The AI assistant answers clients who message your Telegram bot: it reads the question, looks the answer up in your knowledge base and drafts a reply in the client's language. Anything missing from the knowledge base is handed over to a human instead of being invented. The assistant runs on Claude models (Anthropic) through your own API key — the platform does not provide one.
Settings live in the admin panel: AI assistant in the left menu (https://app.appexbots.com/app/admin/bot-ai-assistant). They are set per bot: pick the bot in the header first.
The assistant is built into the Reply Bot: client messages must reach your support group with topics. If the Reply Bot is not enabled yet, start there.
| Mode | What the client sees | What managers see in the topic |
|---|---|---|
| Off | nothing | nothing |
| Shadow | nothing | a draft marked "🤖 … · AI draft (shadow mode)" with a 📤 Send to client button |
| Live | a confident answer right away | a copy marked "🤖 … · sent to the client"; when the assistant is unsure — "🙋 … · human needed" with a draft and the button |
The "…" in the note is the question topic: bot_setup, catalog, payment, delivery, billing, bug or other — handy for spotting what clients ask about most.
Start in Shadow mode: read the drafts in the topics for a couple of days, extend the knowledge base and instructions, and only then switch to Live.
Managers do not configure anything: the assistant lives in the same topics the Reply Bot forwards client messages to. A typical day looks like this:
The assistant replies in the language the question is written in, regardless of the client's Telegram interface language; the article link is picked in that language too. In Live mode a "sent to the client" note means the client already has the answer — no need to repeat it.
Each draft in a topic has a 📤 Send to client button. Pressing it sends the draft text to the client on behalf of the bot, exactly as if a manager had written it; the note gets a "✅ Sent to client" mark and the button disappears. Any member of the support group can press it.
If the draft needs changes, do not press the button: copy the text, edit it and send it as a regular message in the topic — the Reply Bot delivers it to the client.
The button appears under every draft in Shadow mode and, in Live mode, only under answers marked 🙋.
The assistant works with any of five providers; pick one in the Model provider field. The API key and the model always belong to the chosen provider, so paste its key after switching.
| Provider | Where to get the key | Default model | Notes |
|---|---|---|---|
| Claude (Anthropic) | console.anthropic.com → API keys | claude-opus-5-5 |
the most accurate answers from the knowledge base, base caching |
| OpenAI (ChatGPT) | platform.openai.com → API keys | gpt-5-mini |
an "OpenAI API key"; a ChatGPT Plus subscription does not work |
| DeepSeek | platform.deepseek.com → API keys | deepseek-chat |
the cheapest option |
| Qwen (Alibaba Cloud) | Alibaba Cloud Model Studio → API keys, international region | qwen-plus |
a key from the international (Singapore) region; the China region does not work |
| Kimi (Moonshot AI) | platform.moonshot.ai → API keys | kimi-k2-turbo-preview |
|
| Z.AI (Zhipu GLM) | z.ai → API keys, international platform | glm-4.6 |
the China platform bigmodel.cn uses a different API address and does not work |
To replace the key, paste a new one into the same field. To revoke it, switch on Remove the saved key and save; the assistant stays silent until a new key is added. A key saved for one bot is not available to other bots or shops.
The Model field takes a model identifier of the chosen provider, exactly as written in its documentation. Leave it empty for the default from the table above (for Claude, the platform default, currently claude-opus-5-5). For OpenAI, DeepSeek, Qwen and Kimi any model with function calling works; without it the assistant cannot look at the catalog.
Claude models:
| Model | When to choose it |
|---|---|
claude-sonnet-5-5 |
recommended: accurate answers from the knowledge base at a moderate price |
claude-haiku-5-5 |
cheapest and fastest; fine for short FAQ-style answers |
claude-opus-5-5 |
the most careful answers to complex questions, the most expensive |
To switch to Sonnet, type claude-sonnet-5-5 and press Save — new questions go through it within a minute, no restart needed. The current list of identifiers is in the Anthropic documentation. A typo in the identifier is not reported in the admin panel: the assistant simply stops answering — double-check the spelling.
The knowledge base is cached on Anthropic's side, so the cost of a question depends mostly on the length of the question and the answer, not on the size of the base.
The Knowledge base field is the only source the assistant answers from. Whatever is not there, it "does not know" and hands over to a human.
Worth including:
Write plain text, split it into sections with headings or blank lines, one fact per line. Up to 1,000,000 characters fit. Changes are picked up within 10 minutes after saving.
For the platform (registration) bot the knowledge base is an export of this documentation; the field is a fallback.
You do not need to copy products into the knowledge base: the assistant of a shop bot searches the catalog itself at the moment of the question. It has two tools, a product search by name and a product card, and calls them when a buyer asks about availability, price, sizes, composition or specifications. In the card it sees the description, the price and the old price, the stock of every variant (size, colour), attributes, categories and tags. Hidden products are never shown to buyers, and stock comes from the catalog at the time of the question, so "do you have size 42?" gets an up-to-date answer.
Put into the knowledge base only what the cards do not contain: delivery and payment terms, opening hours, returns, frequent questions. With a catalog the assistant answers even when the knowledge base is empty — but then everything that is not about products goes to a human.
For the search to work, product names should match how buyers call them: "Nike Air sneakers" is found by "sneakers" and by "nike", but not by an article number in the name. One question with a search and a card costs 2–3 model calls.
Instructions define the character: who the assistant is, whom it speaks for, the tone, what it never does. One text for all languages — the assistant replies in the language of the question. Leave it empty for the platform defaults.
Regardless of the instructions, the platform always requires the assistant to:
Example instructions for a shop:
You are the assistant of the "Daisy" shop in Telegram. Be polite, short and friendly.
If the client asks about a specific order, ask for the order number and hand the question to a manager.
Never promise discounts or delivery times that are not in the knowledge base.
Answers per client per day caps spending: after N answers in a day the assistant goes quiet for that client, and their messages reach managers as usual. The default is 30, 0 means no limit. Days are counted in UTC.
Changes to the mode, the key and the model apply within a minute.
/ — the assistant skips those.