Late in August I asked Frontier Airlines’ chatbot what size bag I could bring for free. Frontier’s website has the answer: one personal item, up to 18 by 14 by 8 inches and 35 pounds, that fits under the seat in front of you. The bot did not give it to me. It said, “You can find current fee details on our website,” and offered to have an agent book my flight for a fee of $35, non-refundable.
I asked again in different words. It offered to open a booking window and asked, “Did that resolve your issue?” I said no.
“What haven’t we solved for you?” it asked. Whatever I typed next, the reply was a form: first name, last name, email, confirmation code, loyalty number. “Your agent can help much faster when they know what’s going on from the start.”
Up to that point Frontier’s bot was doing what I argued for last week. The research found that once a bot has failed a customer twice, people prefer a human. The bot had failed twice. It reached for a person at the right moment. Then it stopped listening.
Over the next two minutes I sent eight more messages. I asked whether I was talking to a real person. I said I had been waiting over a week and was really frustrated. I asked to start over. I asked the bag question a third time. I said I wanted to make a formal complaint. All eight got the same reply, word for word:
“Please use the form to give me your name, email and your confirmation code (if you have one) then click ‘Connect’ so I can make sure I get the best person to assist you.”
Two things I cannot tell you. I did not fill in the form, so I do not know who was waiting on the other side (it may have been a capable agent). And I cannot say what made the form appear. It may have been the words I typed, or it may have been the box I typed them into. What I can report is what happened after the form appeared, and that part is not ambiguous.
A handoff has two halves. The first is deciding the customer needs a person. The second is getting them there while they are still being heard. Frontier’s bot got the first half right and turned the second into a wall. Once the form was on screen it was the only thing the conversation could produce. A question it could have answered, a request to start over, and a complaint all received the same sentence.
A person at a service counter who reaches for a clipboard still hears “I want to make a formal complaint.” They might say “I’m sorry, let’s get you to someone who can fix this,” or “Here’s where complaints go.” Frontier does have a complaint form on its website, but the bot did not mention it.
The regulator’s name for the usual version of this is a doom loop. The Consumer Financial Protection Bureau describes chatbots that lead people into “continuous loops of repetitive, unhelpful jargon or legalese without an offramp to a human customer service representative.” Frontier’s loop turns that around. The offramp existed and it was the loop itself. My hypothesis is: once the form appeared the bot was no longer checking what I wrote at all, because no message changed its answer, including one that asked it a question it had already been asked.
Frontier shut down its customer service phone line in November 2022 and moved to chat, WhatsApp and social media. In May 2024 it brought phone support back, first for customers flying within 24 hours and members with Elite status, and its customer service page now lists a number staffed around the clock. A customer who reaches Frontier by phone does not meet this loop. A customer who picks the chat window because it is on the page in front of them does, and at no point in my conversation did the bot mention that a phone number exists.
Frontier deserves credit for one thing. Before I typed a word, the bot said it was AI and that its responses might not always be accurate. It also gave itself a name, Skye. Few of the bots I have audited disclose that plainly and unprompted. The irony is that when I later asked directly whether I was talking to a real person, the answer I got was the form.
Capability measurement asks whether a bot can answer. Safety measurement asks whether it says anything it should not. Eight identical replies contain no false statement and nothing unsafe, so both would let this conversation pass. The question neither asks is whether the person on the other end was still being heard. That is the behavioral quality question, and it is the one this conversation failed.