We ship an AI chatbot and we think it's good. We also think a lot of teams switch one on in places it actively makes support worse - and then conclude the technology is useless, when really it was pointed at the wrong problem. Here's where we'd tell you not to bother.
Somebody whose order is late, whose payment failed twice, or whose service is down does not want to be triaged. They want a person to acknowledge the problem. A bot in that moment reads as a company avoiding them - and it is remembered long after the issue is fixed. If your widget sits on a page people reach when something has gone wrong, route it straight to a human.
Medical, legal, financial and safety questions are the wrong place for a model that can be confidently wrong. The cost of a plausible incorrect answer is not a bad review - it's harm, and sometimes liability. Keep AI away from advice a regulator would care about, no matter how good the documentation is.
Grounded AI answers well from what you've written down. During an outage, a recall, or the week after a confusing release, the questions arriving are precisely the ones nobody has documented yet. That's when a bot is least useful and most likely to improvise - and exactly when your customers are least forgiving.
A chatbot is not a set-and-forget appliance. It needs documentation written, answers reviewed, handoff tuned. At a handful of chats a day a person can simply answer all of them, better, with less effort than maintaining the bot. Canned responses and a proactive greeting will get you most of the speed for none of the upkeep.
If the honest goal is to make reaching a human hard, a chatbot will achieve that - and customers recognise it immediately. Deflection that comes from genuinely answering the question is good for everyone; deflection that comes from friction just moves the anger to a public review.
The mirror image of the list above.
The single biggest difference between a chatbot people tolerate and one they resent is how quickly it steps aside. Set generous handoff triggers, let "talk to a human" work at any point, and pass the full transcript across so nobody repeats themselves. A bot that admits defeat early is judged far more kindly than one that keeps guessing.
Read next
How handoff works - and how to set the triggers.
The quieter option
Reply suggestions give your team the speed without putting a bot in front of customers.
ZChat gives you the installable server, web dashboard, website widget, and desktop agent tools in one self-hosted product you buy once and keep. Run it on infrastructure you trust and connect AI only if and how you want it.
Deployment
Install on Windows or Linux, behind IIS or Nginx, in a VM, or in Docker if that fits your stack.
Commercial model
One-time purchase, perpetual license, and no monthly per-agent bill attached to growth.
AI Flexibility
Use Ollama locally or connect OpenAI and Anthropic with your own provider accounts.