Agents & automation

1. 1The agent already stuck to the conversation — the one it was handed over to, the one you are trying out from the “Test” panel, or the agent that answered the earlier messages in the same thread.
2. 2The department the conversation was sent to — either because the bot handed it over there, or because that department is the workspace entry point. One of THAT department’s active agents answers, and the customer’s message decides which one.
3. 3The agent marked as default (the entry point) — but only if it is switched on. A default agent that is off is skipped entirely, as if it were not marked at all.
4. 4If none of the above applied: every active agent in the workspace is a candidate, and the customer’s message decides which one fits best. Creation order (oldest first) is only the tie-breaker, for when the message gives nothing to decide on.

If no agent is switched on, nobody answers — the channel stays silent. That is taken as your decision, not as an error.

If the chosen department has no active agent, the choice simply moves on down the list. A conversation is never left unanswered just because a department is empty.

Once picked, the agent stays on that conversation: the following messages go straight to it instead of being re-decided every time. That is what keeps the bot from changing personality between messages.

• Write in sections with headers: who it is, how it sounds, what steps it takes, what it must not do.
• Point form, not paragraphs. One instruction per line, steps numbered in the order they happen.
• Write behaviour, not knowledge: how to answer, not what the model already knows.
• Never restate a setting — business hours, timezone, channel and documents reach the agent automatically.

1. 1CONTEXT — who it is talking to, and in what situation.
2. 2ROLE & TONE — how it sounds and how long its replies run.
3. 3WHAT IT DOES — the steps, numbered, in the order they really happen.
4. 4BOUNDARIES — what it must never say or promise.

• Auto-close — the AI marks the conversation "Closed" when the customer confirms it's resolved (e.g. "thanks, that's all"). It leaves the active list, and a new customer message reopens it automatically. You can also close manually via "Mark resolved".
• Auto-assign to a teammate — the AI hands the conversation to a team member: the status becomes "Human" (the bot goes silent) and the teammate gets an email with a direct link. Name the person in the instructions ("assign to Maria") or let the system pick.
• AI-updated customer status — the AI moves the contact through the Contacts pipeline (New → Contacted → Qualified → Won / Lost), silently, based on the conversation.
• AI-filled contact details — the AI saves the customer's name, email, phone and company when they come up in the conversation (name on the contact, the rest on the note).
• AI-updated tags — adds/removes tags on the contact, preferring the existing workspace tags.
• Route to an AI department — the bot does not go silent: it keeps answering as the chosen department’s specialist (the department prompt takes over). Different from the human handoff, where the bot stops.

How to write the instructions — examples that work well:

1. 1Close: "Close only when the customer explicitly confirms it's resolved and has no more questions. Never close when a human follow-up was promised."
2. 2Assign: "If the customer asks about invoices or payments → assign to Maria. If they want a custom quote → assign to Daniel. Otherwise, round robin."
3. 3Status: "Real buying intent → Qualified. Confirmed order or booking → Won. Clear refusal → Lost."

• Assignment methods: round robin (fair rotation, default) or least_open (the teammate with the fewest open conversations).
• The AI finds teammates by name or email, with or without diacritics.
• Limits: no scheduled assignments ("assign at 5pm") and no online/offline awareness yet — the pick runs across the whole team.
• After an assignment the bot stays silent on that conversation until you return it to AI from the dashboard.

Conversations settings — closing note and categories:

• AI-generated closing note — when a conversation is closed (by the AI or via "Mark resolved"), the AI writes a 2-3 sentence summary (what the customer wanted, how it was resolved) shown on the conversation in the inbox.
• Conversation categories — define your own list (e.g. Support, Sales, Complaint); on close the AI picks the best fit, shown as a chip on the conversation.
• Custom contact fields — from Contacts → ⋯ → "Custom fields" you define your own fields (Text / Number / Date / List, e.g. "Birthday", "Plan"). Edit them on each contact; the AI fills them automatically from conversations when the "Contact details" action is on.

The inbox copilot — when you take over the conversation:

• ✨ Suggest reply — the AI drafts your next reply from the conversation and the knowledge base; you review, edit and send it. Nothing is sent automatically.
• AI Prompts — transforms the text you already typed: professional / friendly tone, grammar fixes, RO/EN translation, shortening.
• Conversations routed to an AI department show a colored chip with the department name, in the list and in the detail view.

• Auto — automatic routing: simple questions go to the cheap model, complex ones to the strong model.
• Haiku — most economical. Perfect for knowledge-base questions (FAQ, pricing, policies).
• Sonnet — most powerful. For complex reasoning and expert tasks.

For customer-facing bots (WhatsApp, website, Telegram) the recommendation is Haiku — it answers simple questions efficiently at a fraction of the cost. The model is chosen from the agent’s “AI model” card.