AI Employees and AI settings
Set up an AI Employee that answers customers from your own knowledge, and manage AI keys, credits and budgets.
On this page
An AI Employee is a named AI agent that answers inbox conversations from your own knowledge. It sends replies on its own or drafts them for approval, and hands the conversation to a person when it can't help.
Find them under AI Employees in the sidebar. You need Use AI employees (owners, managers and editors have it by default) and a plan that includes AI Employees.
Before you start: AI must be set up
AI Employees, writing help and AI workflow steps run on one of two things:
- Included AI credits that come with the workspace's plan, or
- your own provider key from OpenAI, Anthropic (Claude), Google Gemini, DeepSeek, Mistral, xAI, Groq, Azure OpenAI, Amazon Bedrock or any OpenAI-compatible service.
If AI isn't set up, an organization owner or admin does it in Settings → AI models & credits (see below).
Create an AI Employee
- Go to AI Employees and click New AI Employee.
- Give it a Name and a Role (Customer Support, Community Manager, Sales or Custom).
- Write the System prompt: its core instructions. Add a Tone (for example "friendly and concise") and lists of things it should and shouldn't do, one per line.
- Choose the Reply mode: Send automatically, or Draft for human approval.
- Set up the hand-off (see below) and the Tools it may use.
- Click Create.
Give it knowledge
On the AI Employee's page, under Knowledge, choose Add knowledge from:
- Text: paste it in. Headings such as "# Shipping" help, because answers are matched section by section.
- Document: PDF (up to 20 MB, text PDFs, not scans), Word .docx, .txt, .md or .csv (up to 5 MB).
- Web page, or Website (crawl) with Pages to import (up to 50 pages from the same site).
Click Add knowledge. Use Test your knowledge to check what it finds, and Playground to try a customer message.
Tools and approvals
Tools decide what the AI Employee may use while answering:
- Read tools only look things up: Search knowledge, Read contact profile, Read contact fields, Read past conversations, Read conversation status (status, priority, assignee, snooze, SLA), Read business hours (open now, next opening, holidays), Browse saved replies and Use saved replies (so it reuses answers your team approved), and (if your administrator has turned web search on) Search the web and Open web pages.
- Action tools change the conversation or the contact: Add labels, Assign conversations, Write internal notes, Set priority, Open or resolve conversations, Snooze conversations, Schedule follow-ups (snooze plus a note for the team), Escalate to a team, Tag contacts and Update contact fields (only fields your workspace already uses, never consent, email or phone). New AI Employees start with read tools only; turn action tools on one by one.
Action tools wait for a teammate's approval unless you tick them under Run without approval. A card marked Needs approval then appears in the conversation, showing exactly what the AI wants to do. Click Approve or Reject (you need to be able to reply on that account). An approved action still needs your own rights: for example changing status needs the account's status right, and contact edits need permission to manage contacts. Requests expire if nobody decides in time or the customer writes again.
Organization owners and admins can add connected-app and MCP tools under Apps & MCP tools. If you don't see it, your Luminobot administrator hasn't turned it on yet.
Draft or send automatically
- Send automatically: replies go straight to the customer.
- Draft for human approval: each reply appears in the conversation with a Draft badge. Click Approve & send or Discard.
Assign conversations to it
It only answers conversations assigned to it. In a conversation, open the AI Employee panel, pick one in Assign an AI employee… and click Assign. Click Unassign to take it off. Workflows can do it with the Assign AI employee step (see Build a workflow).
By default it doesn't reply inside group chats; turn on Reply in groups if you want it to.
Hand-off to a person
The AI hands the conversation to your team when:
- the customer writes one of your Escalation keywords,
- Escalate on negative sentiment is on and the customer is upset, or
- it can't answer safely.
The panel then shows Handed off with the reason. You can send the customer a Handoff message, and the team gets an internal note starting "AI summary at hand-off". The workflow trigger AI hand-off to human lets you route these chats automatically.
Rules it always follows
- WhatsApp Business (Cloud API): WhatsApp only allows AI assistants that stick to your business, so off-topic requests (homework, essays, code) get a polite reply that it only helps with your products, orders, bookings and support.
- 24-hour window: on Messenger, Instagram and WhatsApp Business, automated replies can't be sent more than 24 hours after the customer's last message. A person has to answer instead.
Settings → AI (organization owners and admins)
Open Settings (click your organization name at the bottom-left of the sidebar) and choose AI. Here you manage AI for every workspace:
- Included AI credits: each workspace's plan, and how many credits it has used this month. Calls made with your own key never use credits.
- Your provider keys: click Add key, choose the Provider, paste the API key, pick a Model and click Test connection. Tick Use as the organization's default key for the main one. Keys are encrypted; your provider bills you directly.
- Monthly budget: set a Monthly token limit and a Monthly cost limit (USD) (0 = unlimited), plus optional Token caps per AI feature. Click Save budget. AI stops with a clear message once a limit is reached.
- Usage & costs: AI calls, tokens and cost.
If AI stops with "AI credits used up", the workspace has used this month's included credits. Add your own key, buy more credits or wait until next month. See Troubleshooting.