In this article:
- What changes between setup and advanced configuration
- How do you configure when the chatbot hands off to a human?
- How do you restrict what the chatbot is allowed to say?
- How do you keep answers consistent for specific questions?
- How do you configure multilingual behavior?
- How do you keep the chatbot learning after launch?
- Which settings to review, and how often
Our guide to setting up an AI customer support bot covers the six steps that get LiveAgent’s chatbot live: create the Flow, connect knowledge sources, apply it to the widget. That’s the minimum to launch. It’s not the configuration that determines whether the bot is actually good three months in. If you’re still comparing chatbot options rather than tuning one you already run, our roundup of the best AI chatbots for 2026 is the better starting point.
This guide picks up from there. Every setting below is a real field or component you’ll find in LiveAgent or FlowHunt , not a general best-practice list.
What changes between setup and advanced configuration
Setup answers “does the bot work?” Configuration answers “does it work the way our team actually needs it to?” That difference shows up in four places: when the bot escalates, what it’s allowed to say, how consistent its answers stay across sessions, and whether it gets better or stays static after launch.
A bot that passed its first-week test can still frustrate customers a month later if none of these four are revisited. The knowledge sources you connected at launch age, the questions customers ask shift, and the default handover threshold that felt right on day one might already be sending too many, or too few, conversations to your agents.
How do you configure when the chatbot hands off to a human?
Open the Chatbot object in LiveAgent and set the “Required answers to initialize chat with human agent” field. This number is the minimum count of AI responses given before the “Request chat with human agent” button becomes available to the visitor. Set it too low, and the bot escalates questions it could have handled, wasting agent time on requests that never needed a person. Set it too high, and frustrated customers cycle through unhelpful replies before reaching anyone.
There’s a second path that doesn’t depend on this count at all: the AI Agent can decide on its own, mid-conversation, that a query is too complex and transfer automatically, without waiting for the required-answers threshold. This is the option to lean on for support topics where a wrong answer is costly (billing disputes, account security) and a fixed count feels too rigid.
Once a conversation transfers, agents pick it up from the Chats overview section using the “Stop chatbot and join” button, and the full conversation history carries over into the same ticket, the mechanics covered in detail in our AI human handover feature guide . Nothing about that handoff is customer-configurable, but knowing it happens this way explains why testing the transfer flow end to end matters, not just the bot’s answers up to that point.
Ready to grow your business?
Start your free trial today and see results within days.
How do you restrict what the chatbot is allowed to say?
By default, LiveAgent’s template AI Agent allows a visitor to ask for a human agent directly, at any point in the conversation, separately from the required-answers threshold. If that’s not what you want, for example, a bot meant to fully resolve a narrow set of questions before any human option appears, the fix is editing the system message of the AI Agent component in FlowHunt, not a LiveAgent-side setting.
The system message is also where tone and persona actually live. There’s no separate “personality” dropdown: whatever instructions you write here (formal vs. casual, how to handle a question outside scope, whether to apologize before escalating) is what the bot follows. FlowHunt supports over 15 AI models, and a model switch can shift how literally the bot sticks to that wording, but the model itself isn’t where tone comes from.
How do you keep answers consistent for specific questions?
Knowledge sources answer general questions well, but for anything your team wants worded identically every time, pricing statements, policy language, a specific disclaimer, use FlowHunt’s Questions & Answers feature instead. It lets you define a fixed answer for a specific query, so the bot returns that exact wording rather than generating a fresh paraphrase each time someone asks.
This matters more than it sounds. A knowledge base written for human browsing, spread across five linked help articles, doesn’t automatically produce one clean, consistent bot answer when the AI pulls from all five at once. Fixed Q&A entries close that gap for the handful of questions where wording precision actually matters to your team.
Join our newsletter
The latest tips and deals delivered straight to your inbox.
How do you configure multilingual behavior?
The chatbot automatically detects a visitor’s language and replies in it, with support for over 100 languages depending on the AI model connected. This isn’t a setting to turn on, it works by default, but it’s not uniform across every language. Some limitations can apply to less widely spoken languages, since model quality for those languages varies by provider.
Before rolling the bot out to a market with a less common language, test it directly with the actual phrasing your customers use, not a direct translation of your English test questions. A literal translation can pass while a native phrasing the model hasn’t seen as often produces a noticeably weaker answer.
How do you keep the chatbot learning after launch?
Set up LiveAgent’s self-learning loop
so that every escalated ticket teaches the bot something instead of just being a one-off fix. Create an automation rule in LiveAgent that fires when a ticket is resolved and tagged with Update_AI, sending the ticket ID to FlowHunt via an HTTP request or webhook.
From there, the AI agent runs four steps automatically: it reads the full ticket thread to see what the chatbot missed, compares that to the human agent’s actual solution, generalizes the fix into a reusable rule rather than a one-time answer, and saves it as a structured memory entry (topic, trigger, prerequisites, resolution) the chatbot can retrieve on the next similar question. Personally identifiable information gets stripped out automatically before anything is saved.
The result compounds. Escalations your team resolves in week one become knowledge the bot has in week two, without anyone manually rewriting a knowledge source article.
| Setting | Where you configure it | What it controls |
|---|---|---|
| Required answers to initialize chat with human agent | Chatbot object, LiveAgent | Minimum AI responses before handover becomes available |
| AI Agent system message | AI Agent component, FlowHunt | Tone, persona, and what the bot is/isn’t allowed to say |
| Questions & Answers | FlowHunt | Fixed, word-for-word answers for specific queries |
| AI model | AI Agent component, FlowHunt | Response style and language coverage |
Update_AI automation rule | Automation rules, LiveAgent | Triggers the self-learning loop on resolved tickets |
Which settings to review, and how often
Check the handover rate and the specific questions triggering escalation weekly for the first month after any configuration change, then monthly once the pattern stabilizes. A climbing handover rate usually means a knowledge gap the self-learning loop hasn’t closed yet, not a setting that needs rebuilding from scratch.
If the bot’s tone drifts from what the system message specifies, that’s a sign to revisit the wording directly rather than switching AI models to compensate. And if escalations cluster around one recurring topic, that’s the exact signal the Update_AI tag is designed to catch, so confirm your agents are actually applying it before assuming the self-learning loop isn’t working.
Once handover timing, the system message, multilingual coverage, and the self-learning rule are all set deliberately instead of left on defaults, the difference between “the bot works” and “the bot works the way our team needs it to” mostly closes on its own.
