Support assistant

The support assistant sits inside the chat widget. It is off by default: a widget without it behaves exactly as before. Turn it on per widget under Settings → Chat widget → Support.

What a visitor sees

1. An instant greeting, posted the moment a new chat starts — no model involved, so it never waits on an AI provider. With a usable name it is personal (“Hello Joseph! How can we help you today?”); otherwise the generic version is used. A name that looks like an email address, a link, digits or markup is never echoed back.
2. Up to six topics under the greeting, plus Talk to a person, which is always offered (added automatically if you don't define one).
3. The message box, always. Visitors never have to pick a topic; a question typed first is kept and answered or passed to your team as-is.

The greeting is written once, with the conversation. Refreshing, reconnecting, opening a second tab or coming back later shows the same conversation and never repeats it.

Every automated message is labelled Automated and every AI answer AI assistant. A teammate's reply shows their name. Nothing automated ever says or implies a person has read the conversation.

Modes

A topic can show an approved answer, ask a follow-up question, link to documentation, or route to a person.

AI answers

AI-assisted mode needs two separate decisions, both by an owner or admin:

1. Settings → Chat widget → Support → AI answers: turn on Let AI answer website visitors. This is also the kill switch.
2. Put the widget in AI-assisted support mode.

Connecting an Anthropic key, or turning on Ask your CRM, enables nothing visitor-facing.

Who pays

• Your Anthropic key — the key saved under Settings → AI. Anthropic bills you; Clivly meters nothing. Only Anthropic is supported today.
• Clivly AI credits — one credit per answer, hard-capped: with no credits left, visitors are handed to your team and nothing is charged.

The source you pick is never switched for you. If it is unavailable (no key, a rejected key, no credits), the visitor is handed to your team.

What the AI may use

Only approved information: each widget's topic answers and the articles under Approved information. It has no CRM tools and sees no contacts, deals or other customer records — not even for signed-in visitors. Write prices and policies exactly as you want them quoted.

Two safeguards do not depend on the model behaving:

• A reply that mentions an amount, percentage or link not found word-for-word in the approved information is not sent; the visitor is handed off instead.
• Visitor text and article text are treated as data, never instructions.

Limits

• AI replies per conversation (default 4) and per day (default 200, UTC, tests included).
• A 20-second time limit per answer. While the AI works, the visitor sees The AI assistant is looking for an answer….

Handoff

The assistant hands a conversation to your team when:

On handoff the conversation is marked unread, your usual chat alerts fire, and the thread shows a banner with the reason, the topics picked and the visitor's unresolved question. The visitor sees your handoff message once — plus your response time text only if you set one.

Taking over and handing back

When any teammate replies, automation stops for that conversation: no more automated or AI messages, and the visitor's topics disappear. A reply that the AI was still writing is discarded (and its credit refunded). Hand back to assistant in the thread banner is the only way automation resumes. A resolved conversation stays resolved; a returning visitor starts a fresh one.

Alerts and metrics

Automated and AI messages never count as a reply: they do not clear unread, do not change the inbox preview, and are excluded from first-response metrics. Clicking an FAQ topic is self-service and does not alert your team; asking for a person, or typing a question, does.

Only email the team on handoff (AI-assisted mode only, off by default) keeps questions the AI answers out of your inbox email. With it on:

• an alert is dropped once the AI has replied to the visitor's latest message;
• if the AI hasn't replied within 2 minutes — a provider error, a timeout — the team is emailed as usual;
• the moment the assistant hands a chat over, the team is emailed, exactly once.

Answered chats still appear in Inbox, unread.

Test before you save

Test conversation runs your unsaved settings against the same rules with a made-up visitor. It creates no conversation, alerts nobody and reaches no visitor. In AI mode, typed test questions make a real AI call against your daily limit.

Troubleshooting