AI Chatbot
Train it on your website, FAQ, and knowledge base in minutes. It answers visitors 24/7 in your tone of voice, keeps a clear handoff path to humans, and gives your team visibility into the questions customers ask most.
Train
Crawl your website, paste FAQ text, point at help docs — ready in minutes.
Answer
Answers from your approved content, with source links visitors can trust.
Hand off
Escalate to a human with the full context already in their queue.
The first screen
Most chatbots fail because the first screen is a blank input. Ours gets visitors talking with starter prompts, a clear bot identity, and a friendly disclaimer about what AI can and can’t do.
Greeting
"Hi! What can I help with today?"
Sets tone — friendly, focused, not gimmicky.
Starter prompts
3-6 clickable suggestions
"How do I install?" · "Pricing" · "Talk to a human" — trains the visitor on what the bot is good at.
Bot identity
Name, avatar, disclaimer
Set realistic expectations — visitors know they’re talking to AI.
Start screen — greeting, prompt chips, bot identity, disclaimer.
Handoff
When the visitor asks for a human, when confidence drops, or after three unresolved turns — the bot escalates. The agent picks up the same thread with the full transcript already in front of them.
- "Talk to a human" / "agent" keyword detection
- Confidence threshold (configurable)
- N-turn unresolved trigger
- Department routing on handoff
Safety
Every answer is grounded in your training sources. If the bot doesn’t know, it says so — and offers to escalate. No made-up policies, no invented prices, no off-brand advice.
- Answers based on approved training sources
- Source links back to the original pages
- Out-of-scope detection & safe fallback
- Per-message audit log for every answer
Before you switch it on
An AI assistant grounded in your content is only ever as good as that content. If a page states an old price, the bot will state the old price confidently, and confidence is exactly what makes a wrong answer expensive. So the work before launch is content work, not prompt work: crawl what is genuinely current, and leave out anything you would not want quoted back to a customer verbatim.
Then decide, in writing, what the assistant is not for. Most teams keep it away from anything touching an individual account, refunds, or promises about delivery dates, and route those straight to a person instead. Test privately against the questions your customers actually ask before any visitor sees it, and set the handoff triggers deliberately rather than accepting the defaults.
- Write the disclaimer plainly — say it is AI, and say when a human takes over
- Schedule the re-crawl, so a price change reaches the bot without anyone remembering
- Set the confidence threshold and the unresolved-turn limit before launch, not after a complaint
Our guide on setting AI chat scope and guardrails works through the boundary decisions in detail.
Keeping it honest
AI chat is not a project that finishes. The questions visitors ask drift as your product changes, and the assistant only stays accurate if somebody looks at what it said. A short weekly review is enough, and the question-insight list makes it quick: the questions asked most often, and the points where the conversation handed off to a person.
The most valuable output of that review is a content to-do list. A question the assistant could not answer is usually a missing help article rather than a model problem, and fixing the article helps your search traffic and your human agents at the same time. When an answer is wrong, resist rewording the prompt as a first move: correct the source the answer came from, then re-test the same question.
See reviewing AI chat answers and keeping AI answers current for the review routine.