I Told an AI Support Bot Twice I Had Already Fixed Its Mistake. It Still Could Not Process That.

I booked a trip to Dallas through Navan, and needed to change the departure airport after the booking was already made. Navan would not allow the change through its own system, so I went directly to Delta and made the change myself. A short time later, Navan contacted me: fix this through official channels, or they would cancel the itinerary entirely.

I contacted support. What answered was an AI chat. I explained plainly that I had already made the change and wanted to keep it. I said it a second time, just as plainly, in case the first version had been unclear. The bot could not process it. It could not confirm the fix already existed. It could not let the thread go. It kept running its own process, threatening cancellation, as if a fact I had stated twice in plain language simply had not been said. Eventually it escalated me to a human, who resolved it in a few minutes.

This was not an isolated incident. I had separately given Navan direct feedback that it would not show me flights out of New York, forcing me to book with Delta directly again. No acknowledgment that the feedback was received. No indication it was being worked on. No follow-up, ever. Two different tools, two different narrow gaps, the same missing piece both times: nothing in either system was built to recognize its own failure and close the loop.

Every technical leader watching AI tools improve month over month has felt the pull toward trusting them with more, not just internal operations, but actual customer-facing interactions end to end. The rough edges are assumed to be smoothing out version by version, and whatever gaps remain get treated as minor, temporary friction, not something worth actively hunting for. That assumption gets more dangerous the closer these tools sit to the customer, because customer service is the kind of thing a company can be built entirely on, and the kind of thing that crashes a company that gets it wrong.

Here is what actually happens. A company deploys AI into customer service because it handles the visible majority of cases fluently, and fluency gets mistaken for readiness. What never gets tested is the narrow edge case, a customer correcting the system's own mistake, a request that falls just outside the trained pattern, because those cases are rare enough that nobody built a test for them and common enough that every real customer base eventually produces one. When a customer hits that gap, the system has no mechanism for recognizing it has failed, so it keeps running its normal process against an abnormal situation, threatening consequences, repeating itself, escalating friction instead of resolving it. The customer's actual experience in that moment reads as a company that does not have anyone paying attention, not confusion about a machine, because from the outside, there is no way to tell a company that has not tested for this gap from a company that does not care that the gap exists.

The cost lands in two places at once: the relationship takes the hit immediately, and the customer often leaves anyway, because trust and retention are not actually separate line items when the failure happens in the one function that was supposed to prove the company was paying attention.

Waiting for the AI to get better at every edge case does not work, that list never actually closes. What works is building an explicit mechanism for the moment the system cannot recognize its own failure: a detection point that flags when a customer's plain statement does not match any expected pattern, and an escalation path that fires automatically, before the customer has to force it by repeating themselves or making a threat stick. The company that has this in place turns the narrow gap into a fast recovery. The company that does not turns it into the exact moment the customer decides nobody is paying attention, and starts looking for a company that is.

Think about the last time a customer had to repeat themselves to one of your AI-driven systems, in support, sales, or onboarding. Was there an automatic point where the system recognized it was stuck and escalated, or did the customer have to force that themselves, the way I had to?

I send a short daily email on exactly this kind of gap, the narrow places AI-driven systems fail that nobody built a test for. Subscribe free: https://technicalleader.coach/daily-email

---

I write about structural leadership for technical leaders in high-stakes operating environments. If you're reading this outside the daily email, subscribe free: https://technicalleader.coach/daily-email