cards.
This course is not legal advice. The sources are official pages: PIPEDA (checked 22 October 2026), Quebec’s private-sector privacy act (checked 22 October 2026), the Directive on Automated Decision-Making (checked 22 October 2026), and the Parliament of Canada pages for bills C-27, C-34 and C-36 (each checked 22 October 2026). For the UK: the government’s AI regulation approach and the Algorithmic Transparency Recording Standard hub (both checked 22 October 2026). The Canadian plans: the government’s news releases on the national AI strategy and the AI transparency consultation (both checked 22 October 2026).
A shop that sells in many regions
Section titled “A shop that sells in many regions”Think of a country with regions, where a shop pays a national sales tax and a regional one. One region has its own rules. A shop that sells across the country charges each buyer the tax of the buyer’s region. The bookkeeper does not memorise every rate. She keeps a list, checks it at set times and when the news reports a change, and asks an accountant when a case is unclear.
Larkfield’s rules work the same way. There are national rules, a region with its own rules and other countries. Someone keeps the list up to date.
The problem: “Canada has no AI law”
Section titled “The problem: “Canada has no AI law””Tomás has read the news. “Canada’s AI bill never passed. So at home, nothing applies to our AI.”
He is half right, and the wrong half is expensive. Canada’s privacy law applies to every one of Larkfield’s systems that uses personal information. Quebec’s privacy act has a rule for decisions made only by a machine. And two federal bills could change the picture. If Larkfield waits for an “AI law”, it misses the laws that already apply.
The idea: what binds today, and what is only proposed
Section titled “The idea: what binds today, and what is only proposed”Canada, like the EU, has laws, bills and government policies. The kinds from the landscape map help most here: for each text, ask whether it binds, and whom.
Text description of the diagram
The diagram has two groups. In the federal group, one text is in force: PIPEDA, the privacy law for companies, with its accountability and consent principles. Bill C-27, with the proposed AI and Data Act, did not pass. Two bills are not law. C-36 is on privacy, with an explanation of automated decisions. C-34 is on online safety, and covers some chatbot services. The Directive on Automated Decision-Making binds only the government. The national strategy “AI for All” and a consultation on AI transparency are plans. In the Quebec group, one text is in force: the private-sector privacy act, section 12.1, on decisions based only on automated processing. Larkfield, a Canadian company with customers in Quebec, has solid arrows to PIPEDA and to section 12.1. A dotted arrow, “watch”, goes to the bills.
Canada’s AI bill did not pass. Bill C-27 included the proposed Artificial Intelligence and Data Act (AIDA). It was still in committee when the session of Parliament ended in January 2025, so it did not become law. The course found no federal law in force made only for AI. A lawyer confirms this on the day you ask.
Privacy law applies today. PIPEDA is Canada’s federal privacy law for private companies. Two of its principles matter most for AI:
- Accountability. An organisation is responsible for personal information under its control, and it names a person who is accountable. It stays responsible when it sends the information to another company for processing, such as an AI vendor (Schedule 1, principle 4.1 and 4.1.3).
- Consent. Collecting, using or disclosing personal information needs the person’s knowledge and consent, except where that is inappropriate (Schedule 1, principle 4.3).
Quebec has a rule for decisions made by a machine. Section 12.1 of Quebec’s private-sector privacy act covers a decision based only on automated processing of personal information. The business that makes it must tell the person. It must do so no later than when it tells them the decision. On request, it explains the information used and the main reasons and factors. The person can submit observations to a staff member who can review the decision. This has been in force since 22 September 2023. Quebec’s privacy regulator can impose a penalty when a business does not inform the person or does not let them submit observations (section 90.1).
Two bills to watch. Bill C-36, a federal privacy bill, was introduced on 15 June 2026. As drafted, it would give people a right to an explanation of a decision made with an automated decision system. Bill C-34, the Safe Social Media Act, was introduced on 10 June 2026. It covers social media and some chatbot services. Neither is law, and both may change.
For the government only. The federal Directive on Automated Decision-Making applies to federal institutions, not to private companies. Before an institution uses an automated decision system, it publishes an Algorithmic Impact Assessment, with an impact level from I to IV (section 6.1.1).
Plans. Canada launched a national AI strategy, “AI for All”, in June 2026. It does not create binding AI rules by itself. A public consultation on AI transparency ran from July to September 2026.
Sector rules. Some sectors, such as banking or health, have their own regulators. This course did not check their rules. If your organisation works in one, ask the lawyer which sector rules touch your AI.
