---
title: "AI Customer Service Agent: Support Chatbot With Handoff"
description: "Create an AI customer service agent that answers from your pages, looks up live orders and hands upset customers to your team. Free to set up and test."
canonical: https://luckpanda.app/chatbot/ai-customer-service-agent
updated: 2026-10-14
---

# Help angry and confused customers, even at 2 am

Canonical URL: https://luckpanda.app/chatbot/ai-customer-service-agent
Language: en
Updated October 14, 2026

Create a customer service agent for your website. It answers from your help pages and your live order data, says plainly when it does not know, and hands the chat to your team when a customer is upset or asks for a person.

## Customers do not hate help. They hate being stuck.

Waiting, repeating the order number, and an answer that does not match their account. Those three turn a question into a complaint.

- **87%**: of customers say access to a human agent is essential when companies use generative AI in customer service.. Source: [Gartner press release, 2 September 2026 (survey of 3,566 customers, February and March 2026)](https://www.gartner.com/en/newsroom/press-releases/2026-09-02-gartner-finds-only-27-percent-of-customers-would-try-a-chatbot-again-after-a-negative-experience) (retrieved October 14, 2026)
- **27%**: of customers would try a chatbot again after a negative experience.. Source: [Gartner press release, 2 September 2026 (survey of 3,566 customers, February and March 2026)](https://www.gartner.com/en/newsroom/press-releases/2026-09-02-gartner-finds-only-27-percent-of-customers-would-try-a-chatbot-again-after-a-negative-experience) (retrieved October 14, 2026)

**What the agent does**
- Replies in seconds, at any hour, in the customer's language
- Looks the order up in your own system instead of asking for it twice
- Hands the chat to a person, with the whole conversation, when it should

**What makes customers angry**
- Waiting until the morning for a one-line answer
- Typing the same order number to a bot and then to a person
- A bot that loops, or will not let them reach a person

## What happened in that chat, step by step

The chat at the top of this page uses settings every LuckPanda project has. Nothing in it is a special feature for the demo.

1. **It looked the order up.** A live connection to the shop's order system answered with the real status. The agent did not guess a date.
2. **It apologised once, then gave facts.** One short apology, then what happened and what happens next. No paragraph of sorry.
3. **It offered a person.** The customer could have stopped there. They asked for more, so the agent passed the chat on.
4. **It did not promise a refund.** "Refund" is one of the shop's handoff topics, so that request went straight to a person.

> The agent does not detect anger or read tone. It does not need to: a fast answer from your real data calms most people, and the words you choose, such as refund, complaint or cancel, always reach a person.

## A person is always within reach

Customers can always get to your team. Your team can step into any chat with one click.

1. **They ask, they get a person.** With the customer support goal, a customer who asks for a person is handed over. No category quiz first.
2. **Topics that always go to you.** List words such as refund, complaint or cancel. When a customer uses one, the chat goes to a person.
3. **Handoff after a few replies.** Set a number of bot replies. If the customer is still asking after that, a person takes over.
4. **Honest about your hours.** Outside your business hours it still answers. When it hands off, it says when your team will reply.

**Your team**
- Chats waiting for a person are in one list, across every project
- A browser alert with a Take over button, and an email if you want one
- Take over in one click. The bot stays quiet while you have the chat
- Reply as your business, then hand the chat back to the bot

**The customer**
- Sees "Priya joined the chat", with your teammate's name, when a person takes over
- Sees your teammate's own name, without the AI label
- Never repeats the story: your teammate reads the whole chat

## It answers from your pages and your live data, or says it does not know

- Before it answers about prices, policies or features, it searches the pages, documents, text and FAQs you gave it.
- Its rules forbid inventing features, prices, links or policies. When your sources have nothing, it says so and offers a person.
- Live connections fetch current data while it answers: order status, stock, bookings, account details. Use any HTTP API or an MCP server.
- Connections are read-only by default. Credentials are encrypted and never shown to the model or written to logs.
- Under each reply in a saved chat you see what it looked up, so you can fix the page it misread.

No chatbot can promise it will never be wrong. Ask it twenty questions your customers really ask before you go live, and read the chats afterwards. A wrong answer usually means a page is missing, old or unclear.

## It never promises a refund. A person decides.

An AI that says "your refund is done" when it is not does more damage than a slow reply. So the agent does not do it.

**What the agent does**
- States your written policy, from your own page
- Passes refund, cancellation and complaint requests to a person
- Looks things up through read-only connections

**What it does not do**
- Refund, cancel, change an order or take a payment
- Make up an exception, a discount or a deadline
- Ask for a card number or a password

One honest limit: if you untick Read-only on a connection, the bot may send changes to that system, and there is no approval step for each call yet. Keep anything that touches money read-only. In-app support for signed-in users is stricter: any change becomes a confirm card that the person must press.

