AI Chatbot and AI Agent Development in Canada

About AI Chatbot and AI Agent Development

A chatbot that answers from your own help docs, catalogue and policies instead of making things up. An agent that can check an order, book a slot or update a CRM record, and nothing you didn't allow. We've shipped 18 production AI integrations, and every one has a running-cost cap.

What the engagement includes

• Answers come from your content only, with the source shown, so it can't invent a refund policy you don't have
• A confidence threshold and a clean hand-off to a human when the question is outside what it knows
• An evaluation set of real customer questions, run before every change, so a prompt tweak can't quietly break answers
• Hard monthly spend limits and per-request cost tracking, so the model bill can't surprise you
• Agents get the narrowest permissions that work, and anything touching payments needs a human to approve
• Your data stays within the provider settings you choose, and nothing is used for model training

Air Canada's chatbot ruling, the federal generative AI code and Ontario's rule on errors with electronic agents

The Canadian case every customer-facing chatbot is measured against is Moffatt v. Air Canada, decided by British Columbia's Civil Resolution Tribunal on 14 February 2024. A customer whose grandmother had just died asked the airline's website chatbot about bereavement fares and was told they could book at the regular fare and apply for the reduced rate within 90 days of the ticket being issued. The airline's own bereavement page, linked from that very answer, said the policy did not apply to requests made after travel. Air Canada argued it could not be held liable for information provided by one of its agents, servants or representatives, a chatbot included. The tribunal called that a remarkable submission, held that a chatbot is still just a part of the company's website and that the company is responsible for all the information on it whether it comes from a static page or a chatbot, found the airline had not taken reasonable care to ensure its chatbot was accurate, and asked why customers should have to double-check information in one part of a website against another. It awarded $650.88 in damages plus interest for negligent misrepresentation. It is a small claims decision rather than an appellate ruling, but it is the reasoning a Canadian customer will cite. Three design choices follow. When the bot answers a question about policy, price, eligibility or refunds, it quotes the current approved source and links it rather than paraphrasing a rule into a promise, and where the source does not cover the question it says so and hands over to a person. Every answer is logged with the model version, the prompt configuration and the passages it drew on, so a screenshot produced months later can be checked against what the bot actually said and why. And the policy pages and the knowledge base the bot reads are published from one place, so the two cannot drift apart. How far your own terms limit that exposure is for your counsel.

Canada has no statute telling a chatbot to say it is a chatbot, but two federal documents ask for it. The Voluntary Code of Conduct on the Responsible Development and Management of Advanced Generative AI Systems, published by Innovation, Science and Economic Development Canada in September 2023, separates developers of a system from managers, the organisations that put it into operation, control its parameters and access and monitor it, which is what a company deploying a support bot is. For managers of systems available for public use it asks that systems that could be mistaken for humans be clearly and prominently identified as AI systems, that operation be monitored for harmful uses after release, including through third-party feedback channels, and that cyber security risks, data poisoning among them, be assessed. Its signatories include a bank, payment networks, telecom carriers and large software vendors, so a bot built for one of them can expect to be asked how it meets those measures. The federal, provincial and territorial privacy commissioners' principles for generative AI, issued on 7 December 2023, reach into the conversation itself. A public-facing tool should make sure people know they are interacting with generative AI. Using conversational bots to deliberately nudge people into divulging personal information they would not otherwise have shared, sensitive information in particular, is listed among the practices of concern. Prompts should not be retained, used for secondary purposes such as further training, or disclosed unless that is authorised, and there should be a way for people to access and correct information collected about them while they used the system. In the build that becomes a label in the chat header and the first message rather than a line in the terms, prompt design reviewed for questions that fish for more than the task needs, a transcript store with its own retention period, model provider terms that rule out training on your customers' conversations, and a route from the chat to an access or correction request. Whether either document binds your organisation is for your counsel.

Once the bot can do something rather than only answer, Ontario's Electronic Commerce Act, 2000 decides when what it did binds the customer. The Act defines an electronic agent as a computer program used to initiate an act or respond to electronic documents or acts, in whole or in part, without review by an individual at the time, which describes a chatbot that books an appointment, changes a plan or places an order. It provides that an offer or acceptance may be expressed by touching or clicking an icon or by speaking, and that a contract may be formed by the interaction of an electronic agent and an individual, so a typed yes in a chat window or a spoken one to a voice bot can conclude a contract. Section 21 is the part a chatbot has to be built around. An electronic transaction between an individual and another person's electronic agent is not enforceable by that other person if the individual makes a material error, the agent gave them no opportunity to prevent or correct it, they promptly notify the business on becoming aware of it, and, where they received something because of the error, they return or deal with it as instructed and do not materially benefit. A language model makes that failure easy, because it fills in a quantity, a date or a plan tier from a misread message and completes the action in the same turn. So before any action with money or a commitment attached, the bot restates the specific terms, item, quantity, date, price and total, in a structured confirmation the customer accepts as a separate step, with an obvious way to change any field, and the transcript keeps that confirmation and the acceptance together. A voice bot reads the same summary back and waits for an explicit answer rather than treating a pause or a filler word as agreement. And undoing a mistaken booking is a tool the bot can call, not a promise to pass a message on. Whether a given transaction is enforceable is for your own counsel.

How we work with Canadian clients

Two things shape Canadian projects more than anything technical. The first is language: if you serve Quebec, French is a legal requirement rather than a nice addition, and the rules tightened considerably under Bill 96. The French version has to be available on equal terms with the English one, which is a design constraint from the first wireframe rather than a translation task at the end.

The second is anti-spam law. CASL is stricter than most teams expect, and the penalties are real. Consent has to be express or clearly implied, records of it have to be kept, and every commercial message needs a working unsubscribe and accurate sender identification. We build those into the product rather than leaving them to whoever configures the mailing tool later, because that is where the exposure usually sits.

Questions Canadian clients ask

AI Chatbot and AI Agent Development: common questions

Other markets we work in