How AI agents avoid hallucinations: knowledge bases, citations and the Air Canada lesson

The biggest fear business owners have about AI agents is simple: what if it makes something up? It is a reasonable fear. Large language models can produce fluent, confident answers that are wrong. In customer service, a wrong answer is not just embarrassing. It can be legally binding.

The Air Canada lesson

In November 2022, Jake Moffatt asked Air Canada’s website chatbot about bereavement fares after a death in his family. The chatbot told him he could book a regular ticket and apply for the reduced bereavement rate within 90 days. That was not Air Canada’s policy. When he applied, the airline refused.

He took the case to British Columbia’s Civil Resolution Tribunal. In Moffatt v. Air Canada, 2024 BCCRT 149, the tribunal ruled against the airline.^[1] Air Canada had argued that it could not be held liable for information provided by its chatbot, effectively treating the chatbot as a separate entity. The tribunal rejected that, and ordered Air Canada to pay a partial refund of CA$650.88 plus interest and fees.

The amount was small. The principle was not: your business is responsible for what your AI agent tells customers.

What is an AI hallucination?

In a widely cited survey published in ACM Computing Surveys, Ziwei Ji and colleagues define hallucination as generated content that is nonsensical or unfaithful to the provided source content.^[3] They describe it as a pervasive problem across language generation tasks, from summarisation to dialogue, not a rare glitch.

Hallucinations happen because a language model is trained to produce plausible text. When it does not have the facts, it can produce something that sounds right instead of saying "I don’t know".

How to stop an AI agent from guessing

1. Ground answers in your own documents

Retrieval-augmented generation (RAG) was introduced by Patrick Lewis and colleagues at Facebook AI Research and University College London, presented at NeurIPS 2020.^[2] Instead of relying only on what a model memorised during training, the system first retrieves relevant passages from a trusted knowledge source, then generates an answer based on them.

For a business, that knowledge source should be your approved content: prices, policies, opening hours, services, terms, FAQs.

2. Cite the source

Every answer should be traceable to a specific document or passage. If staff review a conversation, they should see exactly where the answer came from.

3. Set a confidence threshold

If the retrieved information does not clearly answer the question, the agent should not try. Below a set confidence level, it says honestly that it does not have that information and offers a next step.

4. Offer a human path

"I don’t have the answer to that, but I can transfer you to our team, take a message, or arrange a call back." That is a far better customer experience than a confident wrong answer.

5. Log and close the gaps

Every unanswered question is a signal. Log it, review it, approve an answer, and add it to the knowledge base, so the agent can answer next time.

Questions to ask any AI vendor

1. Where does the agent get its answers: model memory or our approved documents?
2. Can we see the source of every answer?
3. What happens when the answer is not in our content?
4. Can we set or tune the confidence threshold?
5. How do we review and approve new answers?
6. Who is responsible for keeping the content up to date?

This is the design behind the Northline AI knowledge base: answers only from approved content, with citations, a confidence threshold, a human handover when it is unsure, and a gap log your team can review.

The bottom line

You cannot make a language model infallible. You can make sure your business’s AI agent only speaks from your approved information, shows where every answer came from, and says "I don’t know" when it should. After Air Canada, that is not just good practice. It is basic risk management.