luckpanda
Start free
English
Support

Help angry and confused customers, even at 2 am

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.

Build my chatbot
  • Free to set up and test
  • Answers only from your pages and live data
  • A person can take over any chat

Free to set up and test. Going live costs $4.99 (₹199 in India) a month.Billing is in US dollars, or Indian rupees for a business in India.

Claire AIAI assistant

Sample chat on a fictional online shop, at 11:40 pm

WHERE IS MY ORDER #4471. This is the third time I am asking.
I'm sorry you have had to ask three times. Order 4471 left our warehouse on Monday. The courier shows it out for delivery tomorrow, before 6 pm. Would you like a person from our team to follow it up as well?
Looked up order 4471 (live connection: Orders API)
yes. and if it does not come tomorrow I want a refund
Understood. Refunds are decided by our team, so I have passed this chat to them with your order number. Nobody will ask you to start again. They are back at 9 am and will reply in this chat.
Handed to a person: "refund" is a handoff topic
Team offline: said when they are back
What customers say

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) · 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) · 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
An angry message, handled

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 upA 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 factsOne short apology, then what happened and what happens next. No paragraph of sorry.
  3. It offered a personThe 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.

Human handoff

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 personWith the customer support goal, a customer who asks for a person is handed over. No category quiz first.
  2. Topics that always go to youList words such as refund, complaint or cancel. When a customer uses one, the chat goes to a person.
  3. Handoff after a few repliesSet a number of bot replies. If the customer is still asking after that, a person takes over.
  4. Honest about your hoursOutside 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
The Chatbot module in LuckPanda for Fernhill Studio on a phone: the Chats inbox of Fernhill Studio showing the Studio website project: a project filter, the Projects button, search, the Needs you, Leads, Assistant and All views, and visitors such as Yuki, Lucia and Pieter, each with the last message, a status seal such as Lead or Assistant and the project.
Grounded answers

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.

Refunds and actions

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.

Confused customers

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

i paid but it says my booking is pending?? i dont get it
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
and if they say no?
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).
A bill you can predict

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 creditsTokensAt the list rateOn the largest pack
Our planning figure (cautious)about 60,000$0.30$0.15
Expected with prompt caching (being measured)about 10,000$0.05about 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 chargePricing modelPublished price
Intercom FinPer outcome (a resolution is one kind)99 US cents per outcome
Zendesk AI agentsPer automated resolutionFrom 1 US dollar per automated resolution
Tidio LyroPer conversationFrom 50 US cents per conversation
LuckPanda ChatbotFlat 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
Your data

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.
Free tools

Free tools for support teams

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

Go deeper

Pages for your business and your comparison

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 per conversation at the list rate, and $0.15 on the largest pack. With prompt caching, our expected figure is $0.05 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.

Keep reading

Related pages

For AI agents

Build it with your AI agent

Give an agent your site and it can create the chatbot, read your pages and test it through LuckPanda's API and MCP. It prepares the work; going live stays with you.

Use LuckPanda's MCP server (https://luckpanda.app/mcp) to build a website chatbot for me. Ask for my website address, call chatbot_create_from_website with it, wait until the pages are learned (chatbot_source_list), then run a few test questions with chatbot_test_message and show me the answers. Do not put it live: I will do that myself.

Instructions for AI agents

Needs a LuckPanda agent key (Settings › Developers) with the chatbot.manage grant. MCP endpoint: https://luckpanda.app/mcp (Streamable HTTP, bearer key).

  • chatbot_create_from_website { url }: creates a draft chatbot named after the site and starts reading it.
  • chatbot_source_list: pages learned, failed or skipped, with a reason for each.
  • chatbot_test_message: asks the bot a question as the owner. Test chats never email anyone.
  • Going live (chatbot_update with status live) is a person's action; an agent gets APPROVAL_REQUIRED.

Try it on your website

Paste your address and watch the chatbot learn your pages. No account and no card.

This page in other languages