You may remember the Air Canada case: the airline’s chatbot gave a customer incorrect information about bereavement fares, the customer booked based on that information, and when Air Canada refused to honor it, a court held the airline liable for the chatbot’s mistake.
It’s an instructive case. Not because it’s unusual — it’s a preview of what happens when language models are put in charge of anything involving numbers, prices, or calculations without a reliable engine underneath.
LLMs are excellent at language. They are not reliable calculators. They don’t look up live data. They generate plausible-sounding answers — and in the domain of pricing, availability, deadlines, and quantities, “plausible-sounding” isn’t good enough.
The solution isn’t to distrust AI. It’s to use the right tool for each part of the problem: language models for language, databases for facts, deterministic engines for calculations. An AI that knows when to delegate is far more trustworthy than one that tries to do everything itself.
The Air Canada chatbot didn’t fail because AI is bad. It failed because the wrong component was asked to do something it isn’t designed for.
Originally published on LinkedIn.