AI Compose & briefings
Requires the Escalation module.
A handoff is expensive for two reasons: the agent has to read the whole conversation before they can help, and then they have to write. Promptly gives agents three AI helpers for exactly those two problems — a briefing that replaces the read, and AI Compose plus copilot suggestions that take the first pass at the write.
All three are switched on under Dashboard → Settings → Escalation → AI Behavior.
The AI briefing
With Generate handoff summary for agents enabled, the AI writes a briefing the moment a conversation is escalated. It appears as a collapsible AI briefing panel at the top of the conversation in the Inbox.
The panel gives you:
- Intent — what the customer is actually trying to do.
- Key details — the entities pulled out of the conversation: product names, order numbers, dates.
- What the bot tried — so you don't repeat a failed answer as your opening line.
- Sentiment (negative / neutral / positive) and urgency (low / normal / high) badges.
- Detected language — the language to reply in, which matters when your team handles several.
- Suggested first response — a ready-made opener. Copy it, or click Use to drop it straight into the reply box.
The briefing is generated in the background and never delays the handoff. If it isn't ready yet you'll see "Briefing pending — the AI is still summarising this conversation…", and the panel fills itself in when the summary arrives. It reads the last 30 messages and gives up after a few seconds rather than making an agent wait — a missing briefing is a missing convenience, never a blocked conversation. The panel collapses on its own once you start replying.
🖼️ [Image] — The AI briefing panel expanded, showing intent, key details, and a suggested first response.
AI Compose / Polish
The reply box has an AI button whose label changes with what's in the box:
- Compose — the box is empty. Click to generate a full reply from the conversation and the briefing.
- Polish — you've typed something. Whatever is there is treated as a steering hint, not as text to preserve. Type
offer refund, apologise for the delay, no shipping costand you get a written-out message on exactly that line.
Polish is the more useful of the two in practice: it's faster to jot three bullet points than to edit a paragraph the AI guessed at.
What Compose sees: the recent conversation (roughly the last dozen messages), a note of which attachments were sent by whom, and the AI briefing if one exists. It replies in the language detected for the conversation.
Two things worth knowing:
- AI Compose never sends anything. The draft lands in the reply box for you to read, edit, and send yourself. There is no auto-send path in a handed-off conversation.
- Compose writes in a support-agent voice, not your widget's brand persona. It is deliberately separate from the bot's system prompt — you're writing as a person on your team now, not as the assistant. Expect to adjust the wording if your brand tone is distinctive.
🎬 [Video] — Typing a rough note, clicking Polish, and editing the AI draft before sending.
Copilot suggestions
AI copilot suggestions for agents is the third helper and the only proactive one. With it on, the inbox drafts a suggested reply on its own after the customer writes, so a suggestion is often already waiting when you open the thread. If there is nothing to answer — the last message is yours, not the customer's — no suggestion is offered.
Compose and copilot overlap on purpose: the copilot handles "the customer just asked something, get me started", Compose handles "I know what I want to say, write it properly".
Cost note
These features call an AI model, and under Promptly's bring-your-own-key model the tokens are billed to you by your own provider. Briefings and copilot drafts are short, but they are not free — leave Generate handoff summary on (it saves the most agent time per token) and switch copilot off if you want to trim usage.