# Captain assistant
| Item | Details |
|—|—|
| Who it is for | Administrators configure it; agents use Copilot / rewrite |
| Menu location | Left sidebar **Captain**; model switches under **Settings → Captain** |
| Plan dependency | Requires Captain. Reply counts and FAQ counts are billed monthly / against stock |
Captain is the workspace AI assistant: it can **auto-reply to customers** on selected inboxes, and help agents **rewrite, summarise and suggest replies** in the conversation view.
Each workspace can create up to a platform limit of assistants (about 25). Each **inbox can bind only one** assistant.
—
## 1. Capabilities at a glance
| Capability | Who uses it | Description |
|—|—|—|
| Inbox auto-reply | Customers | When a conversation is Pending and the inbox has an assistant bound, Captain answers from the knowledge base |
| Hand over to a human | Customers / agents | When the assistant cannot resolve the issue or a rule matches, the conversation is passed to a person |
| FAQ | Administrators | Q&A knowledge retrieved during auto-reply |
| Documents | Administrators | Upload files or web URLs; parsing produces FAQ drafts |
| Scenarios | Administrators | Define “for this kind of question, add a label / call an API / hand over” |
| Tools | Administrators | Let the assistant call HTTP APIs (look up orders, etc.); count is capped |
| Playground | Administrators | Try the assistant without messaging customers |
| Copilot | Agents | Ask the AI beside the conversation, with current-thread context |
| Rewrite / summarise / suggest reply | Agents | Generate copy in one click; you confirm before sending |
| Speech-to-text | Agents | Transcribe voice messages (requires the matching switch) |
Auto-replies consume **Captain reply quota** (monthly). Copilot overview stats usually do not use that quota. Vector retrieval of knowledge is not billed as “replies”, but generating an answer is.
—
## 2. Before you enable it (administrators)
1. Confirm the plan includes Captain and this month’s reply quota is not exhausted.
2. **Settings → Captain**: choose models for editor, assistant, Copilot and transcription, and whether assistant / Copilot / transcription are on.
3. Optional: connect your own OpenAI-compatible key (BYOK). Matching calls may bypass platform billing; on failure they may fall back to the platform model (per current settings).
4. Prepare knowledge: product URLs, FAQ sheets, Help Centre articles.
Never put secrets in conversations or articles.
—
## 3. Creating an assistant
1. Open **Captain**. If none exists yet, follow the creation wizard.
2. Enter a name and persona / instructions (state clearly who you are, what you must not promise, and when to hand over).
3. Bind **one or more inboxes** (each inbox may have only one assistant).
4. Set **audience and hours** (for example auto-reply only in business hours; outside hours leave for humans).
5. Turn the assistant on.
### 3.1 When auto-reply fires
All of the following must be true:
1. Conversation status is **Pending**
2. No human agent and no Agent Bot is the assignee
3. The inbox has a bound, enabled assistant
4. The contact matches audience rules and is within allowed hours
5. Monthly Captain reply quota is not exhausted
So: once you assign the conversation to yourself, Captain usually stops. Keep it Pending and do not claim it early if you want the assistant to answer first.
—
## 4. Knowledge: FAQ and documents
### 4.1 FAQ
– Add questions and standard answers under **Captain → FAQ**.
– The assistant retrieves these when answering.
– After conversations end, the system may **suggest new FAQs**; approve or reject them under “FAQ suggestions” so bad wording does not enter the knowledge base.
FAQ count is limited by the plan’s **faqs** quota.
### 4.2 Documents
– Add a **web URL** or **upload a file**; the system fetches/parses and proposes FAQ candidates.
– Review manually before approving. Delete or update stale pages, or the assistant will cite old policy.
Published Help Centre articles can also be a knowledge source if vector search is enabled (plan feature “Help Center Embedding Search”). A steadier pattern: Help Centre for customer self-service, FAQ for Captain’s scripted answers.
—
## 5. Scenarios and tools
**Scenarios**: describe a flow in natural language, for example “customer wants a refund → label refund → hand over”. Scenarios can use built-in actions (look up FAQ, hand over, label, note) and custom tools.
**Tools**: register an HTTP endpoint (tracking, subscription lookup). Notes:
– Expose only read-only or tightly scoped write APIs the assistant truly needs
– Verify in a test environment before production
– Custom tools are capped (about 50)
– If a tool fails, instruct the assistant to hand over rather than invent order status
Use Playground with real questions for several rounds before enabling for customers.
—
## 6. Guardrails and system settings
In the assistant’s **Settings / guidelines / guardrails** you can:
– Restrict topics that must not be discussed
– Define handover wording
– Tune reply style and language
– Choose whether to write “memory” after resolve, and whether to extract FAQ suggestions
After changes, regression-test in Playground so the assistant does not suddenly become too quiet or too wordy.
—
## 7. How agents use it
1. Open any conversation.
2. **Suggest reply / rewrite / summarise**: polish a draft, then send after you confirm.
3. **Copilot**: ask in the sidebar, for example “From history, what did this customer complain about last time?”
4. Voice messages: if transcription is on, wait briefly for the transcript.
These features help you, but you remain responsible for what is sent. Confirm refunds and legal commitments manually.
—
## 8. Usage and quality
– Check this month’s `captain_responses` and `faqs` under **Settings → Billing**.
– Charts on the Captain overview help with auto-resolution rate and handover reasons.
– When quota runs out, auto-reply stops and conversations may stay Pending — hand over or upgrade promptly.
—
## 9. Frequently asked questions
**Assistant stays silent?** Check: feature on, inbox bound, conversation Pending, not already claimed by a person, quota, business hours, audience rules.
**Assistant answers badly?** Tighten instructions, delete bad FAQs, add guardrails, force handover on uncertain questions.
**Use with Agent Bot or Dialogflow?** Think of bot assignees as mutually exclusive per inbox: do not let several auto systems fight over one conversation. Pick one primary auto-reply path.
**Customers see the name “Captain”?** Change the assistant display name / handover copy to your brand wording.
—
## Related articles
– [Conversation workspace](./03-conversation-workspace3-conversation-workbench)
– [Help Centre](./08-help-centre8-help-centre)
– [Channels and inboxes](./12-channels-and-inboxes12-channels-and-inboxes)
– [Billing and usage](./26-billing-and-usage26-billing-and-usage)