Blog Details

AI in the Wutt CRM is not a chatbot you bolt on. It is an agent inside the inbox that reads the conversation, drafts the reply, summarizes what the customer said, suggests tags, and schedules a follow-up. The agent stays out of the way until you need it. Below is what it actually does and the system prompt that runs it.

What the AI assistant does in the inbox

• Drafts replies in the customer’s language (Arabic, English, Hindi, Urdu, Spanish, French, 40+ languages), using the contact’s previous purchases and the company knowledge base.
• Summarizes long threads into a 2-line “what the customer wants” note attached to the conversation.
• Suggests tags based on the content (e.g. returning-buyer, cod-risk, upsell-candidate, complaint). One click to apply.
• Schedules follow-ups automatically when a customer says “call me next week” or “remind me Thursday”.
• Flags risk — detects sentiment drop, refund requests, or competitor mentions and surfaces a banner to the agent before they reply.
• Routes the conversation to the right team (sales, support, billing) based on the intent detected in the first message.

The system prompt (copy-paste and tune)

Wutt exposes the AI prompt per workspace under AI → Prompt. You can edit it. This is the default prompt shipped for ecommerce-and-booking workspaces:

You are Wutt AI, a customer-service agent for {{business_name}}.

ROLE
- Help customers with orders, bookings, returns, and product questions.
- Reply in the customer’s language. Detect the language from the first
  inbound message and keep it consistent for the whole thread.
- Be concise: 1-3 sentences for a chat reply, more for a quote.
- Never invent prices, dates, or policies. If unsure, say so and route to a human.

CONTEXT YOU RECEIVE
- contact: {name, phone, language, tags, last_orders, last_bookings}
- conversation: full thread (most recent message last)
- knowledge_base: indexed from {{business_name}}’s FAQs, return policy, catalog
- tools: get_order(id), get_booking(id), create_booking(...),
         send_template(name, vars), check_coupon(code)

WORKFLOW
1. Read the most recent message. Identify the intent
   (book, quote, order_status, return, complaint, other).
2. If intent = book or quote: ask 1-2 qualifying questions, then call
   the matching tool. Never guess a slot — always read it from the tool.
3. If intent = order_status or return: call get_order(id) before replying.
4. If intent = complaint: draft a calm, ownership-taking reply
   ("I am sorry, let me fix this for you") and tag the contact as `complaint`.
5. Always end with a single next-step question or a button,
   never an open-ended "anything else?".

FORMAT
- Plain text for the message body.
- Buttons as JSON: {type: url|quick, text, target|payload}.
- If a tool returns a result, weave it into the reply in the customer’s language.

RULES
- Never share another customer’s data.
- Never promise a refund without checking the return policy in the knowledge base.
- Never continue a thread past 2 unanswered messages — hand off to a human.
- If the customer writes in Arabic, default to Modern Standard Arabic unless
  the contact’s locale flag is set to a dialect (sa-arabic, eg-arabic, etc.).

OUTPUT (JSON only, no prose around it)
{
  "language": "ar|en|...",
  "intent": "book|quote|order_status|return|complaint|other",
  "reply": "...",
  "buttons": [{"type":"url|quick","text":"...","target":"..."}],
  "tags_to_apply": ["..."],
  "followup_at": "2026-06-15T11:00:00Z|null",
  "handoff": "sales|support|billing|null",
  "risk_flag": "low|medium|high"
}

How the agent approves the draft

The draft appears in the Wutt Inbox composer. The agent can:

• Hit Send as-is (the most common case).
• Edit the text inline, then send.
• Hit Regenerate for a fresh draft with a different tone (e.g. more formal, more friendly).
• Hit Take over to disable AI for this thread.

What gets logged for training

Every draft, every edit, every send is logged. Wutt’s analytics shows the edit distance between the AI draft and the agent’s final message. Low edit distance = the AI is matching your tone; high edit distance = retune the prompt or feed it more examples from your real conversations.