AI reply classification and drafts
Every reply is sorted into interested, question, referral, out of office, not interested or unsubscribe. Categories drive labels, notifications, campaign steps and analytics, and your team can correct them. AI drafts help you answer faster.
When a lead replies on any channel, GraphKit reads the reply and sorts it into a category. The category appears on the conversation, can add an inbox label, can notify your team, can stop or branch a campaign, and feeds the positive reply rate in analytics. Anyone who can reply can change it.
Categories
| Category | Means |
|---|---|
| Interested | Wants a call or a meeting, asks for pricing or details because they are considering it, or says yes. |
| Question | Asks something before deciding (what is this, how does it work). |
| Referral | Points you to someone else ("talk to my colleague"). |
| Out of office | An away notice or auto-reply. |
| Not interested | Declines, not a fit, already has a solution, not now. |
| Unsubscribe | Asks you to stop messaging or to be removed. |
Only the lead's latest reply is classified; earlier messages in the conversation are used as context. Group chats and your own messages are never classified. Replies imported from the last 90 days when an account first connects are not classified, so history does not use your allowance.
Opt-out keywords (for example "stop" on WhatsApp) are classified as Unsubscribe straight away, without AI. See Suppression and privacy.
Where categories appear
- Conversation list. A small chip on each classified conversation. A sparkle means the AI set it.
- Conversation header. The category with a menu to change or clear it.
- Reply intent filter. The inbox's left side filters by category, alone or with any view, channel or label.
- Labels. With Add the category as an inbox label on (the default), the conversation gets a label named after the category. The label is replaced when the category changes, so a conversation carries one category label at a time.
Correcting a category
Open the conversation and choose the category in the header, or Clear category. Corrections:
- apply to the conversation's latest reply, and take effect at once;
- drive the same signals as the AI (label, notification, campaign steps, suppression for Unsubscribe);
- keep what the AI said, so your workspace's correction rate can be measured.
A new reply from the lead is classified again, even after a correction, because the lead's intent may have changed.
Confidence
The AI gives each answer a confidence. Every answer is shown and labelled, but:
- below 60 %, it does not notify anyone or move a campaign;
- an Unsubscribe answer suppresses the lead on every channel only at 90 % or more, because a suppression stops every campaign for that lead. Below that, the conversation shows Unsubscribe for your team to confirm: choosing Unsubscribe yourself suppresses the lead.
Notifications
An Interested reply (from the AI at 60 % or more, or set by a teammate) sends a Positive reply notification to the people who can see the conversation, in the app and on the channels your team routed it to. A reply is announced once, even if the category is changed later. See Notifications.
Campaign signals
Two campaign features use categories:
- Stop on positive reply (campaign settings). Stops a lead once a reply is classified Interested. Use it with Stop on reply turned off, so out-of-office and question replies let the sequence continue while a real yes ends it.
- Reply category condition (sequence builder). Branches on the category: Yes as soon as a reply is classified in one of the categories you pick (for example Interested or Question), No when the wait runs out. You can limit it to replies on one channel. Classification takes a few seconds after the reply arrives, so give the condition a wait of at least an hour. With Stop on reply on, leads stop before reaching the Yes branch, and the builder warns you.
When a reply is classified, the lead's open campaign enrollments are re-checked right away.
Analytics
The Reply intent card on the analytics page shows how many replies landed in each category in the period, and the share that were Interested. The positive replies metric (and positive rate) counts replies in conversations classified Interested, plus conversations carrying a positive label (Interested, Positive, Meeting booked and the others listed in Analytics). Campaign analytics count positive replies the same way.
AI drafts
In a conversation, ask for a draft and GraphKit writes a reply from the conversation, the lead's name, headline and company, the campaign, and your team's guidance. You can add a one-off instruction ("suggest Tuesday"). The draft goes into the composer for you to edit; it is never sent on its own. Drafts follow the lead's language, stay short for chat channels, and leave a [placeholder] instead of inventing prices, dates or links.
