Buyers describe the same bot failures to us in almost every first meeting: it repeats the offer after the customer said no, it starts talking before the customer has finished, it loops back to a question it already asked, and the call runs five minutes against a three-minute target. Some are model problems and are getting better; some are design problems that were always fixable. The last kind is the customer's problem being harder than any script, and that one is what a person is for.
The list, as buyers say it
I keep a note of the complaints people bring into a demo about bots they have already tried, in their words. "It kept repeating the offer after I said no." "It started answering before I'd finished." "It asked me the same question twice." "OBD call jaisa laga." "Five minutes twenty for a service request, our AHT is three." "It sounded like it was reading." One head of operations said a customer on a recording had simply said "unintelligent bot" and hung up, and he played it to us, which I respected.
Model problems, getting better
Talking over the customer is mostly turn detection: the model decides the customer has finished when they have paused to think. Sounding like it's reading is mostly the voice. Both have improved a great deal in a year and both still fail on a bad line with a slow speaker in a noisy room, which is what a lot of Indian calls are. We measure per-turn latency on every call, and we don't claim it never happens.
Design problems, always fixable
Repeating the offer after a no is a script problem. Someone wrote the flow as a funnel and the funnel does not have an exit for no. Asking the same question twice is a state problem: the bot did not record the first answer in a place the second prompt reads. The five-minute call is usually both, plus a bot that confirms every field back to the customer because nobody trusted the recognition. What fixes them is someone reading the first week's transcripts with the client and changing the flow. That is the reason we insist on a named owner for that week.
The third kind
The third kind looks like a loop but isn't. The customer has a problem the script does not contain: the policy was mis-sold, the EMI was debited twice, the father died and the car is his. The bot circles because there is nowhere in its world for this to go, and every circle makes the customer angrier. A better model circles more politely, and the fix is a person, on the same call, with the transcript, inside a minute. That's why we put people on the queue rather than selling the bot alone, and why the plateau sits at 60 to 80% rather than 95.
What to do with the list
If you have a bot running now, pull twenty recordings where the customer hung up and sort them into the three kinds. Model, design, third kind. In my experience the third kind is a bigger share than anyone expects and the design kind is bigger than the vendor admits. Then ask whoever runs the bot which of the three they will fix, and by when. If the answer to the third kind is a better model, that is the moment to ask the handover question. Send me your twenty if you want a second pair of eyes.