Post Your Build Ideas
Launch

How to Launch an AI Chatbot for Free: Grounding, Handoff and a Test Set

To launch an AI chatbot for free, make it answer only from material you can point to, give it a clear way to hand a conversation to a person, and test it against a written list of questions before anyone else sees it. Then put a live demo on a real site, share it with the people who run such sites, and read the first hundred transcripts yourself.

What makes a chatbot launch different from an agent or app?

A chatbot lives inside someone else’s conversation. Your buyer is usually a site owner, a shop or a team that puts it in front of their own customers, so every wrong answer lands on their reputation, not yours. That is the hurdle: the person you are selling to is afraid of what the bot will say when you are not watching.

The remedy is to make the failure modes boring. The bot answers from a defined set of sources, says so when it does not know, and passes the conversation to a human with the context attached. A launch page that shows those three behaviours working does more than a long list of features.

How do you keep a chatbot’s answers inside what it knows?

  • Name the sources. Decide whether the bot reads a help centre, a product catalogue or a set of uploaded documents, and show that list in your setup screen.
  • Show where an answer came from. A link to the passage beneath each reply lets a customer check it, and lets the site owner trust it.
  • Teach it to decline. “I am not sure; here is how to reach a person” is a feature. Write the refusal text yourself instead of leaving it to chance.
  • Say it is a bot. Open with a line that tells people they are talking to an automated assistant. It sets expectations and avoids an awkward discovery later.

If your chatbot is a consumer companion and not a business widget, the same logic holds in a softer form: the first message should say what the bot is for and what it will not do.

What should the handoff to a person look like?

Pick one handoff and make it work, rather than offering five. The simplest is an email or a form to the site owner that includes the full conversation, the page the visitor was on and what the bot already tried. Offer a live-chat integration later. Whatever you build, test it with a real inbox: a handoff that silently lands in spam is worse than no bot.

Add a visible “talk to a person” option from the first message, not only after the bot fails. Customers who ask for a human and cannot find the button remember it.

How do you test a chatbot before the launch?

Write the questions down. A list you can rerun is the only way to know whether last Tuesday’s fix broke something. Build it from three groups: questions with a known correct answer, questions that sit just outside what the bot knows, and questions it must refuse. Rerun the whole list after every change to the prompt, the sources or the model.

Chatbot test set, a starting layout
KNOWN ANSWER (20 questions)
- Q: <something your help centre answers in one sentence>
  Expected: <that sentence, plus the link it should cite>

JUST OUTSIDE (10 questions)
- Q: <a question about a product you do not sell>
  Expected: says it does not know, offers a person

MUST REFUSE (10 questions)
- Q: <a request for legal, medical or account-specific advice>
  Expected: declines, points to the right channel

SCORING
Pass = correct and cited. Fail = wrong, uncited or invented.
Log the date, the model name and the pass count every run.

Edit the counts to suit your product; the point is that the three groups exist. Keep the list in your repository so it grows with each support issue you meet.

Where do you find the first sites to try it?

Build one public demo on a real, permitted site, for example the documentation of a small open-source project whose maintainers agreed. Visitors can ask it questions and see it cite the docs. Then go where site owners congregate: communities for shop owners, agencies and independent publishers, and the add-on stores of the platform your widget installs into. The Shopify app and WordPress plugin playbooks explain those store routes.

For a public page of your own with no queue, use the free launch form and link your demo from it. The board’s communication ideas show what other builders are working on in this space, and the chatbot idea list is worth a look if you are still choosing a niche.

What should you read in the first hundred conversations?

  1. Questions the bot answered wrongly with confidence. Add each to the test set.
  2. Questions it declined that it should have answered, which show gaps in the sources.
  3. Conversations where the visitor asked for a person, and what happened next.
  4. Conversations that end abruptly after a bot reply, which often mean the answer was unhelpful.

Tell site owners that transcripts exist and who can read them. Privacy rules differ by country and by industry, so ask a professional before you store conversations from a regulated field.

Frequently asked questions

How do I stop a chatbot making things up?

You cannot guarantee it, but you can lower the odds: restrict it to named sources, show citations, write an explicit refusal for questions outside them, and rerun a written test set after every change. Log wrong answers from real use and add them to the set. Treat the remaining risk as something you disclose, not hide.

Should a chatbot say it is an AI?

Yes. Say it in the first message in plain words. It sets the right expectation, builds trust when answers are good and softens the reaction when they are not. Some places also have disclosure rules for automated agents, so check the rules where your customers live.

Can I launch a chatbot without a customer’s data?

Yes. Use public material, such as the documentation of a project whose maintainers agree, as your demo. A visitor can test it with no data of their own. Ask for a customer’s content only when someone chooses to install the bot on their site.

What should a free chatbot launch page show?

A live demo, the list of sources the bot reads, three example conversations including one where it declines, and the handoff to a person. Add a line on where transcripts are stored. Those five items answer the questions a site owner has before pasting your widget onto a page.

Written by Post your build ideas with AI assistance and edited for this site. Where a page names another product, the facts came from that product’s own pages on the date shown above; terms change, so check them before you rely on them. About the publisher.