Air Canada paid for its chatbot's wrong answer. Here's why every business chatbot needs regression tests, and what that means in plain words.
You raised a price on Monday. You updated the website, told the team and moved on. On Thursday night a customer asks your chatbot about it, and the bot cheerfully quotes the old number. Now you get to choose: accept the old price, or explain why your own website gave someone the wrong one.
Big companies have had this conversation in public. In February 2024, a British Columbia tribunal ordered Air Canada to pay a customer C$812.02 after its chatbot described a refund policy that the airline's own website contradicted. If there's a chatbot on your site, chatbot accuracy testing isn't an engineering luxury. It's how you find out the bot is wrong before a customer does.
I build AI chat agents for a living, so I'm not neutral. That's why I'd rather show you where these tools go wrong.
What actually happened to other people's chatbots
Air Canada. A customer asked the airline's chatbot about bereavement fares. According to the tribunal decision, the bot said he could apply within 90 days of buying his ticket. The bot even linked to Air Canada's bereavement page, which said the policy didn't apply once travel was completed.
Air Canada argued it couldn't be held liable for what the chatbot said. The tribunal called that argument "remarkable" and said the chatbot was "still just a part of Air Canada's website." The airline paid C$650.88 in damages, plus interest and fees.
New York City. The city launched its MyCity business chatbot in October 2023 on Microsoft's Azure AI. In March 2024, The Markup found it telling business owners they could take a cut of workers' tips and that landlords didn't have to accept Section 8 housing vouchers. Both are wrong under city law. It carried a small disclaimer, and gave different answers to the same question.
Cursor. In April 2025, users of the coding tool Cursor kept getting logged out when they switched computers. A support bot called "Sam" told them this was a new one-device policy. There was no such policy. Co-founder Michael Truell posted an apology on Reddit: "We have no such policy," according to The Register. By then, Fortune reported, some customers had already cancelled.
In fairness, the models themselves aren't perfect even under ideal conditions. Vectara's hallucination leaderboard asks AI models to summarize documents using only the facts in each one. In the September 22, 2026 update, the best of 108 models still added unsupported facts 1.8% of the time, and 50 of the 108 did so more than 10% of the time. Caveats: Vectara sells AI search tools and built the grading model, and summarizing isn't chatting. Still, the error rate is never zero.
The price-change problem, and why fixes break things
Most mistakes a small business chatbot makes aren't exotic. They're stale.
A chat agent answers from a knowledge base, which is a library of your pages and documents it reads before replying. If your website changes and that library doesn't, the bot is confidently out of date. In our own product, the website rescan setting is "Manual only" by default, with daily, weekly or monthly options. I'm saying that plainly because it's exactly the kind of setting that bites. If your prices change often, set a rescan schedule, or click Rescan the day you change them.
Then there's the second trap. You fix the price in the agent's instructions, and the same edit quietly changes how it answers about parking, pets or deposits. Nobody checks, because nobody thinks to.
That's what regression tests are for. A regression test is a saved question with a known right answer that you re-run after every change, to check nothing that used to work has broken. Chatbots need it more than most software, because the same AI can answer the same question differently twice in a row.
How we handle it: answers you can check, changes you can test
Our answer is a simple principle: answers you can check and improve. Here's what's in the product today.
Show AI Details. On any AI reply in your conversations, you can open Show AI Details and see which agent and version wrote it, the instructions it followed, any tools it called, and the knowledge base context it used, with a confidence score. When the bot quotes an old price, you can see where that number came from.
Teach the agent. When a staff member answers something in the mobile app that the AI couldn't, they can tap Teach the agent this answer. It becomes a short question and answer in the knowledge base within a minute or two, with personal details left out. To be clear, there's no approval step here, so owners and admins should check those learned answers now and then, and delete any that go out of date.
Improvements, tested before they go live. The system flags chats with low ratings or failures and proposes a fix. You keep a regression suite: real questions with verified expected answers, plus at least one "held-out" case the fix was never shown. Before you can publish, cases that depend on the AI run three times against both the current version and the proposed one. Critical checks must pass every run, and quality must match or beat what you have now. Then you Approve and Publish. Approval alone changes nothing. If something still looks off, Roll back returns to the previous version.
So in the price scene, you'd add "How much is the two-bedroom?" with the new price as the expected answer. If a later change drifts back, it fails before a customer sees it.
Honest limits: rollback doesn't undo emails or CRM updates already sent, and some things, like external tools and webhooks, can't be captured in a test yet. The details are in Improve your chat.
The bigger pattern
The Air Canada decision put it simply: a company has to take "reasonable care" that what it tells customers is accurate. A disclaimer didn't save MyCity's reputation, and a polite bot name didn't save Cursor's.
The responsibility is shared. We can build tools that show their work and test their changes. You still have to keep your website current and write down the answers that matter most. Testing doesn't make a chatbot perfect. It makes mistakes visible early and keeps fixed ones fixed.
Want to see a regression test run on your own prices and policies? Book a 20-minute demo and we'll walk through it on your site.