When a chatbot actually helps, and when it just makes people angrier

I was testing a chatbot for a client last Tuesday, typing in the kind of questions a real customer would actually ask. Third question in, it looped back to 'Sorry, I didn't quite understand that, would you like to speak to someone?' Which is a fine answer, except it had already said that twice.

We'd built better ones before. So I sat there for a bit trying to work out what was actually different this time, because it wasn't the AI model. It was practically the same one we use elsewhere.

The difference was what the chatbot was allowed to know. This one had a script and a FAQ document and nothing else. It couldn't check an order, look up stock, or tell someone their package left the warehouse on Monday. So every question that mattered got the same shrug, dressed up in polite AI language.

The good ones aren't smarter, they're just wired in

We built a chatbot for a client who sells parts for HVAC systems, the kind of business where a technician is standing next to a broken unit at ten at night trying to find a part number. That chatbot is connected to their actual product database. It doesn't guess. It checks. If someone types a model number, it pulls the real specs and the real stock count, and if it isn't sure, it says so and offers to email a person in the morning instead of inventing an answer that sounds confident and is wrong.

That's the whole trick, really. A chatbot that's honest about the edges of what it knows feels helpful even when it can't solve your problem, because it doesn't waste your time pretending. A chatbot that's been told to sound helpful no matter what just keeps generating plausible sentences until you give up and start typing in capital letters.

People don't get annoyed because a chatbot can't help them. They get annoyed because it acts like it's helping when it isn't, and it takes three exchanges to find that out instead of one.

The other thing I noticed testing that first one: there was no obvious way out. No visible 'talk to a person' option, no email address in the chat window, nothing. You had to ask for a human, and then hope it understood you were asking. A good handoff is almost more important than a good answer. If the bot genuinely doesn't know, the fastest route to a real person is the difference between a customer who's mildly inconvenienced and one who's furious.

None of this is really about how advanced the underlying AI is anymore. Most of the models are good enough now. It's about whether someone bothered to connect it to real data, tell it what it's allowed to say when it doesn't know something, and build a door out of the conversation that actually opens.

The client with the FAQ-only bot wasn't wrong to build one. They just built the front door without connecting it to the house.