When Alison Gil turned to United Airlines' chatbot for a straightforward answer about her $200 travel credit, she got a response in seconds. It was also completely wrong. The bot told her the credit would last five years. Weeks later, an email from United warned it would expire within months. The incident exposes a gap that customer support professionals face daily: AI chatbots can reduce handle times, but their errors can destroy trust and create escalations that human agents must then untangle.
Gil had spelled out exactly what she needed. "I thought that I'd get a fast answer. Spelled out exactly what I was looking for," she said. A screenshot shows the chatbot stating her TravelBank credit "stays active for 5 years from the date it's deposited." When a reminder email arrived setting the expiration for September 27, 2026, she pushed back. United's response offered no explanation for the discrepancy. "No, they were just like, well, sorry, the bot told you that, but it's one year and we can't change it," Gil said.
When the bot becomes a liability
Gil's frustration points to a structural problem. "It's just frustrating that there seems to be no accountability in it. Their chatbot can apparently give out whatever answer it wants," she said. She is not alone. Online forums contain reports from other customers who received inaccurate information from airline chatbots, including one Chicago-area traveler who said a question about a refund fee led a United chatbot to cancel their entire trip.
V.S. Subrahmanian, a professor of computer science at Northwestern University, said mistakes by AI systems are not unexpected. "We know human operators will make mistakes. We expect AI to make mistakes as well, but there should also be some responsibility for those mistakes," Subrahmanian said. He has worked in the AI field for decades and sees a missing piece in how companies deploy these tools.
"I think the one thing that chatbots lack, which many humans have, but not all, of course, is the ability to dynamically listen to what you're saying and tell if it sounds right or not," Subrahmanian said. "There needs to be some regulatory structure that has some system of checks and balances." For teams managing AI for Customer Support, that gap between automation and judgment is where most escalations begin.
What United did next
After United apologized for the confusion and told Gil the credit could not be extended, NBC 5 Responds contacted the airline. United acknowledged the failure. The airline said its "AI chatbot provided this customer with the wrong information for her situation. Our teams work to correct these issues as quickly as possible when they occur, but we do provide a disclaimer informing customers that information may not always be complete or relevant."
United explained that TravelBank fund expiration varies depending on how the funds were received and confirmed the chatbot gave incorrect information. The airline ultimately provided Gil with a $200 electronic travel certificate. "You guys answered me so quickly. I feel like you took me seriously. I mean you guys seem really willing to help and I appreciate that," Gil said.
The accuracy problem in context
While no studies specifically measure the accuracy of AI customer service chatbots, at least two studies of major AI chatbots found roughly half of responses contained errors, misleading information, or missing context. That statistic should concern any support leader whose team relies on deflection metrics as a primary success measure. A deflected ticket that delivers wrong information does not reduce workload - it creates a second, angrier contact later.
Consumer advocates recommend that when a chatbot is the only customer service option, users should save screenshots of the conversation. If the information turns out to be wrong, they will have proof. For supervisors and workforce planners building AI Learning Path for Call Center Supervisors programs, the same principle applies to internal monitoring: if you are not capturing and auditing chatbot transcripts, you are flying blind on accuracy.
Why this matters for customer support professionals
Every inaccurate chatbot response eventually lands on a human agent's desk. The customer is already frustrated, the clock is ticking, and the agent must now fix a problem the company's own tool created. Support leaders should build workflows that flag bot conversations containing policy-related answers for random audit, not just sentiment analysis. A disclaimer buried in fine print does not reduce the operational cost of an escalation - it just shifts the burden to the team least equipped to push back. If your chatbot handles money, dates, or cancellation terms, spot-check its answers weekly. The alternative is paying twice: once in automation savings, and again in service recovery.
Your membership also unlocks:
