AI Support Assistant

The AI Support Assistant is a support bot that answers players' questions right in your chat, grounded in the help materials you upload in the admin panel. It works in two modes that can run at the same time in one project.

Both modes use the same knowledge base, the same answer pipeline and the same safety checks. Every answer is verified against your materials before it is published. A claim the assistant cannot confirm is replaced by an honest refusal that points the player to your support contacts. The assistant replies in the language of the player's message.

Each conversation becomes a support ticket in Watchers: the player can resolve or reopen it from the thread, and your team sees the tickets in the admin panel.

Mode 1: answering on request

In this mode the assistant answers only when a player addresses it. A message counts as addressed when one of the following is true:

• The text mentions the bot by name: @Support how do I withdraw?. The name must stand as a whole word, so a bot called Support does not react to @Supporter.
• The message is a reply to one of the bot's own messages. No mention is needed.
• The message is written in the bot's personal room (the one-on-one chat opened from the personal bot button).
• The message is written inside a support thread that the bot already opened. The player keeps talking without mentioning the bot again.

What happens to the question. When a player mentions the support bot in a group room, Watchers marks that message as private: only the author and the bot see it, the rest of the room does not. Mentions written by moderators or by other bots (for example an announcement "write to @Support") stay public and do not trigger the bot. At the same moment Watchers opens a support ticket for the player with the status Open.

Where the answer goes. The place depends on the room and on the "Reply privately" toggle of the bot:

Inside the thread. The thread carries the bot's name in the header and shows a one-time notice "AI can make mistakes" that the player dismisses. The assistant remembers the question that opened the thread, so follow-ups like "and how long does it take?" are understood in context. After its first answer the assistant writes a short title for the ticket in the player's language, so your support team can read the ticket list without opening every thread. A title that someone already set is kept.

Resolving the ticket. The player closes the conversation with the Resolve button in the thread header. In a resolved thread the assistant no longer answers from the knowledge base: it replies that the request is closed and that mentioning the bot in the room opens a new one. The player can reopen the thread with the Reopen button, and the assistant answers again. Active and past support chats are listed in the player's support history in the chat.

Rate limit. One player can send the assistant up to 10 messages per minute. Above that the assistant replies that there are too many requests and asks to wait.

Mode 2: proactive monitoring

In this mode the assistant watches every message in the public rooms of the project and steps in when a player describes a support problem without addressing anyone. Personal rooms and support threads are not monitored, they already belong to Mode 1. Messages that mention the bot or reply to it are also left to Mode 1, so a player never gets two answers to one message.

How the assistant decides. Each new message passes a chain of checks, and a failure at any step means silence, never a wrong public reply:

1. Messages from bots, including the support bot itself, are skipped.
2. A classifier reads the text and decides what kind of message it is and how confident it is. Uncertain verdicts are dropped.
3. Only two kinds of messages lead to an answer: complaints about the platform (deposits, withdrawals, payouts, bonuses that did not arrive, games that do not work, service) and ordinary messages that are in fact a support question (account, payments, bonuses, promo codes, games, verification). Everything else, such as insults, spam, personal data or politics, is left to moderation.
4. A rant with no concrete question or fixable problem ("this casino is a scam, it never pays") is not treated as a support request and gets no answer.
5. Per-player and per-room limits are applied so the assistant does not flood a room: at most 2 unsolicited answers per player per minute and 6 per room per minute.

How the assistant answers. The answer is generated from the knowledge base exactly as in Mode 1. The reply is always private, whatever the "Reply privately" toggle says: the assistant opens a personal thread under the player's message and answers there. Nobody else in the room sees that the bot reacted. The player can continue in the thread without mentioning the bot, and from that point the conversation is handled as in Mode 1: the assistant remembers the opening exchange, Watchers opens a support ticket for the thread, the ticket gets a title, and the player can resolve it.

What the player sees. Under their own message a "1 reply" thread counter appears. Opening it shows the assistant's answer with the usual thread header and the "AI can make mistakes" notice. Moderation runs independently: a complaint with words like "scam" may be auto-hidden from other players by your moderation settings, while the author still sees it and still gets the assistant's thread.

Enabling the mode. Proactive monitoring is switched on per project by the Watchers team. Contact your account manager with the project name.

Knowledge base and answer generation

The assistant answers only from the documents you upload for it. In both modes the answer goes through the same steps:

1. Retrieval. The question is matched against the fragments of your documents in an isolated per-project index. Greetings and small talk are recognized and answered without a search.
2. Ranking. The candidate fragments are ranked by usefulness for the question and the best ones are passed to the model.
3. Generation. The model composes a short plain-text answer from those fragments, following the bot profile (brand, operating countries, persona, tone of voice, extra rules) and the conversation history where there is one. The answer language always follows the language of the question, even when the documents are written in another language.
4. Verification. A separate check confirms that every factual claim in the draft (numbers, timelines, procedures) is supported by the fragments and that the answer is safe and on topic. If the answer cannot be confirmed, it is replaced by a refusal in the player's language that suggests contacting your support.

What the assistant never does. It does not invent facts that are not in your documents. It does not promise response times or actions it cannot perform (escalating a case, resending an email, changing an account). It does not discuss another player's account. Instructions hidden in a player's message ("ignore your rules") do not change its behavior. Questions outside the support scope, such as sports results or general knowledge, are politely declined in one sentence.

Documents. Upload them on the Documents tab of the bot in the admin panel. Supported formats are PDF, Markdown and plain text, up to 25 MB per file, up to 50 documents or 200 MB per bot. A document goes through the statuses Queued, Ready and Failed. A failed document shows the reason (no text layer, too short, unreadable encoding, unsupported type) and does not answer questions until replaced. Each document gets a quality score from 0 to 5 and warnings such as "Looks like an internal document" or "Some fragments are low quality". Re-uploading an identical file is ignored, deleting a document removes it from the knowledge base immediately.

What to upload. Materials written for players: FAQs, how-to guides, terms summaries. Keep internal documents out: call-center scripts, escalation matrices and operator instructions leak into answers as if they were commitments to the player. Avoid time-sensitive content such as lists of current promo codes or dated promotions; describe the process instead (where codes are published, why a code may not work). One question per entry and explicit numbers retrieve best.

Setup

Everything for Mode 1 is done in the admin panel, section Chatbots.

1. Check the Bot API add-on. It must be enabled for the project (APIs and Integrations). Without it the bot cannot connect to the chat.
2. Create or open a chatbot. Give it the name players will mention, an avatar and, if needed, a list of rooms it is restricted to.
3. Turn on "Enable AI Support Assistant". This marks the bot as the support bot of the project. Only one active bot per project can be a support bot: the admin panel refuses a second one until the first is deactivated. A bot cannot be a sports assistant and a support assistant at the same time.
4. Choose "Reply privately". On: answers to mentions in group rooms open a private thread under the question. Off: answers are posted publicly with a quote. Personal rooms and proactive answers are private either way.
5. Fill in the AI Agent section. Brand Name and Operating Countries go into the bot form. The Watchers team can additionally configure a persona name, a tone of voice and extra rules for your assistant ("Do not discuss competitors"). Tone and rules shape style and add restrictions; they cannot weaken grounding, safety or privacy rules.
6. Save, then open the Documents tab and upload the knowledge base. The tab appears after the bot is saved as a support bot. Watch the statuses: the assistant answers from a document only when it is Ready.

Changes apply within a minute without a restart. Every save of the bot and every document change is synchronized to the assistant automatically.

Turning the assistant off. Deactivate the bot or switch off "Enable AI Support Assistant". The bot stops answering in this project; its profile and documents are kept, so switching it back on restores the knowledge base. Deleting the bot deletes its documents and knowledge base permanently.

Support tickets in the admin panel. Tickets opened by the assistant's conversations are available in the AI Assistants dashboard, where each request shows the room, the bot type (Support), the bot name and the player, and through the Admin API (admin/support). A ticket holds the player, the opening message, the thread, the generated title and the status (Open or Resolved).

Limits and behavior

Languages. The assistant answers in the language of the player's last message, including refusals and small talk. Documents can be in any language; facts are translated when needed. The ticket title follows the player's language too.

Privacy of the conversation. A mention of the support bot in a group room is hidden from other players. Threads opened by the assistant are visible only to their owner. Interaction logs, where enabled, mask personal data and store player ids as salted hashes.

Moderation. Moderation runs independently of the assistant. A player's message can be auto-hidden for other users by your moderation rules while the assistant still answers its author. The assistant's own answers are published as bot messages and never trigger another bot, so two bots never talk to each other.