Why most website chatbots fail — and the four things the good ones do
Most chatbots exist to deflect tickets, and customers can tell. The four things that make the difference.
There is a reason people reflexively close chat widgets. For a decade, most of them existed to stop customers reaching support, not to help them. The technology has changed considerably; a lot of the design has not. Four decisions account for most of the gap between a bot people use and one they dismiss.
1. It knows your actual business
A bot answering from a generic model with no grounding will be confidently wrong about your pricing, your delivery timelines and your return policy — the three things people ask about most. It needs a maintained knowledge base drawn from your real pages and policies, and it needs to say 'I do not know, let me get someone' when the answer is not in there. That sentence is a feature, not a failure.
2. The handover is one click away, always
- A visible option to reach a human in every state of the conversation.
- Full transcript passed to the agent, so the customer never repeats themselves.
- Honest handling of out-of-hours — a form and a stated response time beats a loop.
- Automatic escalation on frustration signals, repeated questions, or anything touching billing disputes.
3. It can do something, not just say something
The step change is not better answers, it is actions. Check an order status against the real system. Book a slot on a real calendar. Raise a ticket and return the number. Start a return. A bot that can only talk is a search box with a personality; a bot that can act removes work.
4. It is measured on resolution, not containment
Containment rate — the share of chats that never reach a human — is the metric that produces bad bots, because the easiest way to improve it is to make escaping harder. Measure resolution rate, satisfaction after bot-only conversations, and the rate at which handed-over chats are resolved on first contact. Those three cannot be gamed by trapping people.
The launch sequence that works
- Start with the ten questions that make up most of your inbound volume. Answer those extremely well.
- Run it in suggest-only mode to your agents for two weeks and correct it before customers see it.
- Turn it on for a single page or segment first.
- Review the transcripts weekly for the first month. The gaps in your knowledge base are all in there.
The services behind this post.
Frequently asked questions.
Usually it changes what the team handles rather than how many people you need — routine status questions go away and the harder conversations get more time.
Less than most teams assume. Ten well-written answers covering your highest-volume questions is a viable launch; the transcripts tell you what to add next.
Talk to the team that does this every day.
Tell us what you're trying to grow. You'll get a plan, not a pitch.
Contact Us