An AI agent in customer service: what works, what breaks

On February 14, 2024, British Columbia's Civil Resolution Tribunal ordered Air Canada to compensate a customer because the chatbot on its website had invented a refund policy. The amount was modest: $812.02. The lesson was not. For a small business considering an AI agent for customer service, that ruling is the most instructive document published to date: it draws, with the precision of a legal decision, the line between what you can hand to an agent and what you never should.

Many business owners still file these tools with the frustrating chatbots of the mid-2010s, those decision trees that answered beside the point and sent customers in circles. The technology has changed in kind since then. The ruling, however, remains fully current, because it is about responsibility, whatever the engine underneath. This article turns it into concrete design rules: what works in production, and what breaks.

The short answer, for the busy

Yes, an AI agent can handle part of a small business's customer service, provided it is built with explicit limits. It performs well on questions whose answer already exists in your documentation (policies, delays, pricing), on order status lookups, and on triaging requests before handing them to a human. It breaks on policy exceptions, angry customers, and anything that resembles a promise. The risk is real: in February 2024, British Columbia's Civil Resolution Tribunal ordered Air Canada to pay $812.02 because the chatbot on its website had invented a bereavement fare policy (Moffatt v. Air Canada, 2024 BCCRT 149). A company is responsible for all the information published on its website, conversational agent included. The design rules that follow: answers grounded in your data, automatic escalation to a human, full conversation logging, and a hard ban on inventing policy.

What the tribunal actually said

In November 2022, Jake Moffatt booked two flights between Vancouver and Toronto after his grandmother's death. Before booking, he asked Air Canada's chatbot how bereavement fares worked. The bot's answer: book now, then submit your refund request within 90 days of the ticket being issued. That was wrong. The official page on the same website stated the policy did not apply after travel. Mr. Moffatt paid $1,630.36 for his two tickets, filed his claim within the window the bot had given him, and was refused.

Before the tribunal, Air Canada argued it could not be held liable for information provided by its chatbot, presenting it in effect as a separate legal entity responsible for its own actions. The tribunal called this a "remarkable submission", which, in a legal decision, is not a compliment. Its answer fits in one sentence: a company is responsible for all the information published on its website, and it makes no difference whether that information comes from a static page or a chatbot.

The tribunal also rejected the fallback argument, that the customer could have checked the correct page (the bot even linked to it). Air Canada did not explain why that page would be inherently more trustworthy than its own chatbot, or why a customer should have to cross-check one section of a website against another. The verdict: $650.88 in damages for negligent misrepresentation, plus interest and fees, $812.02 in total.

Keep the structure of that reasoning in mind, because it contains every design rule that follows: what your agent says, you say.

What has changed since the 2015 chatbot

The decision never states what technology powered Air Canada's bot, and that silence is instructive: liability applies whatever the engine. What has changed since is what you can build.

The classic chatbot followed a decision tree: pre-written answers, triggered by keywords, and a dead end the moment a question left the script. An AI agent in the proper sense works differently: it composes its answers from your reference documents (this is called grounding), it can query your systems to check the status of an order or a file, and it applies escalation rules that hand the conversation to a human when specific conditions are met.

Canadian businesses already use them: according to Statistics Canada, among companies using AI in the second quarter of 2025, 24.8% deployed virtual agents or chatbots, the third most common use. The irony of the Air Canada file is that the bot failed exactly where a well-designed agent is constrained: it produced an answer that matched no official company document. Air Canada paid for a single design flaw: a chatbot with permission to improvise policy.

What works in production

Three families of tasks hold up, and they share one property: the right answer already exists somewhere, or the task requires no judgment.

Questions with documented answers. Opening hours, return policies, delivery times, what a package includes, which documents are needed to open a file. If the answer lives in your pages, guides, or internal procedures, a grounded agent restates it faithfully, in French and in English, at 3 a.m. and on statutory holidays. It is the most profitable and least risky use case.

Status lookups. "Where is my order?", "Has my file been processed?" With read access to your systems, the agent answers in seconds questions that eat a disproportionate share of inbound calls. It reads a value and passes it along; it interprets nothing.

Triage before handoff. The agent greets the request, asks the qualifying questions, gathers context, and routes everything to the right person with a summary. The customer has not repeated their story three times, and your team receives a file that is already assembled. We covered these scenarios in our tour of AI agent use cases for SMBs.

What breaks, every time

Three situations again, with the opposite property: the right answer does not exist in advance, or it commits the company.

Policy exceptions. "I know the deadline has passed, but my case is special." Granting an exception is a business decision, not a document lookup. An agent that grants exceptions is writing policy in real time, and the Moffatt file shows what a bot-written policy is worth: it binds you.

Angry customers. Anger calls for listening and real power to make things right. An agent that replies with perfectly uniform calm to a furious customer makes things worse. Detected frustration should be an escalation trigger, never a conversation to continue.

Promises. Refunds, goodwill gestures, delivery commitments, confirmation that a product will suit a specific use. Since February 2024, Canadian case law is clear: your agent's promise is your promise.

The design rules that follow from the ruling

A declared scope. The agent states what it can do and politely declines the rest. A clean refusal beats a plausible wrong answer; the Moffatt file puts an exact price on the second option.

Strict grounding. Every answer must come from your reference documents, and from a single version of them. Air Canada's bot contradicted the very page it linked to: two sources of truth is zero sources of truth. If the information is not in the source, the agent says so and hands off.

Explicit escalation triggers. An exception request, a rising tone, money at stake, a second failure to understand: each of these signals should route the conversation to a human, with the context already gathered. Escalation is part of the agent's job, on the same footing as answering.

Full logging. Mr. Moffatt won because he had a screenshot. Your conversation logs play the same role, in both directions: they are your evidence in a dispute and your raw material for fixing the agent. An unlogged conversation is an undocumented risk.

A ban on inventing policy. This is the rule that summarizes the others. A well-designed customer service agent simply has no authority to state a business rule that does not appear in your documents. No guessing, no rounding, no "showing flexibility".

Where to start

If you keep one thing from the Moffatt ruling, keep the tribunal's test: no difference between a static page and a conversational agent. So treat your agent as a publication, with the same care as your website: current sources, defined scope, regular review of the logs. And start with the tasks where it excels, documented answers and triage, before considering anything else. The same reasoning applies on the phone, where a voice agent follows the same rules with one added constraint, real time.

That is exactly the kind of scoping we do in a first conversation: identify the requests your team actually receives, separate what is documentable from what requires judgment, and design the escalation triggers before the first answer. 30 minutes, no commitment.

→ First conversation, no commitment

This article explains a legal decision to help a small business ask the right design questions; it is not legal advice. Moffatt v. Air Canada, 2024 BCCRT 149, is a British Columbia decision: how it applies to your situation deserves a professional opinion. The full text of the decision, published on CanLII, prevails.