When a Chatbot Invented a Policy — and a Tribunal Sided With the Customer
In 2022, Jake Moffatt had just lost his grandmother. He went to Air Canada's website, where the airline's chatbot told him he could buy full-fare tickets and claim the bereavement discount afterward, within 90 days. That was wrong. The real policy doesn't work that way — the bot had simply made the answer up. A British Columbia tribunal ordered Air Canada to pay the difference, ruling the company had "not taken reasonable care to ensure its chatbot was accurate."
That's the word worth sitting with: accurate. An AI voice agent answering your phone is a wonderful thing — right up until the day it states, with flawless confidence, something that isn't true. Stanford researchers have measured that language models left ungrounded give a wrong answer in 15 to 30% of customer-service queries. For a small business, that's not an abstract number. It's roughly one in three or four callers walking away misinformed.
The good news? In 2026 we finally understand why these errors happen — and, more usefully, how to spot them before you hand your phone to an AI. Here's what the recent data reveals, and what it means for a business anywhere from Halifax to Vancouver.
Trend #1: Why an AI 'Makes Things Up'
The technical term is hallucination. As IBM explains, a language model doesn't "look up" a fact in a database — it predicts the most plausible next word from patterns it saw in training. Most of the time it lands on the truth. But when a question falls outside its comfort zone, it fills the gap, and a fake policy, a fake price or a fake set of hours comes out with the same confidence as a real one.
The root cause is almost always the same: the agent answers from its general knowledge instead of YOUR information. It has never seen your price list, your calendar or your cancellation rules — so it guesses. And a caller on the phone has no way of knowing that the polite, confident voice they're hearing is improvising.
That's exactly what happened to Air Canada, and it's what happens to any agent you drop onto a line without guardrails. So the question isn't "can the AI get it wrong?" — it can. It's "what stops it from getting it wrong?"
Trend #2: Grounding Collapsed the Error Rate
This is the most important advance of the past year, and somehow the least visible in the marketing. It's called grounding: instead of letting the AI answer from memory, you force it to pull every response from a verified source — your knowledge base, your booking calendar, your price list. The agent no longer speaks "off the top of its head"; it reads your information and says only what's there.
The numbers are striking. A well-grounded system drops the hallucination rate below 5%, versus 15 to 30% for a model left on its own. Adding a grounding layer alone cuts errors by 60 to 75%. But here's a detail few vendors mention: the single biggest predictor of the error rate isn't the model you pick — it's the health of your knowledge base. One Zendesk report estimated that 30% of a typical help centre's articles are more than twelve months old. A perfectly grounded agent pointed at stale information will quote an old price with total confidence. The architecture does everything right; the source is wrong.
The lesson for a small business is a freeing one: reliability doesn't come from a "smarter" AI, but from an agent fed your real, up-to-date information. It's a matter of configuration, not magic.
Trend #3: In Canada, You're Liable for What Your AI Says
The Moffatt v. Air Canada decision set a precedent every business should know. The tribunal rejected the airline's argument that the chatbot was "a separate legal entity responsible for its own actions." As the American Bar Association summarized it, the company remains fully responsible for the information its conversational agent provides — whether it comes from a static web page or a talking AI.
Translation for your business: if your voice agent promises a discount that doesn't exist, confirms a booking you can't honour, or gives wrong product information, you own the outcome — not the software vendor. That's not a reason to run from the technology. It's a reason to choose an agent that won't wander outside your verified information.
And it shifts the real buying criterion. The question is no longer "which AI has the nicest voice?" but "which AI refuses to invent when it doesn't know?"
Why Voice Raises the Stakes on Accuracy
A wrong answer in a written chat, a customer can re-read, question, screenshot. On the phone, they can't. Speech moves fast, the tone is confident, and the caller leaves believing what they just heard. That's what makes a voice hallucination more dangerous than its text cousin: there's no visible record to check in the moment.
