In-Flight Wifi Refunds, Automated End to End, for a UK Airline Group

The client is an airline group based in the UK that sells wifi on board its flights. Passengers who are unhappy with the connection write to a mailbox set up for refunds, in their own words. A typical message runs along the lines of "The wifi was awful, can I have my money back?"

Someone had to read each of those emails, track down the right transaction in the airline's bespoke transaction platform, check whether it qualified, and then either process it or pass it up. All manually, and at volume.

Where the time went

• Staff reading vague emails. Until someone reads a free-text complaint and asks for the purchase details, there is nothing to act on.
• Manual matching. Nothing else could happen until each request was located in the transaction platform.
• Eligibility by eye. For every request, the purchase date, the refund window and the remaining rules were checked by hand.
• Expert time on routine work. Each routine refund handled manually was time taken away from the truly exceptional cases.

Each request now finishes in one of two ways: a refund that has been processed and confirmed by email, or a ticket that puts the case in front of the refunds team.

• Nobody triages by hand between the complaint arriving and the refund or ticket.
• A refund is paid only once the eligibility rules pass, refund window included.
• Nothing fails quietly and nothing is guessed. What the automation cannot finish turns into a ticket.
• The refunds team spends its time on exceptions, not on routine refunds that already qualify.
• No numbers to report. Volumes and timings for this project were not kept, so the figures here describe scope, not savings.

A risk that is easy to miss

A refund robot that guesses does more harm than having no robot. One that fails quietly is worse again: the customer waits, nobody is aware, and the complaint returns angrier. If a transaction cannot be found, the safe outcome is a ticket that hands the case to a person. Build that path first, then the happy path.

A related build that handles many intents: inbox triage and one-click help for agents at a car rental group. There, five types of request share one inbox, and every email gets a classification and a confidence score.