Short notes on other places
Section titled “Short notes on other places”The United Kingdom. The UK government set out its approach in a 2023 white paper and its 2024 response. It gives five principles that the existing regulators apply in their own areas, first without a new law. For government, the Algorithmic Transparency Recording Standard is mandatory for all departments and for some public bodies.
The United States. The NIST framework from the last lesson is voluntary. Some US states have passed AI laws of their own, and their start dates have moved. This course keeps state law general. Ask a lawyer about the states where your customers live.
Other places. Many countries have AI laws, bills or strategies. This course checked only the texts on its map. For any other country, add an entry to the map before you rely on it.
A tiny example: Quebec’s rule and the refund workflow
Section titled “A tiny example: Quebec’s rule and the refund workflow”The support-resolution workflow proposes a refund or a refusal, and a person (an agent or a team lead) approves each action. Farida asks: does Quebec’s section 12.1 apply to a refusal for a Quebec customer?
| Step | Question | Answer |
|---|---|---|
| 1 | Is personal information used? | Yes: the customer’s message, the order, the payment status. |
| 2 | Is it a decision about the person? | Yes: refund or no refund. |
| 3 | Is it based only on automated processing (section 12.1)? | A person approves each action. If the person really reviews it: probably not. If the person approves without reading: unclear. |
| 4 | If yes, what must Larkfield do? | Tell the customer; on request, explain the information and the main reasons; let them submit observations to a staff member who can review. |
| 5 | Not sure | Is an approval in a few seconds a real human review? |
| 6 | Question for a lawyer | If a team lead approves a refusal in seconds, is the decision based only on automated processing under Quebec’s section 12.1? |
Step 3 is the heart of it. The answer depends on a fact and on a meaning. A log can show how long people look before they approve. Only a lawyer can say what the words of the law mean here. The lesson Incidents, complaints and redress looks at a complaint about exactly this.
Keep a rule watch
Section titled “Keep a rule watch”The texts on the map change. The EU AI Act’s dates moved in 2026, and two Canadian bills appeared in June 2026. A rule watch is a small routine that notices such changes and sends the right questions to a lawyer.
Text description of the diagram
The loop starts every three months, and also when the news reports a change. Read each entry of the map at its official source. Then ask: did something change that touches one of our systems? If no, write the new date checked, and wait for the next round. If yes, write the questions: the system, the data, the countries, and what the output changes for a person. Ask the lawyer, with the evidence. Then update the map and the rules per system, and tell the system owners. Then the next round starts.
At Larkfield, Farida owns the map. The steps:
- Every three months, read each entry at its official source.
- When the news reports a change, read that entry the same week.
- Check the inventory: does the change touch one of our systems?
- If it does, write the questions before the meeting with the lawyer.
- Update the map with the new date checked, and tell the system owners.
When to ask a lawyer. Ask before the moment passes, not after:
- before personal data goes to a new company or country;
- before a system decides something about a person with no real human review;
- before AI enters an Annex III area, such as hiring;
- when the rule watch finds a change that touches a system.
Bring facts: the system, the data, the countries, and what the output changes for a person.
When to use it, and what it costs
Section titled “When to use it, and what it costs”Keep the watch from the first system. It costs a few hours every three months, more when a law changes. The lawyer’s time costs more, so arrive with the rows filled in and the questions written. Do not use the watch to decide what is legal for your organisation. It tells you when to ask.
Common mistake
A map that nobody checks
Symptom: the map of laws still shows the EU AI Act’s original high-risk date and no Canadian bill. Nobody knows when it was last checked.
Cause: the map had no owner and no rhythm. It was right on the day it was written.
Fix: name one owner, set the three-month rhythm, and write the date checked on every entry. A map without a date checked is a map you cannot trust.
Module practice: the rules for every system
Section titled “Module practice: the rules for every system”You now use the whole module on Larkfield’s eight systems. Larkfield sells to customers in Canada and in the European Union (constructed). TicketHarbor, MinuteMoss and LedgerFern are made up for this course.
Worked example. The document assistant’s row, from this module’s lessons:
| System | Output used in the EU? | AI Act category and role | Privacy law | Contract | Not sure | Question for a lawyer |
|---|---|---|---|---|---|---|
| Document assistant | Not directly: staff answers | Minimal risk (most likely); provider and deployer | PIPEDA | model provider; the partner shop | where the partner shop’s staff are | Does serving another company’s staff make Larkfield a provider placing a system on the market? |
Your task. Fill the same row for the other seven systems in the rules per system template. Use each system’s card and the map. For each system, write at least one thing you are not sure of, and one question for a lawyer.