Add the pressure of real time. A good voice agent answers in under a second to feel natural — something we broke down in our piece on AI voice agent response time. But that speed must never come at the cost of accuracy. A well-built agent would rather take an extra half-second, or hand off to a human, than deliver a fast, wrong answer.
In short, on the phone a bad answer doesn't just cost a satisfaction point. It can cost a customer, a mistaken booking, or worse — a commitment you never actually made.
What This Means for Canadian Businesses
For a Canadian company there's an extra layer: language. An agent has to be not only accurate, but accurate in both English and French, without crossing its sources or mistranslating a price or a condition. Canada is officially bilingual and its cities speak dozens of languages, so an agent that greets in English, switches to French on request, and — crucially — stays grounded on the same verified information in either language is a real advantage, not a nice-to-have.
The other layer is data. When an agent is grounded on your systems, it touches customer information, so reliability and privacy travel together. We turned that into a practical checklist in our data-privacy questions to ask before trusting an AI voice agent. A serious vendor gives you a clear answer on both fronts — where the answers come from, and where the data goes.
Put plainly, accuracy isn't a technical luxury. It's what separates an agent that builds your reputation from one that quietly erodes it, one call at a time — whether your customers are in Toronto, Calgary, Winnipeg or Moncton.
How to Tell an Accurate Agent From a Risky One
You don't need to be an engineer to ask the right questions. First signal: is the agent grounded on YOUR information? A serious vendor will walk you through how it connects your price list, hours and calendar — and confirm the agent only answers from those sources. If they mostly tout "a super-smart model" and go quiet on your data, be careful.
Second signal: what does the agent do when it doesn't know? The right answer is that it admits it and transfers to a human rather than guessing. Deployments that enforce this confidence threshold cut customer-facing errors by 70 to 85%. An agent that never says "let me connect you with someone" is an agent that, sooner or later, will invent.
Third signal: does the voice inspire trust without pretending to know everything? A good agent sounds natural — a topic we explore in our guide to making a voice agent sound human instead of robotic — while staying honest about its limits. Natural and cautious aren't opposites; together they're the mark of an agent that's been set up properly.
Predictions for 2026-2027
First prediction: "accuracy" becomes the headline selling point for voice agents, ahead of voice quality and price. Vendors will advertise reliability rates the way they advertise latency today.
Second prediction: the public debate over liability intensifies. After Air Canada, more decisions will confirm that the business answers for its AI — and the best-prepared SMEs will be the ones that grounded their agent from day one.
Third prediction: the bar rises for everyone. An agent that invents one answer in three will become as unacceptable as an employee who hangs up on callers. The new standard is an agent that can say "I don't know — let me transfer you."
Frequently Asked Questions
Can an AI voice agent really get it wrong enough to hurt my business? Yes, if you let it answer from memory. That's precisely why a serious deployment grounds it on your verified information and forces it to transfer when unsure. Configured well, the error rate falls below 5%.
How do I know if my agent is 'grounded'? Ask the vendor where the answers come from. They should be able to name your sources: price list, calendar, knowledge base. If the answer stays vague, that's a red flag.
Will the agent sometimes say 'I don't know'? Ideally, yes — and that's a good thing. An agent that acknowledges its limits and hands off to a human protects your reputation far better than one that guesses to look competent.
Is accuracy the same in English and French? It has to be. An agent built for a bilingual market stays grounded on the same verified information in both languages, without mistranslating a price or a condition.
The Bottom Line
The 2026 data tells a reassuring story: voice AI is no longer doomed to hallucinate. An agent grounded on your real information, honest about its limits and configured with care answers correctly, call after call. The problem was never "AI gets things wrong"; it was "unguarded AI gets things wrong."
Want to know whether an agent would hold up for your business — no jargon, no empty promises? Get a personalized analysis and see how the Agent IA Vocal team configures an agent that stays true to your information — and knows to say "let me transfer you" when it should.