Guardrails: Confidence, Limits & Requiring an Email

"Guardrails" are simply safety rules — boundaries that keep your bot behaving well, control costs, and decide when it should stop and hand a visitor to a real person. Think of them like guardrails on a road: they keep things on track. This page covers the Guardrails tab.

Where to find it

Go to Chatbots in the left sidebar and click your bot.
Click Configure.
In the left menu, click Guardrails.

You'll see a note at the top confirming these are safety boundaries — the bot stays inside them and hands off to a human when a limit is reached.

Confidence threshold — when to call in a human

What it is: a slider that sets how sure the bot must be before it answers on its own. Why it matters: it's your quality safety net. If the bot isn't confident, you'd rather a person step in than have it guess wrong. How: drag the slider (shown as a percentage).

Lower % → the bot answers more on its own (more coverage, slightly more risk of a shaky answer).
Higher % → the bot hands off more often (safer, but more work for your team).

When the bot's confidence drops below your setting, it passes the conversation to a human agent. A middle setting is a sensible starting point; adjust after you've watched a few real chats.

Max messages before handoff

What it is: the number of back-and-forth messages allowed before the bot automatically brings in a human. Why it matters: if a visitor is still stuck after many messages, a person should take over. How: type a number in Max Messages Before Handoff. Enter 0 for unlimited (no automatic handoff based on message count).

Daily AI credit limit — cost control

What it is: a cap on how many AI credits this bot can spend per day. Why it matters: it protects you from a surprise bill if the bot suddenly gets very busy (or someone tries to abuse it). How: in Daily AI Credit Limit, enter a number.

Leave it blank → uses the default of 100 credits/day.
Enter 0 → stops this bot from spending any credits.
Each reply costs roughly 1 credit per 1,000 words of text (minimum 1). The bot can never spend more than your account's credit pool actually holds.

Profanity filter — keep it clean

What it is: an on/off switch that blocks and warns on offensive language. Why it matters: keeps conversations professional and safe. How: toggle Profanity Filter on.

Blocked topics — subjects to avoid

What it is: a list of subjects the bot will politely refuse to discuss. Why it matters: keeps the bot on-brand and out of trouble (e.g. competitors, politics). How: in Blocked Topics, type a subject and press Enter. Add as many as you like — for example competitor, politics, adult content. The bot will decline to engage on these.

Keyword triggers — auto-actions on specific words

What it is: rules that fire when a visitor's message contains a word you choose. Why it matters: some words mean "get a human, now" — like refund, cancel, or urgent. How: in the Keyword triggers section:

Type the word (e.g. refund).
Pick what happens: Hand off to agent, Block message, or Custom response (you write the auto-reply).
Click Add Trigger.

Add one trigger per important word. They're listed below so you can remove any anytime.

Require email before chat — read this carefully

This is the one that trips people up, so here's exactly how it works.

What it is: a switch that makes visitors type their email address before the AI will answer anything. Why it matters: it turns every chat into a lead — you capture the email up front. But it also adds a step, so use it intentionally. How: toggle Require Email Before Chat on.

What the visitor experiences when it's ON:

They open the chat and are asked for their email before they can talk to the bot.
They must type a normal, valid email address — something like name@company.com.
Once a valid email is entered, the chat unlocks and the AI starts answering as usual.

Important: the visitor just types a plain email like jane@acme.com and hits enter — that's all it takes to unlock the chat. There's no code, no verification link, no password. If a visitor says "the bot won't answer me," the first thing to check is whether Require Email Before Chat is ON and whether they entered a properly formatted email (it needs an @ and a domain like .com). A typo like jane@acme (missing the .com) won't unlock it.

If you don't want visitors gated behind an email, simply leave this toggle off — the AI answers immediately.

How a handoff actually plays out

When a guardrail triggers a handoff (low confidence, a keyword, or the message limit), here's the flow:

Trigger detected — confidence dropped, a keyword matched, or the message limit was hit.
Bot wraps up — it sends a handoff message and marks the chat "Queued for agent."
Agent notified — your team gets an email and/or browser alert (set this up in the Notifications tab).
Agent takes over — a teammate opens the Inbox, clicks Take Over, and replies directly.
Resolution — the agent wraps up; the visitor may see a quick satisfaction rating.

Learn more in Handing Off to a Human Agent.

After you save

Test everything with the Test Bot preview. Especially: if you turned on Require Email Before Chat, open the preview and confirm you're asked for an email first, then type a valid one and confirm the AI responds. See Testing Your Bot Before You Launch.