On this page
Too long? Read this
- A reply suggestion is an AI draft the agent sends, edits or discards; the customer never receives anything a person did not approve
- Grounded in the same knowledge base as the chatbot, with the conversation as context, in the customer's language
- It is the safe first step for teams not ready for an autonomous bot, and it exposes knowledge base gaps quickly
- Setup takes an hour: knowledge sources, a system prompt for tone, and a plan that includes AI (Growth and above)
Most support teams are not ready to let a bot answer customers unattended, and they are right to be cautious. Reply suggestions are the middle ground: the AI drafts the answer inside the agent's composer from your own knowledge base, the agent reads it, edits it if needed and sends it, or throws it away. Speed goes up, tone stays human, and every mistake is caught before it reaches a customer. This guide explains how suggestions work, when they help, how to set them up in an hour and how to judge their quality.
How a suggestion is made
When an agent taps the sparkle button in the composer, the AI receives the recent messages of the conversation, the contact's basic details, the channel's system prompt (tone, language, what never to promise) and the most relevant passages retrieved from the knowledge base for that channel or widget. It writes a reply in the customer's language and places it in the composer. Nothing is sent. The agent edits, sends or discards; the choice is logged so the team can see how often suggestions are used as written.
Suggestions versus the chatbot
| Reply suggestion | AI chatbot | |
|---|---|---|
| Who sends | The agent | The AI |
| When it acts | On the agent's tap | On every customer message it is allowed to answer |
| Risk | A bad draft is caught by the agent | A bad answer reaches the customer |
| Speed | Seconds saved per reply | Instant, around the clock |
| Best for | Teams starting with AI, sensitive topics, high-value customers | Routine questions at volume, out of hours |
| Knowledge | Same knowledge base | Same knowledge base |
Most teams run both: the chatbot answers hours, prices and status questions and escalates the rest; agents use suggestions on what the bot handed over.
Setting it up
- Plan. The AI assistant is on Growth and above.
- Knowledge base. Under the chat widget or channel's Knowledge Base tab, add sources: your website's FAQ, pricing and policy pages as URLs; pasted text for documents; FAQ pairs for the questions agents answer most. Aim for coverage of the top fifty questions.
- System prompt. Write the voice: "You are the support assistant for {{business}}. Be brief, warm and specific. Answer only from the provided knowledge. Never promise refunds or delivery dates not in the knowledge. Reply in the customer's language." Add house rules (address the customer by first name, no emojis, mention the support hours).
- Model. The default Groq-hosted model works for most; bring your own OpenAI or Anthropic key under AI settings if you prefer.
- Test. Open a conversation, tap the sparkle, read the draft, and repeat for the ten hardest recent conversations. Fix the knowledge base wherever the draft is wrong or vague.
The AI replies help article has the settings table.
When suggestions help most
- Long explanations agents type repeatedly: return policies, service inclusions, document requirements.
- Multilingual inboxes where an agent is weaker in the customer's language.
- New agents, who get an answer that reads like the best agent's.
- Peak hours, when every reply is a minute shorter.
- Sensitive conversations, where a bot should not act but a draft still saves time.
Judging quality
Track three numbers weekly: the share of suggestions sent unedited (a healthy base is above half), the share discarded (above a fifth means gaps), and first-response time before and after. Read ten random suggestions with their conversations: wrong facts point to a missing or stale source; wrong tone points to the prompt; wrong language points to a prompt line about language. Suggestions never invent facts they are allowed to state; when the base has nothing, they say so, which is the right behaviour and the clearest signal of what to add.
Writing a system prompt that produces good drafts
Keep it short and concrete. State who the business is and what it sells in one line. State the tone in three adjectives. State the language rule ("reply in the customer's language; if unclear, English"). List what the assistant must never do: promise a delivery date not in the knowledge, offer discounts, discuss competitors, give medical or legal advice. Give two example exchanges in your voice. Prompts longer than a page produce worse drafts than prompts of ten lines, because the model weighs everything equally; put the important rules first and cut the rest.
Examples of good and bad drafts
A good draft to "do you deliver to Pune?" reads: "Yes, we deliver to Pune; standard delivery takes 3 to 4 days and is free above Rs 999. Want me to check your pincode?" It is specific, sourced from the shipping page, and ends with a question. A bad draft reads: "Thank you for reaching out! We would be delighted to assist you with your delivery query. Please share your location." It is padded, sourced from nothing, and delays the answer. The difference is almost always the knowledge base: the first business had a shipping page in it, the second did not.
Privacy and control
Only the conversation needed for the draft is sent to the model provider, under the DPA, and it is not used to train models. Bring your own key to route through your own provider account. Roles decide who can use the sparkle; the audit log records AI settings changes. The customer never sees the AI unless the agent sends its words.
Suggestions on the Android app
The sparkle button is in the mobile composer too, so an agent answering from the phone gets the same draft from the same knowledge base, edits with the keyboard and sends. Drafts are generated server-side, so a weak connection delays the draft by a second or two but never produces a different answer from the web app.
From suggestions to automation
Once the unedited-send rate is high on a category of question, that category is ready for the chatbot to answer on its own with an escalation rule for anything outside it. Suggestions are how a team learns which questions can be automated, one week of real conversations at a time. The AI chatbot guide covers the next step.