Set the team's guidance in Settings → AI → Guidance for every draft: tone, offer, booking link, things never to say.
The AI allowance
Reply classification and drafts use the same monthly AI allowance as Kit. It belongs to the organization, so every workspace and Kit draw on it together. Each call costs what the model charges for it, so a short classification costs a fraction of a cent. The allowance resets on the 1st of each month (UTC). Settings → AI shows how much has been used, the limit and the reset date.
- On a platform key (GraphKit's key), each call is checked against the allowance before it runs.
- On your organization's own key (added in Kit settings), calls are recorded but never capped. Your provider bills you directly.
- A retried classification is charged once. A draft or classification that fails costs nothing.
- When the allowance runs out on a platform key, new replies are not classified until it resets (you can still set categories by hand), and drafts show "AI allowance used for this month". Kit stops at the same point.
Settings
Settings → AI (workspace admins and agency roles can change them):
| Setting | Default | Effect |
|---|---|---|
| Classify new replies with AI | On | Off: only opt-out keywords are classified; you can still set categories by hand. |
| Add the category as an inbox label | On | Off: categories show as chips and filters only. |
| Allow AI drafts in the inbox | On | Off: the draft button answers "AI drafts are turned off". |
| Guidance for every draft | Empty | Added to every draft request. |
Privacy and data handling
- To classify a reply, GraphKit sends the model the channel and the last eight messages of that conversation (text only, with timestamps). To draft, it sends up to the last sixteen messages, the lead's name, headline, company and location, the campaign name, the sending account's name and your guidance. No attachments, no other conversations, no lead lists.
- Messages are treated as data: instructions written inside a message are ignored.
- Requests go to the provider that serves the model: your organization's own key when it has one (Kit settings), otherwise GraphKit's platform key. Results are stored with the conversation; the model's raw answer is not kept beyond the category, confidence and a short reason.
- Drafts and categories respect inbox privacy: you can only draft for or categorize conversations you can see. Private conversations are visible only to people who can read them.
Configuration (self-hosted and operators)
| Variable | Default | Meaning |
|---|---|---|
AI_REPLY_MODEL | claude-haiku-5-5 | Model for classification and drafts: a curated model key from pricing/kit-models.json, or <provider>:<model id>. |
Keys and the allowance are Kit's: the platform keys OPENROUTER_API_KEY, ANTHROPIC_API_KEY and OPENAI_API_KEY, organizations' own keys, and KIT_MONTHLY_ALLOWANCE_CENTS_PER_SEAT. The model is resolved the same way Kit resolves it: the organization's own key first, then a platform key. When no key serves the model for an organization, classification is skipped, drafts answer ai_not_configured, and Settings → AI shows a notice.
API
| Endpoint | Purpose |
|---|---|
GET /api/app/workspaces/{workspaceId}/ai | Configuration, the key source and Kit's allowance (usage, null when not configured), settings. |
PUT /api/app/workspaces/{workspaceId}/ai/settings | Change the settings (team.manage). |
PUT …/inbox/conversations/{conversationId}/classification | Set or clear (null) the category (inbox.reply). |
POST …/inbox/conversations/{conversationId}/ai-draft | Draft a reply (inbox.reply). Errors: 422 ai_not_configured, 402 ai_allowance_exhausted, 403 ai_disabled, 502 ai_unavailable. |
GET …/inbox/conversations?replyCategory=interested | Filter conversations by category. |
Conversations carry replyCategory and replyCategorySource (ai, user or rule).
Unified inbox
Read and answer every conversation of every connected account in one inbox, with filters, assignment, labels, notes, saved replies, snooze, scheduled send, AI drafts, exports, a command menu, privacy and opt-outs.
Kit, the AI agent
Ask Kit to do anything you can do in GraphKit, from drafting replies to building and launching campaigns. Kit acts with your permissions and asks before anything consequential.