AI Support Agent Launch Checklist: 4 Phases, One Page

An AI support agent launch checklist is a short list of checks you tick before real customers talk to your agent. Work through four phases in order: before (decide what it answers and who owns it), build (sources, tone, handover), test (your questions, unknown questions, a real handover) and launch (one channel first, then read every early conversation).

Most AI support agents do not fail on the day they go live. They fail because nobody wrote down what the agent may answer alone, who owns it, or what happens when a customer asks for a person. This page puts the whole launch on one page. Work top to bottom and tick each line before you move on. If you are still choosing where the agent will sit, start with how to add an AI agent to your website.

What should you decide before you build an AI support agent?

Phase 1 is decisions, not software. Anthropic's guide to support agents starts the same way: outline the ideal customer interaction before you build, because it decides what the agent needs (Anthropic).

1. Write down the five questions customers ask you most. These become your first test set, and they tell you which sources the agent needs before anything else.
2. Decide what the agent may answer alone and what must go to a person. Billing errors, complaints and anything legal usually belong with a person. Write the list down; a list can be tested, a feeling cannot.
3. Name one person who owns the agent after launch. Someone has to read the conversations, fix wrong answers and say yes to the next channel. If nobody owns it, nobody improves it.
4. Collect your best sources: your website, help pages, policies, product pages. The agent answers from what you give it, so gather the pages a good support person would use.
5. Check those sources are current. The agent will repeat whatever they say, including an old price or a retired policy.

How do you build the agent?

Phase 2 turns those decisions into the agent itself. For the knowledge side in more depth, see how to train an AI chatbot on your documents.

1. Point the agent at your website and let it read every page that matters. Your site is usually the most complete and most current source you have.
2. Add anything that is not on the site: returns, delivery, opening hours. Gaps in the sources turn into "I don't know" answers, or worse, guesses.
3. Set the tone so it sounds like your brand, not a generic bot. Pick two or three words for the voice and one concrete rule, such as acknowledging a problem before giving the fix.
4. Set the handover rule: when a visitor asks for a person, a person gets the chat. Tell the agent to say what happens next, so the customer is never left waiting in silence.
5. Connect the channels you already use: website first, then the rest. Start where most of your questions already arrive. One channel that works beats three that half work.

A good handover is one the customer does not have to repeat. Our guide to the best AI chatbot with human handover covers what to look for.

How do you test an AI support agent before launch?

Phase 3 is where you find the wrong answers before your customers do. Anthropic notes that a prompt usually needs testing and refinement before it is ready for real use (Anthropic).

1. Ask it your five questions in your own words, then in a customer's words. Customers rarely use your internal names for things. The second round is the one that matters.
2. Ask it something it should not know. It should say so, not guess. Giving a model explicit permission to admit it does not know is a standard way to cut made-up answers (Anthropic).
3. Ask for a human and check a teammate really receives the conversation. Recognising when a question needs a person, and passing it on, is one of the things Anthropic suggests you measure in a support agent (Anthropic).
4. Try it on a phone as well as a laptop. Many visitors will meet the agent on a small screen. Check that the chat opens, reads well and does not hide your page.
5. Have someone outside your team break it for ten minutes. You know the right words to type. A fresh person asks the odd questions your customers will.

For a fuller test plan, with question sets and what a pass looks like, read how to test a chatbot.

How do you launch it without surprises?

Phase 4 is a slow start and a close watch. The aim of the first week is to learn, not to cover every channel.

1. Switch it on for one page or one channel first. A small start keeps any early mistake small and easy to fix.
2. Read the first real conversations the same day. The first day shows you which questions you did not expect.
3. Fix every wrong answer at its source, not with a patch. If the agent got a fact wrong, correct the page or document it learned from, so every later answer is right too.
4. Tell your team what the agent handles and when they step in. Your team should know which chats will reach them and what the agent has already said.
5. Add the next page or channel once the first one holds. Grow one step at a time, and run your test questions again after each change.

If answers keep going wrong after launch, the cause is usually one of a few common mistakes, listed in why AI chatbots fail.

What does each phase need, at a glance?

How do you run this checklist in Dante AI?

Each phase maps to a place in Dante AI:

• Sources. Add your website and Dante AI crawls the page and follows internal links. Add what is not on the site as file uploads, pasted text or question and answer pairs (Knowledge sources).
• Tone and rules. Write them in the Behavior field, on the Sandbox page while you build and under Publish, then Model, after launch (Agent behaviour).
• Test privately. The Sandbox is a private preview on every plan, including Free. Save changes applies your edits to the live agent straight away (Sandbox).
• Handover. With Human Handover on (Advanced and Pro plans), visitors get a Talk to a human option, the chat lands in your Inbox, and the AI stops replying once a teammate takes it (Inbox).
• Website first. Paste one embed snippet before the closing body tag, as a floating launcher or a full panel, and add your domain to Allowed domains (Embedding your agent).
• Read real conversations. Chat history shows every real visitor conversation on Starter and up; on Free you can read your own Sandbox test chats (Chat history).

The Free plan costs $0 and includes 100 credits a month, enough to work through phases 1 to 3 on your own website, apart from the handover test (Human Handover is on Advanced and Pro). Get started free at dante-ai.com.

Sources

• Anthropic, Customer support agent (outline the ideal interaction first, escalation to a person as a measure, test before real use), read 4 November 2026.
• Anthropic, Reduce hallucinations (let the model say it does not know; restrict it to provided documents), read 4 November 2026.
• Dante AI Help Center, Knowledge sources, Agent behaviour, Sandbox, Inbox, Embedding your agent and Chat history, read 4 November 2026.