Colleen McCarthy, a Delta Platinum Medallion, thought she had secured her trip home. She wanted to cancel only the outbound leg of a New York to Los Angeles award ticket. Keep the return. Drop the departure.
Simple request? Usually, yes.
But she was talking to Delta’s AI Concierge, not a human.
The AI replied: “Your JFK to LAX flight has been cancelled, and your LAX to JFK flight is still booked.”
McCarthy believed it. She saw the confirmation.
Then Delta’s actual reservation system canceled both legs. It redeposited her 32,400 SkyMiles. It refunded the $11.20 in government taxes. And it told her that if she wanted to go home, she’d need to book a new one-way award ticket from LAX to JFK.
Cost? Nearly double what she had originally.
This isn’t just a glitch. It’s a look inside the messy beta test of Delta’s AI Concierge taking dangerous control of live tickets.
The “Safe” Confirmation Was a Lie
Look at the refund confirmation McCarthy posted.
The receipt lists DL752 (JFK-LAX) and DL938 (LAX-JFK). The refund number is “null.” It’s almost poetic for a live beta.
The text message she received explicitly stated the return remained. The system receipt contradicts that text. The AI hallucinated a specific outcome it thought she wanted, or its language model interpreted her request as a full cancellation while her intent was a modification.
If you ask a machine to change a roundtrip into a one-way, you might get a different price. One-ways often cost more. Delta should have quoted that price. It should have asked for confirmation.
Instead, it refunded everything and left her stranded in LA with a digital receipt that said “safe” while the backend said “cancelled.”
How Delta’s AI Concierge Actually Works
Delta launched the Concierge beta in January 2025. It sits inside the Fly Delta app. You can text or speak to it.
Early versions just answered questions. Where is my bag? What are my benefits? Where are my eCredits?
Then Delta gave it hands.
Now, Concierge can take action. It can cancel trips. It can offer rebooking during disruptions. In June, Delta’s marketing chief said rebooking and cancellation tools were “working really well.” The goal? Go from 5% of app users to 100%.
It is still labeled “BETA.”
Where the Logic Breaks Down
A human agent knows that a roundtrip ticket is one document. You can’t just yank one half out without affecting the whole structure. Sometimes, removing an outbound leg requires exchanging the entire ticket.
The AI doesn’t know “structure.” It sees a request. It maps it to a tool.
Somewhere in the chain between the language model and Delta’s reservation system (DCS), the command got translated as “cancel all.” Or the system treated the inbound AI request as a full void. The conversational layer then lied, stating the return was still active.
Is the AI broken? Or is it just too good at sounding confident?
A bad human agent ruins one ticket. A defective AI workflow ruins thousands.
Natural language feels authoritative. It feels like customer service. It doesn’t feel like a code error. But it is neither until you look at the backend.
This Isn’t New. It’s Just Scalable.
I wrote about this over a decade ago. A Delta agent adjusted a passenger’s ticket and accidentally voided the entire e-ticket. The return on a partner airline vanished.
United agents have accidentally cancelled future segments during irregular operations. Virgin Atlantic’s staff once cancelled a confirmed LATAM business award without permission. The space was gone. Virgin said it couldn’t fix it. Eventually, they paid for the replacement ticket, but only after a public mess.
Human agents misunderstand. They key in the wrong commands. They cancel space before confirming they can rebuild it. They give wrong information because they’re tired or overworked.
AI promises consistency. It promises speed. It promises to work when the call centers collapse.
But Delta’s implementation is flawed. Right now, it’s worse than the median human agent because it scales human error across thousands of users instantly.
Other Airlines Are Making Similar Mistakes
American Airlines has tested an AI assistant for weather rebooking. Air India’s “AI.g” handles cancellations. Qatar Airways’ “Sama” does conversational booking.
Air Canada tried a similar stunt. Their chatbot gave incorrect info. Air Canada tried to shift the blame to the bot. A Canadian tribunal rejected that argument. The airline is responsible for what it sells through its website, even if a software bug wrote it.
Who Owns the Failure?
Delta’s AI didn’t just make a mistake. It made a mistake and then refused to correct it.
When McCarthy contacted live reservations, they told her they couldn’t restore the trip. They demanded she pay double.
This is a human failure at the end of a technical chain. Delta’s systems created a mess. The AI gave false comfort. Then, the humans refused to fix it.
Is it a choice to not empower agents to override the bot? Is it a choice to not spend time escalating?
Or is Delta just testing how far they can push customers before the backlash gets too loud?
McCarthy got her miles back. Barely.
The flight? Still canceled.
