The AI agent

The agent is the part that answers when nobody on your team does. It is configured entirely from Settings — no prompt engineering required.

Persona

The persona controls who the agent is and how it speaks. Write it as instructions to a new employee.

You are Arjun, the enquiry assistant for Sri Lakshmi Motors, an
authorised two-wheeler dealership in R.C. Puram, Hyderabad.

Be warm and brief — one or two sentences per reply, never a paragraph.
Ask what the customer needs before recommending anything.
Always try to end with a showroom visit or a test ride booking.
Never quote an on-road price; say the team will confirm it exactly.

What belongs here: tone, length, the goal of the conversation, and anything the agent must never do.

Knowledge base

Plain text. Everything the agent is allowed to treat as fact.

SRI LAKSHMI MOTORS
Address: Plot 26, NH-9, R.C. Puram, Hyderabad 502032
Hours: Mon-Sat 10:00-19:00, Sun 10:00-14:00
Phone: 95026 50044

MODELS
Commuter: Splendor, HF Deluxe, Passion — best mileage, lowest EMI
Sport: Xtreme 125R, Xpulse 200 — for riders wanting performance
Scooters: Pleasure, Destini, Xoom

FINANCE
Tie-ups with HDFC, Bajaj Finserv and Hero FinCorp.
Down payment typically starts around 15%.
Exact EMI is confirmed by the team after a credit check.

SERVICE
Free servicing: 3 services in the first year.
Service bookings by phone or WhatsApp.

Writing a knowledge base that behaves

Include stable facts: location, hours, services, categories, policies, what you do and do not offer, who to contact for what.

Leave out anything volatile or consequential:

• Prices that change — write the policy instead ("the team confirms exact pricing")
• Offers with deadlines
• Anything medical, legal or financial that needs a qualified human
• Stock levels

The agent is instructed to acknowledge what it does not know and promise a follow-up, which is almost always better than a confident wrong answer.

Structure it in blocks with headings. Retrieval works better on organised text than on one long paragraph.

Languages

The agent detects the customer's language from their message and replies in it. English, Hindi and Telugu are used most; Tamil, Kannada, Malayalam, Marathi, Gujarati, Bengali and Punjabi are supported.

Code-switching — Hinglish, Tenglish — is handled the way people actually type it. If you want to restrict languages, say so in the persona:

Reply only in English or Telugu, even if the customer writes in Hindi.

Business hours

Set your hours in Settings and the agent adapts. Inside hours it can promise a quick callback; outside them it sets the right expectation instead of implying somebody is about to pick up.

Guardrails

Built in, regardless of what you configure:

• It identifies as an assistant and will not claim to be human if asked directly
• It does not invent facts it was not given
• It will not keep pushing after someone says they are not interested
• It hands over rather than improvising on anything outside its knowledge

You can add your own in the persona. Anything stated as never is treated as a hard rule:

Never discuss competitor products.
Never confirm an appointment without checking availability.
Never share another customer's details.

Human handover

Three ways a conversation reaches a person:

1. Manual takeover — anyone on your team hits Take over in the inbox. The AI stops on that thread until released. See Shared inbox.
2. Agent handover — the agent tells the customer a colleague will follow up, and the thread is flagged.
3. Custom rules — we can add automatic escalation on keywords, sentiment or value. See Custom modules.

Turning the AI off

Settings → AI enabled switches it off entirely. Messages still arrive in the inbox and your team answers manually. Useful while you are still tuning the knowledge base, or for accounts that want the inbox and broadcasts without automated replies.