I recently spent far too much time talking to multiple AI phone support agents without getting help for the problem I actually called about. What finally got me to a human? I had to tell the AI I had a different problem.
The experience was a perfect example of how AI can make a business more efficient when implemented well and make the customer experience considerably worse when it is not. AI itself was not really the problem. The way it had been designed and deployed was.
Here is what happened, along with a few lessons I think any business considering AI for customer support should take from it.
The problem
I needed to change the bank account associated with my Apple Card. I entered the new account and routing numbers in the app, but they were rejected. I checked the information, tried again, and got the same result. At that point, I called support and entered what became an AI infinity loop.
The automated system gave me three choices. Apple Card was one of them, so I selected it. Then the AI asked me to describe my problem.
“I’m getting an error when I try to add a new bank account to my card.”
The AI responded that it understood I needed help adding a bank account and offered to walk me through the process.
“No, I know how to add it. I’m getting an error.”
It walked me through the process anyway. I followed along, entered the information again, received the same error, and told the AI it still did not work. After a few seconds of silence, it said it would transfer me to a bank specialist.
Excellent. Or so I thought.
Lesson #1: Recognize when the answer did not work
The AI understood part of what I was saying. It heard “add” and “bank account” and correctly matched those words to the procedure for adding an account. The problem was that it apparently had nowhere else to go when that procedure failed. Statements such as “I already did that,” “I’m getting an error,” and “that didn’t work” should change the conversation.
A good support system needs more than a library of answers. It needs to recognize when the first answer failed and know what to do next. Repeating the same instructions is not support. It is a loop.
Round 2: Different voice, same AI
A few seconds later, another voice answered.
“Please describe the issue you’re having.”
I explained the problem again, and once again the AI started explaining how to add a bank account.
At this point I started wondering whether there was a magic word that would get me to a human. Many phone systems recognize words such as “representative,” “agent,” or “operator,” so I tried:
“Representative.”
Success. The system told me it would transfer me to an agent. Except it didn’t.
Lesson #2: Always give customers a way out
AI should absolutely handle routine requests. That is where it can save time for both the customer and the business, but the customer should never feel trapped inside the automation.
Give callers an obvious way to reach a human and tell them about it. Something as simple as, “If you would prefer to speak with a representative, say ‘agent’ at any time,” would solve much of the frustration. You can even provide the estimated wait time. If someone would rather wait twenty minutes for a person than spend another five minutes talking to AI, let them make that decision.
People usually call support because something has already gone wrong. Your goal should be to reduce their frustration, not add another layer to it. Unfortunately, saying “representative” had not actually gotten me out.
Round three: Another AI Agent
A third AI voice answered.
“Before I can transfer you to a specialist, please describe the nature of your call in a few words.”
I described the problem. It started explaining how to add a bank account. Again.
I hung up.
Lesson #3: Do not make customers start over after a transfer
Having to explain the same problem repeatedly made the experience even worse. If your system transfers a customer, the information already collected should travel with them.
The next AI agent, or preferably the human agent, should already know who the customer is, why they called, what troubleshooting has been attempted, what happened, and why the call was escalated. Customers should not have to retell the same story every time your system moves them somewhere else. That is not really an AI problem. It is a workflow design problem.
Lesson #4: Give the AI an “other” path
There was another problem underneath all of this. The system seemed to decide that because I used the words “add” and “bank account,” I must not know how to add one. But that was not my problem. I knew how to add the account. The system was rejecting it. Those are two different support requests.
Not every customer issue will fit neatly into the categories you anticipated when the system was created. There needs to be a catch-all path for unusual situations, and the AI needs to recognize when the caller’s problem does not match the standard answer. Otherwise, it simply forces the customer into the closest category available, even when that category is wrong.
AI is only as good as the system around it
I do not blame the AI. It was doing what it had apparently been designed to do: recognize certain words, match them to a procedure, and deliver that procedure. What it could not do was recognize that the procedure had failed and choose another course of action.
That is where businesses need to be careful. Installing an AI support agent does not automatically improve customer service. You need to think through what happens when the AI misunderstands the problem, the standard solution fails, the customer asks for a human, or the issue does not fit an existing category. Those are not unusual exceptions. They are part of customer support.
The real test of customer-facing AI
There is a simple question businesses should ask when evaluating an AI support system:
Is this making it easier or harder for the customer to get help?
Metrics such as how many calls AI can handle and how much it can reduce support costs certainly matter, but they should not be the only measures of success. If AI handles most of your routine calls but causes frustrated customers to give up trying to reach you, you may be optimizing the wrong metric. Customer-facing AI should remove friction. It should resolve easy problems quickly, recognize when a problem is no longer easy, and gracefully get out of the way when a human needs to take over.
So, did I ever get my problem resolved?
Yes. I called back and, instead of telling the AI that I could not add a bank account, I said I was having trouble making a payment. I was transferred to a human almost immediately. The fastest way to get help from the support system was to describe a problem I did not actually have.
If your customers have to figure out how to outsmart your AI just to reach you, your AI implementation probably needs some work.