## Send people to the exact page, not to the FAQ

A confused customer does not need a link to your help centre. They need the one page, and one line on what to do there.

*Sample chat on a fictional booking app*

Visitor: i paid but it says my booking is pending?? i dont get it

Chatbot: Pending means the studio still has to accept it. They usually do within two hours, and you get an email when they do. You can check the status on your bookings page. _(From the help article: Booking statuses)_ _(Button: Your bookings)_

Visitor: and if they say no?

Chatbot: Then the payment is released back to your card automatically, as our cancellation page explains. Do you want me to pass this to a person to check this booking now? _(From the page: Cancellations and payments)_

- Buttons under a reply open a page on your site, never a made-up address.
- Deep-link templates, such as /help/{topic} or /orders/{order_id}, let it build the exact address from the customer's words.
- Greetings by page: a different greeting and suggested questions on your pricing, checkout or help pages.
- In your signed-in app, it can see the customer's own account through your MCP server (in-app support).

## One flat plan, plus credits you can see

No price per resolution. You pay $4.99 (₹199 in India) a month to go live, and AI replies use credits from one balance that you top up when you choose.

| What a conversation costs from credits | Tokens | At the list rate | On the largest pack |
| --- | --- | --- | --- |
| Our planning figure (cautious) | about 60,000 | $0.30 (₹26.97 in India) | $0.15 (₹13.50 in India) |
| Expected with prompt caching (being measured) | about 10,000 | $0.05 (₹4.63 in India) | about half of that |

The planning figure is the average reply measured in production on 12 October 2026 (about 6,900 tokens) times eight replies. Since 13 October, each bot reuses a cached copy of its instructions, and we expect a typical conversation to use about 10,000 tokens. We will replace that line with measured numbers. The Usage & credits page shows the real figure for every chat. List rate: $5 per 1,000,000 tokens.

- A hard cap: credits are bought in advance and never expire. At zero the bot stops calling the AI model. You are never billed for more.
- At zero, customers are not left hanging: they see a short apology, are asked for their email, and the chat waits for a person.
- A banner and an email when the balance falls below 20%, and again at zero.
- Every business gets 10,000 tokens free for a first test. Packs start at $10 (₹899 in India).
- Replies your team writes use no tokens.

| How others charge | Pricing model | Published price |
| --- | --- | --- |
| Intercom Fin | Per outcome (a resolution is one kind) | 99 US cents per outcome |
| Zendesk AI agents | Per automated resolution | From 1 US dollar per automated resolution |
| Tidio Lyro | Per conversation | From 50 US cents per conversation |
| LuckPanda Chatbot | Flat plan plus prepaid credits | $4.99 (₹199 in India) a month, credits from $10 (₹899 in India) |

Prices as each vendor published them on 14 October 2026: fin.ai/pricing, zendesk.com/service/ai/chatbots-customer-service, tidio.com/lyro. Each vendor counts a resolution or a conversation its own way. LuckPanda publishes this page.

**See every price in one place** [Chatbot pricing](https://luckpanda.app/chatbot/pricing)

## Your data stays yours, and you can leave

- We do not train or fine-tune any model on your content, your documents or your customers' chats.
- Credentials for live connections are encrypted and never shown again. Visitor network addresses are kept only as a one-way hash.
- The website reader follows robots.txt and never logs in.
- Cancel any time. The chat stops on your website, and your projects, sources and chats stay in your account.
- No fee to remove our name: "Powered by LuckPanda" is a switch in the widget settings.

- [Security and data handling](https://luckpanda.app/chatbot/security): What it reads, where it is stored, and what it does not do.

## Free tools for support teams

Use them today, with or without the agent. No account needed.

- [De-escalation script generator](https://luckpanda.app/chatbot/tools/de-escalation-script-generator): Calm, specific replies for an angry customer.
- [Empathy statements library](https://luckpanda.app/chatbot/tools/empathy-statements-library): Phrases by situation, plus a tab that adds empathy to your draft.
- [Canned responses library](https://luckpanda.app/chatbot/tools/canned-responses-library): Saved replies for the questions you answer every day.
- [Support ticket response templates](https://luckpanda.app/chatbot/tools/support-ticket-response-templates): Replies for common ticket types.
- [Refund policy generator](https://luckpanda.app/chatbot/tools/refund-policy-generator): Write the policy your agent answers from.
- [Return policy generator](https://luckpanda.app/chatbot/tools/return-policy-generator): A clear return policy, with a chatbot FAQ.
- [Chatbot readiness checker](https://luckpanda.app/chatbot/tools/website-chat-readiness-checker): Can a chatbot read your website?

