AI Agent
An assistant trained on your own business that answers customers, with handoff to a human.
The AI Agent answers customers on your behalf, using information you give it about your business. It is not a general chatbot — it knows what you tell it and hands over when it does not know.
Setting it up
Open AI Agent and fill in what it should know.
Who it is
| Field | What it is for |
|---|---|
| Agent name | What it calls itself. Default is "Chat agent" |
| Greeting message | The first thing a new customer sees, up to 1,000 characters |
| Business name and type | So it describes you correctly |
| Currency | For anything it says about prices |
What it knows
This is the part that decides whether it is useful:
- Business description — what you do, in plain words
- Products and services — what you sell, with prices if you want it quoting them
- Policies — returns, delivery, warranty, payment
- FAQ — the questions you answer every day
- Ground rules — what it must never do. Never promise a delivery date, never quote a discount, never discuss a competitor
Ground rules matter more than people expect. An agent that invents a refund policy costs you more than one that says "let me check with the team".
Importing from your website
Point it at your website URL and it pulls content in as a starting point. Read what it imported before publishing — a website says things in marketing language that make poor answers, and it will not know what is out of date.
How it speaks
| Setting | Options |
|---|---|
| Tone | Friendly, professional, concise, playful |
| Language | Auto, English, Hindi, Hinglish |
| Reply length | Short, medium, long |
| Maximum reply | Characters per reply, default 1,000 |
Auto language matches whatever the customer wrote in — a question in Hinglish gets a Hinglish answer. For most Indian businesses this is the right setting.
Handing over to a human
Three ways a conversation reaches a person:
Handoff keywords. Words that bypass the agent entirely — "agent", "human", "complaint". These never reach the AI, so they cost nothing and are never misread.
The fallback person. Who the conversation goes to on handoff. Set this; an unassigned handoff sits in a shared inbox waiting for someone to notice.
Turning the bot off for one conversation. From the inbox, any team member can switch the agent off for that customer and take over. The agent stays on for everybody else.
The handoff message is what the customer sees when this happens. Make it a promise you keep: "One of our team will reply shortly" is fine if someone actually will.
Publishing
Settings are saved as a draft and go live when published. Live conversations keep using the published version until then, so editing is safe at any time.
Always use the test panel before publishing. Ask it the five questions your customers actually ask and read the answers properly.
Automatic replies
The agent only answers on its own when automatic replies are switched on. Off, you can still test it and use it inside an automation's AI reply node, but it will not answer customers by itself.
Starting with it off, testing for a few days inside automations, then switching it on, is the sane order.
What it costs
Every reply is a request to a language model, so the agent costs money per answer. The model used is configurable; a smaller model is cheaper and usually enough for FAQ-style questions.
Handoff keywords are free — they never reach the model. Putting your common escalation words there saves money and gives a better answer.
What it cannot do
It does not know your orders. It answers from the information you wrote, not from your database. "Where is my order" gets a generic answer unless you hand off.
It does not learn from conversations. If it answers something wrong, fix the business information — it will not correct itself.
It cannot message first. It replies inside the 24-hour window like any free-text message. Starting a conversation needs a template.
Getting good answers
Write the FAQ as questions and answers, in the customer's words — "do you deliver to Pune" rather than "serviceable pincodes".
Be specific about prices. "Starts at ₹499" is answerable. "Competitive pricing" is not.
Test with real messages. Take ten genuine customer messages from your inbox and run them through. Wrong answers point at missing information.
Re-read it monthly. The commonest failure is an agent confidently quoting a policy that changed in March.
Next
- Automations — using the agent inside a flow
- Inbox — taking over from the agent