Show one possible answer
This is the course’s reading, which a legal specialist reviews. It is not legal advice, and a different answer is fine if it gives its reasons.
| System | Output used in the EU? | AI Act category (most likely) and role | Privacy law | Contract | Question for a lawyer |
|---|---|---|---|---|---|
| Ticket triage service | Yes: EU customers’ tickets | Minimal risk: not an Annex III area; agents send the drafts. Provider and deployer | PIPEDA; the GDPR may matter for EU customers | model provider | Do the transparency duties apply to drafts that an agent sends? |
| Support-resolution workflow | Yes: refunds for EU customers | Minimal risk: refunds are not credit scoring. Provider and deployer | PIPEDA; Quebec’s section 12.1; the GDPR’s rule on decisions based only on automated processing may matter | model provider | If a team lead approves in seconds, is the decision based only on automated processing? |
| Support MCP server | Indirectly | Calls no model; part of the workflow’s controls | PIPEDA | — | none |
| Weekly digest script | Yes: EU customers’ messages | Minimal risk as a use; the data transfer is the problem | PIPEDA accountability: Larkfield stays responsible | none | What must Larkfield do now about data sent to a company with no contract? |
| Reply suggester (TicketHarbor) | Yes | Minimal risk; Larkfield deployer, vendor provider | PIPEDA; the GDPR may matter | vendor terms: training on data by default | Is an opt-out by email enough, or must it be in the contract? |
| Meeting-notes tool (MinuteMoss) | Possibly | Minimal risk; deployer | PIPEDA | vendor terms | What must we tell people outside Larkfield before a call is recorded and summarised? |
| Invoice reader (LedgerFern) | Possibly (EU suppliers) | Minimal risk; deployer | little personal data; bank details | vendor terms, not checked | none, after the vendor check |
The surprise: as the course reads it, none of the eight is high-risk under the EU AI Act. The biggest exposure is the oldest kind. Customer data went to a company with no contract, and a refusal may have had no real human review.
Knowledge check
Section titled “Knowledge check”Check your understanding
Tomás says: 'Canada's AI bill never passed, so nothing applies to our AI at home.' Name two texts that already apply to Larkfield's systems, and one that does not apply yet.
Show the answer
Apply already: PIPEDA, Canada’s federal privacy law for companies, for every system that uses personal information. Larkfield stays responsible when a vendor processes it. And Quebec’s section 12.1 may apply, for decisions about Quebec customers based only on automated processing, as a lawyer would confirm. Not yet: Bill C-36 or Bill C-34, which are bills, not law.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
Name three moments when Farida should ask the lawyer before the moment passes, and what she should bring.
Show the answer
For example: before customer data goes to a new company or country. Before a system decides something about a person with no real human review. Before AI enters an Annex III area such as hiring. And when the rule watch finds a change that touches a system. She brings facts: the system, the data, the countries and what the output changes for a person, with the questions written down.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
Module completion check: choose one Larkfield system. State its EU AI Act category (most likely) and Larkfield's role, with the reason, and one thing you are not sure of.
Show the answer
One possible answer, for the support-resolution workflow: minimal risk (most likely). Refunds for a shop are not one of the Annex III areas, and the system does not profile people for access to essential services. Larkfield is provider and deployer, because it built the workflow and uses it. Not sure: is a team lead’s approval in seconds a real human review? It matters for Quebec’s section 12.1 and for the GDPR’s rule on decisions based only on automated processing. That goes to the lawyer.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Recap
- Canada’s AI bill did not pass, but PIPEDA applies today: accountability, also when a vendor processes the data, and consent. Quebec’s section 12.1 covers decisions based only on automated processing.
- Bills C-36 and C-34 are not law; watch them. The Directive on Automated Decision-Making binds only federal institutions.
- The UK gives principles to its existing regulators; in the US, the NIST framework is voluntary and some states have their own laws.
- A rule watch has one owner, a three-month rhythm and a date checked on every entry. It tells you when to ask a lawyer, and what to bring.
Module 2 in one paragraph. You sorted the rules into six kinds and learned to ask, for each one, whether it binds and whom. You read the EU AI Act: its four-level picture, provider and deployer, and its dates as moved in 2026. You saw the NIST framework as a guide, ISO/IEC 42001 as a routine and ISO/IEC 42005 as a check for one system. Then you looked at Canada and Quebec, and set up a rule watch. As the course reads the rules, none of Larkfield’s systems is high-risk. The urgent gaps are a missing contract and a review that may not be real.
You can now say which rules may apply to each system. But the rules assume that you know every system, and Larkfield’s registry lists only half of them. Next, Module 3 builds the AI inventory.
Next: The AI inventory