Recently, one of our own solutions consultants had his flight home cancelled. What followed will sound familiar to anyone who runs an airline contact center: an hour on hold with an IVR that couldn't parse his booking reference, 30 more minutes with a human agent, a rebooking onto a worse flight than one he found himself, and no word about hotel or meals. He got home 22 hours later. The route is normally a 2-hour direct.
Here's the uncomfortable part: the airline did nothing wrong. The weather wasn't their fault, and no compensation was owed. They followed policy, and they still lost a customer for good. That gap between policy followed and passenger taken care of is what our recent session, From cancelled to confirmed, is about. The full recording is now available on demand, and this post covers what it shows.
Disruption is a funnel, and it leaks at three moments
Irregular operations (IROPS) compress thousands of individual emergencies into a few hours. Across years of airline deployments, we've seen the journey break at three specific moments.
Engage. A disrupted passenger who hears an IVR menu assumes nothing good is coming and starts mashing zero. Recognizing the flyer from their phone number, acknowledging the disruption before they explain it, and verifying identity quickly changes that math. Even done well, though, roughly 30% of callers still ask for a human here.
Intelligent matching. This is where rebooking succeeds or dies: which flights get offered, in what order, and whether the system understands that a family of four with paid seat upgrades is not the same reservation as a solo economy traveler. About 35% of callers drop off at this stage when the options miss.
Close and follow up. The flight is only part of the disruption. Hotel, meals, and compensation eligibility need to happen in the same conversation, before the passenger has to go hunting for them.
What the demo call actually shows
The centerpiece of the session is a live call with a fictional carrier's Ada-powered AI agent, on Voice. The agent recognizes the caller from their phone number, acknowledges the cancellation before they explain it, and offers to handle three things: rebooking, hotel and meals, and the compensation claim.
Then it works the problem. It weighs 52 itineraries across the airline's own and partner network. When the caller declines a leg, it re-ranks and comes back with a later departure that keeps their business cabin. It confirms the rebooking with the aisle seat intact, books a hotel located inside the terminal so there's no ground transfer, issues meal vouchers, and files a $700 care credit request, routed to a human specialist who reviews only the approval decision with the full call context attached. A summary arrives by SMS in the caller's language.
Elapsed time: about three and a half minutes. In the real-world version of this story, the passenger was still on hold.
The rebooking your API suggests isn't the one your passenger wants
One finding from the session deserves its own heading. In a deployment we described, the airline's own booking API returned technically valid rebooking options, ranked in an order passengers hated. It buried same-day flights. It proposed layovers that were either alarmingly short or five hours long.
The fix was an intelligence layer on top of the existing API: re-ranking results for same-day arrival, sensible connections, seat and upgrade retention, and keeping families booked together. That ranking logic gets built with each airline, because every network and every policy is different. Once it's proven on Voice and Messaging, the same logic can feed the airline's own self-service rebooking.
"What stops it from promising a refund it shouldn't?"
The liveliest part of the session was the Q&A, where teams from 19 airlines pushed on the question that matters most: control. The answer has three layers, and none of them is "trust the model."
First, the critical steps run deterministically: identity verification, refunds, and cancellations. You decide the exact wording and the exact conditions, the way your non-AI systems work today. Second, before any response reaches a passenger, a separate check compares it against your instructions; anything non-compliant gets sent back and regenerated. Third, every conversation passes through a flagging and QA review afterward, so anything unusual is surfaced for a human to look at.
The session also answers a question we hear from every CX leader who runs disruptions today: how do you measure an experience passengers rarely fill in surveys about? Alongside the post-call survey, Ada analyzes the full transcript of every conversation to produce a synthetic NPS: a read on how the interaction actually went, for every call rather than the few who respond.
It sits in your stack, not in place of it
None of this requires ripping anything out. Ada sits between your existing contact center platforms (NICE, Amazon Connect, Genesys, LivePerson) and your reservation systems, connecting to Sabre, Amadeus, or the layer your team has already built on top, through APIs and MCP. When a conversation needs a human, it hands off with full context.
Watch the session
The recording, including the full demo call and the Q&A on guardrails, non-partner rebooking, and building the business case, is available on demand: https://www.ada.cx/event/on-demand/from-cancelled-to-confirmed-see-how-agentic-ai-handles-airline-travel-disruptions-end-to-end-on-demand/
And if you'll be at World Aviation Festival in Lisbon this October, come find us at booth #1-221. We'd rather show you this on your own routes than describe it.