--- title: "AI Assistant (Instance Agent) — overview" description: "How the built-in AI Assistant in ConnectiveOne helps administrators and configuration teams understand the platform, find features in the UI, and make changes safely." --- # AI Assistant (Instance Agent) overview You are configuring a bot, hunting for a toggle in **Settings**, or showing a new teammate where things live — and you do not want to scroll through dozens of documentation pages. The **AI Assistant** is a chat panel in ConnectiveOne: it answers from platform documentation, knows which page you are on, and can guide you with a UI tour or perform actions for you (when your role allows it). This page is **not** a click-by-click manual. It explains the **idea**: who the assistant is for, how it differs from Copilot, how **Ask / Agent / Plan** modes work, what **Skills** are, and who can do what. --- ## Who this is for - **Administrators and root users** — find a feature quickly, change settings in plain language, or follow a UI tour. - **Integrators** — scenarios, Custom Data, broadcasts, roles: answers in the context of the module you have open. - **Supervisors and team leads** — understand a module before training others. - **Operators** — only if an administrator granted access to the **AI Assistant** module; usually the assistant is for configuration teams, not for replying to customers in chat. --- ## In plain terms: what the AI Assistant is Imagine a colleague who knows ConnectiveOne documentation well and sits beside you while you work in the platform. You ask — they answer. You cannot find a button — they point on screen. You need something configured — they do it, if you gave them permission. In the UI this is the **AI Assistant** (in instance settings — **Instance Agent**): - it opens from the **spark** icon in the top bar; - it works **on top of the current page** — you stay in the section you are in; - **customers in channels never see it** — it is a tool for your team inside the platform. The assistant does not replace operators or speak on behalf of your company in customer chats. It helps **those who configure** ConnectiveOne. --- ## Do not mix it up with Copilot and Fast Line Pro ConnectiveOne has several AI tools — they complement each other but work in different places: | | AI Assistant | Copilot | Fast Line Pro | |---|---|---|---| | **For whom** | Configuration team | Operator in a customer dialog | Customer on the first line | | **Where** | Side panel in ConnectiveOne | Operator panel, inside the chat | Widget, messengers, bot scenarios | | **Answers from** | Platform docs + your data | Knowledge base + customer thread | Company knowledge base | | **Who sends text to the customer** | No one | Operator after review | AI agent automatically | **In short:** Fast Line Pro faces customers externally, Copilot helps operators inside a dialog, the AI Assistant helps admins and integrators inside the platform. --- ## Why teams find it useful ConnectiveOne documentation is large — that is normal for a platform like this. The assistant eases typical pain points: - **“Where is this in the UI?”** — an answer in the context of your page, or a UI tour with highlights. - **“I am afraid to break something”** — **Plan** mode shows steps and risk first; changes only after **Approve**. - **“The same manual setup again”** — describe the task in words; the assistant runs steps if you have edit permission. - **“We have one in-house expert”** — documentation knowledge is available to everyone with access. You do not need a separate documentation tab for one parameter — ask in chat without leaving your current section. --- ## How a conversation with the assistant works The logic is straightforward: 1. You open the **AI Assistant** and see the **Context** tag — the assistant knows which module you are in. 2. If needed, you pick **Mode** (**Ask**, **Agent**, or **Plan**). 3. You write in **Ask a question about your platform...** — you can add **`/`** to pick a **skill**, or attach a file. 4. The assistant searches ConnectiveOne documentation (you may see status like **Searching documentation…**). 5. You get text, a UI tour, a plan, or system changes — depending on mode and permissions. During generation you can click **Stop**. Chat history is stored **in your browser** — conversations may not appear on another computer. --- ## Three modes — what they are for Below the input — **Mode**. This is not “difficulty” but **how much you trust automatic actions**: | Mode | Idea | When to pick it | | --- | --- | --- | | **Ask** | Answers and explanations only | You want to understand without touching settings | | **Agent** | Answers + actions in the system | Normal work when role permissions allow it | | **Plan** | Plan first, then execution | Several steps, bulk changes, you want to see “what will happen” | ### Plan mode — your draft checklist In **Plan** mode an **Execution plan** block appears: steps, **Low / Medium / High risk** labels, **Approve**, **Revise**, and **Cancel**. **Nothing changes in the platform until Approve.