## Pages for your business and your comparison

- [Live order and account data](https://luckpanda.app/chatbot/use-cases/live-order-and-account-data): Connect your order system for "where is my order".
- [Support chatbot from your website](https://luckpanda.app/chatbot/use-cases/customer-support): Answers from the pages you already wrote.
- [After-hours enquiries](https://luckpanda.app/chatbot/use-cases/after-hours-enquiries): What happens when your team is offline.
- [Help centre and FAQ](https://luckpanda.app/chatbot/use-cases/help-centre-and-faq): Turn your help pages into answers.
- [Multilingual support](https://luckpanda.app/chatbot/use-cases/multilingual-support): Customers write in their language.
- [D2C and ecommerce](https://luckpanda.app/chatbot/for/d2c-ecommerce): Orders, returns and delivery questions.
- [SaaS](https://luckpanda.app/chatbot/for/saas): Docs, billing and account questions.
- [LuckPanda vs Intercom Fin](https://luckpanda.app/chatbot/vs/intercom-fin): Flat price or a price per outcome.
- [LuckPanda vs Tidio](https://luckpanda.app/chatbot/vs/tidio): Dated, sourced comparison.
- [AI sales agent](https://luckpanda.app/chatbot/ai-sales-agent): The same agent, set up to convert visitors into customers.

**See it answer from your own site** Paste your website address. You will see the agent learn your pages and can ask it anything. No account and no card.

## Questions people ask

### Will it make things up?

It searches your pages, documents and FAQs before it answers, and its rules forbid inventing prices, policies or features. When your sources have nothing, it says so and offers a person. No chatbot can promise it is never wrong, so test it with real questions first. Under each reply in a saved chat you can see what it looked up.

### What if a customer just wants to talk to a human?

They ask, and the chat goes to a person. You can also list topics that always go to your team, such as refund or complaint, and hand off after a set number of bot replies. The chat shows as waiting for a person, and you can take it over in one click.

### Can it look up an order or an account?

Yes, if you connect it. Add an HTTP API with a URL template such as /orders/{order_id}, or a remote MCP server, and the agent fetches live data while it answers. Connections are read-only by default. Without one, it only knows what is in your content.

### Will it promise a refund or cancel something on its own?

No. It states your written policy and passes the request to a person. Out of the box it can only look things up. If you ever untick Read-only on a connection, it could send changes, and there is no per-call approval step yet, so keep anything involving money read-only.

### Does it detect angry customers?

No. It does not read tone or measure sentiment. It answers the question with your real data, and the words you choose, such as refund, complaint or cancel, always hand the chat to a person.

### What happens at night or on weekends?

It keeps answering from your pages and live data. Set your business hours, and when it hands a chat over outside them, it tells the customer when your team will reply. You get an email or a browser alert, as you choose.

### Do customers have to repeat themselves when a person takes over?

No. Your teammate sees the whole chat, what the bot looked up and the page the customer was on, and replies in the same chat.

### How much does it cost per conversation?

Going live is $4.99 (₹199 in India) a month. AI replies use credits: our cautious planning figure is $0.30 (₹26.97 in India) per conversation at the list rate, and $0.15 (₹13.50 in India) on the largest pack. With prompt caching, our expected figure is $0.05 (₹4.63 in India) per conversation; we will publish measured figures. Your real cost per chat is on the Usage & credits page.

### What happens when the credits run out?

The bot stops calling the AI model, so you are never billed for more than you bought. Customers see a short apology, are asked for their email, and the chat waits for a person. You get an email when the balance falls below 20% and again at zero. Credits never expire.

### Is my data used to train AI models?

No. We do not train or fine-tune any model on your content or your customers' chats. The relevant text of a conversation goes to our model provider under a commercial API agreement that does not train on it. The security page lists what we store and where.

### Does it work on WhatsApp or Instagram?

Not today. It answers on your website, and inside your own signed-in app. If your customers write to you on WhatsApp, the website agent can answer the same questions before they reach your phone.

### Which languages does it speak?

It replies in the language the customer writes in, from your content in whatever language it is. Test it in the languages your customers use before you go live.

### Can I cancel any time?

Yes. Cancelling stops the next renewal. When the subscription ends the chat stops on your website, and your projects, sources and chats stay in your account. There is no fee to remove "Powered by LuckPanda": it is a switch in the widget settings.

Try it on your website: https://luckpanda.app/chatbot/try