Reef TRHThe airline's worst reviews were not about legroom, delays, or a bag that went to the wrong city. They were about the inbox. A passenger would email to move a flight, change a payment, or ask for help — and by the time a support rep worked down to their message, the flight had already left. The reply, when it finally came, was no longer help. It was a refund conversation.
For a carrier running thousands of domestic and international flights, with a reputation built over decades, that lag was doing quiet, steady damage. When we pulled the negative reviews apart, the overwhelming majority weren't about flying at all. They were about waiting for an answer that came too late to matter.
Support itself wasn't bad. It was slow. Reps replied to every email by hand from a knowledge base of set questions and answers, and anything the base didn't cover went up to a manager and waited. The quality was fine; the clock was the enemy. So the thing we were asked to build wasn't smarter answers. It was speed — a useful reply while the passenger still had a flight to catch.
What the agent actually does
We connected an AI agent to the support inbox and to the airline's own systems, the CRM above all. Before it writes a word, it knows what a human rep would otherwise have to dig for: whether the sender is a real, ticketed passenger, when they're flying, and whether they've been in touch before. Then it reads the incoming email and, within minutes, drafts a reply grounded in the same knowledge base the reps had always used.
At first a person approved every message before it left. As the drafts earned confidence, the routine ones started sending on their own — a baggage-policy question, a confirmation query, a "how early should I arrive" — closed without a human ever touching them.
The hard part was knowing when to slow down
Speed was the goal everywhere except where it wasn't. The moment a request touched money — an upgrade, a change of payment method, a cancellation — the agent deliberately stopped being quick. It sent a short holding note to say it was looking into it, and drafted a reply that would not leave the building until a person had approved it.
A brand-new enquiry got a lighter touch. A paying, ticketed passenger got more care, not less. And anything touching a future booking moved into a sensitive tier that often waited for a human. The system's real intelligence wasn't in the answers it rushed out — it was in the ones it held back.
How it earned its way through the aircraft
Nobody hands a bot the whole airline on day one, least of all a company measured in decades. We started deliberately small: a single domestic route to one US destination, with the agent running alongside the human reps rather than replacing them. It had to prove itself in the least risky corner of the operation before it earned any more ground — the way we'd put any first system into production.
It did. After a few months the pilot widened to more domestic destinations, then to international flights, and finally to the premium cabins — business and first — where the expectation is flawless, personal service and the one unacceptable outcome is a passenger sensing they're talking to a machine. That it was trusted there last, and trusted there at all, was the whole point.
What changed
The number the airline actually cared about wasn't first-reply time; it was time to full resolution — the message that closed the passenger's problem for good. It fell by around 70%. The reviews that had been about waiting slowly started to be about something else. The fix for a reputation problem turned out not to be an apology campaign. It was a clock.
None of this was a modelling breakthrough. It was the unglamorous work of wiring an agent into real systems, drawing hard lines around money, and earning autonomy one route at a time — the same distance between a demo and a dependable system that separates the AI that ships from the AI that stalls, and the discipline behind every pilot that actually survives contact with production.
Common questions
Does an AI support agent replace human reps? No. This one started alongside the team and never took over the hard calls — anything touching money or a future booking still routes to a person. It took the volume and the waiting off the humans, not the judgement.
How do you keep it from doing something costly, like a wrong cancellation? Guardrails tied to risk. Requests that touch money or a future booking get a holding reply and wait for human approval; paying, ticketed customers get more caution than anonymous enquiries. It's allowed to be fast only where being wrong is cheap.
How do you roll this out at a large, cautious company? In stages, smallest risk first — one route, running beside the humans, then more destinations, then international, then the premium cabins last. Each expansion had to be earned by the one before it.
The last place the agent reached was first class, the cabin where you would least expect a passenger to tolerate a machine. It got there the same way it got everywhere else: not by being clever on day one, but by being trusted a little more each month. The airline didn't buy an AI that could answer everything. It bought one that knew what it could answer now, and what to leave for a person — and earned the rest.