** Wrong plan — click **Revise** and describe what should be different. --- ## Skills — ready-made “workflows” An **agent skill** is a prepared way to phrase a task for a specific area: settings, clients, bot scenarios, Custom Data, and more. This is **not** operator skill groups in personnel settings. In practice: **`/`** in the input → **Skills** list → pick a skill → add your request → send. Names follow your interface language. Remove a skill — **×** on the chip (**Clear skill**). Skills save time when you already know *what* you need but do not want to write the prompt from scratch. --- ## UI tours and page context **Context** — the tag above the input. Thanks to it, “how do I enable notifications?” in **Settings** vs in **OperatorLine** gives different, relevant answers. A **UI tour** — when you ask *where* to find something, the assistant may offer step-by-step highlights on screen. In chat — **Show me** or **No thanks**. Easier than “third tab bottom left”. --- ## Who can do what: three access levels Access is layered on purpose — root can enable the feature, admins can limit who actually changes data: 1. **Instance** — root enabled **Instance Agent Integration**. Without this, the **spark** button is hidden. 2. **Role** — **Menu → Settings → Users → Roles**, module **AI Assistant (Instance Agent)**: - **Access** — whether the person sees the assistant; - **Allow the agent to create, modify, and delete data (write mode)** — whether it may change data automatically. 3. **Instance read-only** — **Agent Read-Only Mode**: for **everyone**, view and tours only, no automatic changes — even if the role has edit permission. No edit permission in the role? The assistant still answers and can run a tour — but you change settings yourself in the UI. --- ## Web search and public pages (if enabled) By default answers come from **ConnectiveOne documentation** and your platform data. An administrator can optionally allow: - **Instance Agent: Web Search** — search on the public internet; - **Instance Agent: Fetch URL** — read one public page by URL. Both are **off by default** and need an instance toggle **and** a role permission. The assistant does not bypass login or closed site areas. **Fetch URL** returns model-generated text from the page (not raw HTML or a guaranteed verbatim copy). --- ## A few more things worth knowing - **Model** — below the input you can pick one of five models: **Gemini Pro Latest**, **Gemini Flash Latest**, **GPT 5.5**, **Sonnet 4.6**, **Opus 4.7**. Which options are enabled depends on provider keys on the instance and role permissions (Pro requires pro-tools access). - **Files** — paperclip icon: attach a document to the request (allowed types are in file-handling settings). - **Chat history** — **Chat history** in the panel header: separate topics, **New chat**, rename, delete. --- ## Real-world examples **“Where do I enable notifications?”** You are in settings and forgot the section. The assistant explains and offers a UI tour to the right toggle. **“Create a role without scenario access”** An admin describes the task in **Agent** mode — or **Plan** if there are many steps and you want to review the list first. **Onboarding an integrator** Instead of “read 50 pages” — questions in chat in the context of Scenario Builder or Custom Data. **Skills** for scenarios suggest the right way to ask. **Help from a public site** With web access enabled — briefly summarize a partner’s public documentation (without bypassing authentication). --- ## Where to start as a team 1. Root enabled **Instance Agent Integration** — **Menu → Settings → Instance settings → Instance Agent**. 2. Admin configured roles: who sees the assistant, who may allow editing. 3. First weeks — **Ask** mode: get used to answers without automatic changes. 4. The team understands the split: **Copilot** for operators, **AI Assistant** for admins. 5. Complex tasks — **Plan**, especially before bulk changes. --- ## Things to keep in mind - **AI can be wrong** — verify important settings after automatic changes. - **Instance read-only** overrides role edit permission — changes stay blocked. - **Chat history is local** — for audit, look at system changes, not saved browser conversations. - The assistant is **for the team inside the platform**, not for customers in channels. --- ## Where to go next - **[What is the AI Assistant](/en/constructor/explanation/what-is-ai-assistant.md)** — shorter explanation of modes and access. - **[Use the AI Assistant](/en/constructor/how-to/use-ai-assistant.md)** — step-by-step guide for daily use. - **[Configure access](/en/constructor/how-to/configure-ai-assistant-access.md)** — instance, roles, read-only. - **[Web search and URLs](/en/constructor/how-to/enable-ai-assistant-web-research.md)** — for root and role admins. - **[Copilot overview](/en/learn/copilot/copilot-overview.md)** — AI for operators in dialogs. - **[Fast Line Pro overview](/en/learn/fastlinepro/fastlinepro-overview.md)** — first line for customers. - **[Integrator overview](/en/learn/implementer/implementer-overview.md)** — broader platform configuration context.