None of these three words has a legal definition in Canada. What separates them in practice is a real, checkable difference in behaviour — whether the system only talks, or whether it can also act — not the marketing on the product page.
Key takeaways
None of “chatbot”, “AI agent” or “assistant” is defined in Canadian law. The nearest thing to an official Canadian description of the underlying technology is the Cyber Centre’s definition of generative AI itself: “a type of AI that generates new content by modelling features from large datasets that were fed into the model… generative AI can create unique content in many forms, including text, image, audio or software code.” (Canadian Centre for Cyber Security, ITSAP.00.041) That definition covers all three product categories equally — it says nothing about the difference between them, because the difference is a product-design choice, not a distinct technology.
The clearest, checkable distinction is whether the system can act, not just talk. Anthropic’s own developer documentation describes what it calls tool use this way: “tool use (also called function calling) lets Claude call functions that you define or that Anthropic provides… it then returns a structured call that your application executes” — or, for a small set of Anthropic-run tools, that Anthropic executes on its own side. (Anthropic, Tool use with Claude) That capability — calling a function, running a search, updating a record, booking something — is what most people mean when they use the word “agent”: a system that can take a multi-step action toward a goal, not just produce a reply. A plain chatbot, by contrast, generates conversational text and stops there; it does not reach out and change anything in another system on its own.
“Assistant” is the vaguest of the three precisely because it is applied to both ends of the range above. Some products marketed as an “AI assistant” are a chat interface with memory of the conversation and nothing more; others are full tool-using agents that book meetings, search records and update systems, wearing the same label. There is no vendor or regulator convention that reserves “assistant” for either behaviour — a buyer has to ask directly whether the product can call a function or take an action, using something close to Anthropic’s own description above, rather than relying on which of the three words appears on the marketing page.
Whatever a business calls its own tool, Canada’s Voluntary Code of Conduct on generative AI puts a specific obligation on whoever manages a customer-facing system, regardless of the label used for it: “ensure that systems that could be mistaken for humans are clearly and prominently identified as AI systems.” (ISED, Voluntary Code of Conduct) The Code is voluntary and binds only its 46 signatories, so it does not bind every business — but a chatbot, an agent and an “assistant” are all equally capable of being “mistaken for human” in a text conversation, and the norm draws no distinction between the three labels. The disclosure question does not depend on which word is used on the product page.
A retailer deploys a customer-facing tool that answers questions about store hours and return policy by generating a reply from a knowledge base — a chatbot, on the definition above, because it only talks. Six months later the same vendor adds a feature that lets the tool actually look up a specific order and issue a refund without a human clicking anything — at that point it has crossed into agent territory, because it is now calling a function that changes a real system, not just generating a reply about one. If the vendor calls both versions “the AI assistant” in its marketing, the label has not changed even though the underlying capability, and the risk profile the retailer needs to manage, changed substantially.
The moment a system moves from generating a reply to actually issuing a refund, it has crossed from a conversation into an “administrative… or otherwise significant decision” about a real customer account — the exact category Canada’s privacy commissioners say needs “an effective challenge mechanism… allowing them the opportunity to request human review and/or re-consideration of the decision.” (OPC, Principles for responsible, trustworthy and privacy-protective generative AI) A plain chatbot answering a question about store hours never reaches that bar; an agent that can independently issue a refund reaches it the moment the capability is switched on, regardless of whether the vendor still calls it “the assistant.” The practical lesson for a buyer: ask what changes on the customer’s account when the tool acts, not what the product is called, because that is what determines whether a human review path needs to exist behind it.
If the agent’s action includes sending a message, a second statute attaches regardless of what the product is called. CASL prohibits sending a commercial electronic message unless “the person to whom the message is sent has consented to receiving it” and the message sets out an identified sender and “an unsubscribe mechanism”. (CASL, s.6) It makes no difference whether a marketing employee clicked send or an agent generated and dispatched the message on its own initiative: the maximum penalty for a violation is “$1,000,000 in the case of an individual, and $10,000,000 in the case of any other person”, and that liability lands on the business either way. (CASL, s.20(4)) Treadstone Law’s own walkthrough of how a CASL complaint actually plays out for a business is worth reading before that risk is theoretical rather than after. (Treadstone Law, on CASL enforcement generally)
Before adopting a chatbot, agent or assistant, three questions cut through the marketing more reliably than the product name does. Can it change anything in another system on its own, or does it only generate text a person then acts on? If it can act, what specifically can it do — issue a refund, send an email, update a record — and is that list documented anywhere a buyer can actually read it? And who reviews an action it takes before or after the fact, if anyone? A vendor that answers all three clearly has told a buyer more than any of the three product labels ever will on its own.
No. There is no Canadian regulator or standards body definition. It is used across the industry, including by Anthropic’s own documentation, to describe a system that can call functions or tools to take action, rather than only generate conversational text.
Ask directly whether it can call a function, take an action, or only generate a reply — the label does not reliably tell you, since “assistant” is used for both plain chatbots and fully tool-using agents.
No. The Voluntary Code’s wording covers any system that could be mistaken for a human, regardless of whether it is marketed as a chatbot, an agent, or an assistant.
Related: AI vs automation: what’s the difference, what a workplace AI policy should cover, and connecting a chatbot or agent to systems you already run.
A short call is enough to find out what a product actually does before the label on the pricing page decides it for you.
