← All work

Case study 01 · Enterprise · anonymised

Airline AI assistant

A customer chat assistant at a European airline group, growing from answering questions into a travel companion that knows who you are and can act for you. I lead the AI product work on the loyalty and personalisation side.

Role
AI product lead
Scope
Strategy, architecture, cross-team alignment
Status
Ongoing
Note
Anonymised, numbers stay internal
Illustration: an orchestrator routing to loyalty, status, FAQ and safety-check agents
01

Background

Airline chat assistants are good at answering questions. Customers increasingly expect one that recognises them and does things for them: knows their status, explains their miles, guides them through a trip.

That knowledge sits in different systems owned by different teams, and the loyalty side had never operated in a conversational channel at all. Bolting features onto one chatbot doesn't scale past the first few.

02

Solution

Move from one chatbot to an orchestration layer that routes each request to specialist agents, each owned by the team that owns the data.

  • Loyalty becomes an agent, not a feature. The assistant calls it the same way it calls any other capability.
  • Customer signals get a shared interface (MCP), so any team's agent can use them, not just ours.
  • Capabilities go in by risk: the most clearly defined first, open-ended questions stay in discovery, recommendations wait until the basics work.
03

Receipts

The numbers stay internal. What I can show is the reasoning behind the calls I made.

Sequence by risk, not excitementThe first capability is the one that's easiest to define and test. Recommendations are the most exciting and the easiest to get wrong, so they wait.
Test a prototype, not a hypotheticalAsking people what they'd want from an assistant tells you little. Watching them use one tells you a lot.
Build once, reuse everywhereA shared signal layer costs more up front and saves every other team from building its own.
Say no out loudDeprioritising a capability is a decision, so it gets written down with the reason, not quietly dropped.
04

What's next

  • User research with a working prototype in customers' hands.
  • More teams plugging their agents into the same orchestration layer.
  • An idea I'm exploring: proactive guidance for people who fly rarely and need help most.