# The AI assistant: answers, hand-over, copilot and safety Source: https://frosvo.com/docs/ai-assistant Updated: 2026-10-22 The assistant answers visitors from your own knowledge, hands over when it should, helps your team with suggested replies and translation, and is protected by masking and safety checks. The assistant answers visitors from **your** knowledge: the help articles you write and the pages or documents you add under **AI Agent**. It shows the source it used, and it doesn't make things up from elsewhere. ![The AI Agent screen where you add the knowledge the assistant answers from](/docs-img/ai-agent.webp) ## Teaching it - **Articles** — write them in the knowledge base. Published articles also appear in your public help centre. - **Pages** — add your website's pages; Frosvo reads them and keeps the text. Give the assistant a name and instructions for each website (for example the tone to use, or what never to promise). ## Hand-over It hands the conversation to your team when the visitor asks for a person, when it isn't sure, or when your workflows say so — and records why. Plans include a monthly allowance of AI conversations; past it, chats go straight to your team. ## For your team - **Suggested replies**, **rewrite** (shorter, friendlier, more formal) and **summaries** in the composer. - **LiveTranslate** — visitors write in their language; your team reads and replies in yours. ## Getting better Reports show what the assistant resolved, what it cost, and the **knowledge gaps** — questions it couldn't answer, grouped so you know which article to write next. You can review its answers and mark them good or bad. ## Safety and privacy - **Sensitive data is masked.** Card numbers, bank account numbers (IBAN), US social security numbers, passwords, PINs and API keys are replaced with placeholders before the AI sees a conversation. The same applies to the copilot and to the assistant's logs. - **Attempts to trick it are caught.** Messages that try to change its rules ("ignore your instructions…") are spotted and the AI is warned. Knowledge pages that contain instructions like that are left out. - **Every answer is checked before it's sent.** An answer that would reveal its instructions, or that links to a website you never gave it, is held back and the chat goes to your team. In reports, the reason shows as "Answer held back by the safety check". - **When it answers.** In Settings → Widget, choose **Only when the knowledge base clearly covers it** for each website if you want the assistant to stay strictly on your content. Anything else goes to a person, and it costs nothing. - **Voice messages.** When a visitor sends a voice message, the assistant and your team get it written out under the audio, and the assistant answers what was said. Voice replies from your team are written out too, for visitors who can't listen. Transcripts count towards your AI usage ("Voice transcripts"). - **Telling visitors it's AI.** By default the widget labels the assistant's messages "AI assistant" and tells visitors they can ask for a person. Some laws require this, including the EU AI Act. You can turn it off per website. - **Safety check.** Owners can run a set of known attacks against a website's assistant under Settings → Websites, and see how each one was handled. ## Search index and models **AI Agent → Quality → Search index** shows which embedding model your knowledge is indexed with and which model writes the answers. Each answer in the review list records the model that gave it. If the server moves to a new embedding model, sources indexed with the old one are marked outdated. The assistant still finds them by their words, but not by meaning. Owners choose **Re-index outdated** or **Re-index everything**. Sources are re-indexed a few at a time in the background, and the work counts towards your AI budget (it pauses while the budget is used up). You can stop it at any time. ## Keep it in check Run the [safety check](/docs/assistant-safety-check) after you change instructions, set an [AI budget](/docs/ai-budget-and-usage), and read the knowledge gaps in [reports](/docs/reports) every week. Write the missing articles in your [knowledge base](/docs/knowledge-base). ## Common questions ### Does the assistant make things up? It answers from your knowledge and shows the source it used. When the knowledge does not cover a question, it hands over. ### Which plans include the assistant? Not the Free plan. Plans include a monthly allowance of AI conversations. ### Can I make it stay strictly on my content? Yes. In Settings, then Widget, choose Only when the knowledge base clearly covers it.