Customer care agent
An AI agent that answers airline passengers on live chat, X and Instagram, and knows when to step aside.
- Status
- Live, in extended testing
- For
- A leading low-cost airline
- Channels
- Live chat, X, Instagram
- Model
- Claude Sonnet
Three messages, three different decisions
Scripted examples, not real passengers
What the agent decided, and why
The problem
An airline's inbox fills with the same questions around the clock, in more than one language. A small share of those messages are serious enough that an automated answer would be the wrong one.
The agent has to be quick with the first kind and get out of the way for the second.
What it is built to do
- Answer from approved sources only. It checks curated question-and-answer pairs first, then the knowledge base, then documents, then the airline's own website.
- Check identifiers before using them. Booking references, ticket numbers, baggage references and flight numbers each have a format. If what the passenger typed does not match, the agent asks again and never guesses.
- Hand sensitive cases to a person at once. A defined list of categories goes straight to a colleague, with no automated answer in between.
- Make no claims about routes. It never confirms or denies that a route exists. It points to the flight network page, which is always current.
- Reply in the passenger's language.
How it was tested
- Edge cases. Messages that sit between two topics, or that no rule was written for.
- Long, mixed conversations. A passenger who starts with baggage, moves to a refund and ends with a complaint.
- Manipulation attempts. People trying to talk the agent out of its rules.
- A fixed set of reference questions, asked again after every change, to measure whether its answers are still correct.
What the tests found went back into the rules. One example: the agent no longer sends filler text while it is looking something up, so passengers get one clear message instead of two.
What comes next
The planned second phase connects the agent to the airline's customer system, so it can look up a booking instead of only asking about it.
Want something like this for your team?
Tell us what you do by hand today. We will tell you what can be built.
Email okan@okxproject.com