A chat window is the cheapest thing you can add to a website and the easiest one to get wrong. Installed badly, it is a pop-up that interrupts people, collects an email address, and promises a reply that arrives on Monday. Installed well, it answers the question that was actually blocking a sale, at the moment it was blocking it.
The difference is not the widget. It is what happens in the ninety seconds after someone types.
Most questions are the same six questions
Before deciding anything about technology, read your last two hundred customer emails. Nearly every business discovers the same thing: the overwhelming majority cluster into a handful of questions.
Do you deliver to my country. What is the lead time. Is this compatible with what I already have. What does it cost for my quantity. Where is my order. Can I get an invoice with my company details.
These are not hard questions. They are answered identically every time, and the answers already exist—on a shipping page, in an ERP, in a policy document, in the head of whoever has been doing this for six years. The cost is not difficulty. It is repetition, and the delay between a customer asking and a human being free to answer.
That delay is the expensive part. A visitor with a question about lead time is a visitor deciding whether to buy. An answer in ten seconds keeps the decision alive. An answer tomorrow morning arrives after they have already checked a competitor.
What AI should and should not do here
The useful framing is not "can AI replace support." It is "which of these six questions can be answered from material we already have, with no judgement required."
An AI that has been given your actual documentation—shipping terms, product specs, policies, previous answers—handles those confidently and consistently. It does not get tired at 11pm, it does not answer differently on a Friday, and it does not forget the exception for orders above a certain value.
What it should not do is improvise. The failure mode everyone has experienced as a customer is a chatbot that answers fluently and wrongly, because it would rather produce a sentence than admit a gap. That single behaviour destroys more trust than the chat window ever earned. A support AI worth installing is one that is explicitly allowed to say "I don't know, let me get someone"—and then actually does.
The handoff is the whole product
Every chat conversation ends in one of three ways: resolved, abandoned, or handed to a person. The third is the one that decides whether the tool helped or hurt.
A bad handoff restarts the conversation. The customer explained the problem, gets told a human will be in touch, and then a human arrives and asks what the problem is. The customer has now told their story twice and is annoyed—worse off than if there had been no chat at all.
A good handoff carries everything forward. When a person takes over, they see the full transcript, what the AI already answered, what it declined to answer, and who the visitor is if they identified themselves. The customer notices a change of tone, not a restart. In the best case they do not notice at all.
Three things make that work in practice:
- Clear escalation triggers. Explicit request for a human, repeated failure to resolve, detected frustration, or any topic touching money, contracts or complaints. Those should never be handled automatically, regardless of how confident the model sounds.
- A queue humans can actually see. If escalated conversations land in an inbox nobody watches, the escalation is theatre. The handoff needs to be visible and, when nobody is available, honest about it.
- Truth about availability. "Someone will reply within two hours during business hours" is a good message. A blinking green dot at 2am is a lie, and customers learn quickly.
Where the chat window belongs
Not on every page, and not two seconds after arrival. A chat that interrupts someone reading is an obstacle. A chat available where decisions get stuck—pricing, product specs, checkout, delivery terms—is a tool.
The practical version: present but quiet everywhere, prominent on the pages where your analytics already show people leaving. If a page has a high exit rate and a question attached to it, that is where the widget earns its place.
How to know whether it is working
Conversation count is a vanity metric. Three numbers actually matter:
Resolution rate without a human. Of all conversations, how many ended with the customer's question answered and no escalation. This is the number that pays for the tool.
Time to first human response on the ones that did escalate. If this is measured in hours, the AI is not saving your team—it is deferring work and giving customers a worse experience while doing it.
What it could not answer. The most valuable output of a support chat is not the answers. It is the log of questions it failed on. That list is a direct instruction for what to document next, and it is a list most companies have never had.
Start narrow
Do not launch a chat that claims to handle everything. Launch one that handles your six questions extremely well, refuses everything else politely, and hands off cleanly. Customers forgive a system that knows its limits. They do not forgive one that confidently gives wrong information.
This is how we built Andivio AI Support Chat: the AI answers from your own material and resolves what it can, and the moment it cannot—or the moment someone asks for a person—it hands the full conversation to a human who picks up mid-thread rather than starting over. The escalation is not a fallback bolted on afterwards. It is the part we designed first.
Curious what your six questions are?
Send us a sample of your recent customer emails and we'll tell you honestly how much of it a chat could resolve—and which parts genuinely need a person.
Tell us what repeats→Common questions
Should an AI chat replace human support?
No. It should absorb the handful of questions that repeat and are answerable from existing documentation, and hand everything else to a person. Anything touching money, contracts or complaints should escalate regardless of how confident the model sounds.
What makes a handoff to a human good or bad?
A bad handoff restarts the conversation and makes the customer explain the problem twice. A good one carries the full transcript, what the AI already answered and what it declined to answer, so the person picks up mid-thread.
How do I measure whether a support chat is working?
Three numbers: how many conversations resolved without a human, time to first human response on the ones that escalated, and the log of questions the AI could not answer — which tells you exactly what to document next.