research - 2026-07-11

When an AI Assistant Blocks the Customer, the Workflow Is Broken

A founder-led take on a bank support flow where an AI assistant kept repeating the app path and delayed a simple credit limit request until the customer got angry.

Today I ran into a very ordinary problem and got a very expensive kind of bad automation back. I wanted to raise my credit limit so I could pay a car rental deposit. I do not really use credit cards, and the app was already throwing an error while I tried to upload the documents.

The bank chat kept telling me the same thing: do it in the app. The AI assistant asked if it could help, then asked what my question was so it could find the right specialist, then told me again that I could do it in the app. That is the part people miss. The problem was not that the assistant sounded bad. The problem was that it had no real move once I had already said the app was failing.

I asked for a human. The bot stalled. I tried again in Telegram. Same story. Then I called the hotline and got another assistant, “Anna,” who again pushed me back to the app. Only after a lot of shouting did I reach a person. She repeated the app answer too, then checked the case and finally told me the kicker: I had exhausted my request limit for the day and needed to call back tomorrow. No new application had even been sent.

That is not a model problem. That is a workflow problem. If the system cannot recognize “the app is broken,” cannot hand off cleanly, and cannot explain the actual limitation early, then AI is not helping operations. It is just making a wall look nicer. And people do not get calmer when they are bounced between channels by a cheerful assistant.

This is where a lot of teams fool themselves. They add an AI voice, maybe a chatbot, maybe a phone assistant, and assume the service got better. But the real customer flow is the ugly part: errors, refusals, limits, document failures, and the human who has to step in when the happy path breaks. If that recovery path is weak, the AI is not a layer on top. It is the thing customers remember for the wrong reason.

If you are building support automation, check the exception paths first:

- app error or upload failure - limit reached - customer asking for a human - repeated channel switches - refusal without a clear explanation

Those are the moments that decide whether the assistant helps or loses the customer. Not the demo. Not the voice. The handoff.