❯ Question and answer
What happens when an AI agent does not know how to answer a customer?
A well-built agent does not make things up: it tells the customer and hands the conversation to a person on the team, with the full context, so nobody has to repeat anything. That handoff is a rule in every Notkode agent. Without it, the company answers for whatever the bot says, as the Air Canada case showed in 2024.
Updated on
What are the numbers, with sources?
- On February 14, 2024, a Canadian tribunal ordered Air Canada to pay CA$812.02 to a customer who relied on wrong information from its chatbot. The decision says it makes no difference whether information comes from a static page or a chatbot. (Source: Moffatt v. Air Canada, 2024 BCCRT 149)
- Anthropic’s documentation recommends explicitly allowing the model to say "I don’t know", which it says can drastically reduce false information. (Source: Anthropic, Reduce hallucinations)
- Anthropic notes that agentic systems trade latency and cost for better task performance, and recommends the simplest solution possible. (Source: Anthropic, Building effective agents)
Why should an agent never invent an answer?
Because the company is liable for what the agent says. In the Air Canada case, the chatbot promised a discount the airline’s policy did not offer, and the tribunal ordered the airline to pay the customer the difference.
How does a handoff to a person work?
The agent detects it is not confident in the answer, tells the customer it is bringing someone in, and opens the conversation to the team on the same channel. The person gets the history and everything the customer has already said.
- Trigger: the question is outside what the agent may answer, or its confidence is low.
- Notice: the customer knows a person is taking over.
- Context: the team receives the whole conversation, without asking for the details again.
How do I know if the agent hands over too often?
By reading real conversations after go-live. If many go to a person for the same reason, the agent is missing information or an integration; if few go and customers complain, the handoff rule is too loose.
Common mistakes
- Telling the agent to "always answer something": that is how it invents prices, deadlines or policies.
- Handing over without the history: the customer repeats everything and the switch to a person becomes friction.
- Having nobody on call to receive handoffs during the hours the agent is answering.
- Never reviewing the conversations handed to people, which show what the agent is missing.
How does Notkode handle handoff?
Every Notkode agent hands the conversation to a person, with its context, when it is not confident in the answer. After go-live, Notkode monitors the agent for 1 week, makes fine adjustments after seeing it operate for real, documents the flows and trains the client’s team.
Frequently asked questions
Does the customer notice the switch from bot to person?
They should. The agent says it is bringing someone in, and the person takes over on the same channel with the history at hand.
What if nobody is available at that moment?
The agent says when the team will reply and logs the request for them. What it does not do is invent an answer to close the conversation.
Can the agent learn from the conversations it handed over?
Yes, with review. Handed-over conversations show what is missing: new information, a rule or an integration. A person decides what goes in.
Is the company liable for what its agent says?
Yes. In Moffatt v. Air Canada, in 2024, the tribunal held that the company is responsible for all the information on its website, including what its chatbot says.
Last verified on