How to make your AI chatbot
How to make your AI chatbot: choose one job, give the bot reliable business information, define its boundaries, connect only the actions it needs, and test it before adding it to your website. You can use a managed builder instead of coding the whole application. The hardest part is usually designing dependable answers and handoffs, rather than writing the welcome message.
Define one useful first job
Start with a specific task, such as answering pre-sales questions, collecting a quote request, or helping visitors find the right service. Decide what success looks like. A quote-request bot might succeed when it captures the service needed and an agreed follow-up method, rather than when it produces the longest conversation.
Write down what the bot must not do. For example, it should not invent availability, approve a refund, promise a discount, or disclose another customer’s details. Define the route to a person when the request falls outside its job.
A simple decision-tree bot can be enough for fixed choices. An AI chatbot is more useful when visitors ask varied questions in natural language and need answers grounded in business information. Either approach still needs a clear scope and a working fallback.
Prepare a small, dependable knowledge base
Gather approved answers about services, operating hours, service areas, policies, and contact routes. Remove old prices and contradictory documents. Prefer a short collection of current information to an indiscriminate upload of everything the business has ever written.
Keep private records out of a public information bot. Separate general service facts from account-specific information that requires authentication. Name someone responsible for updating each important source when the business changes.
A knowledge base supplies information for answers; it does not guarantee accuracy. ElevenLabs’ documentation explains how documents can provide domain-specific context and how retrieval can make relevant content available to an agent. This is a useful design principle regardless of your chosen builder. See ElevenLabs’ knowledge-base guide.
In VAW, the separate knowledge-base crawler supports up to 500 pages, respects robots.txt and can automatically re-sync website sources. That is distinct from the builder’s homepage-plus-nine-page starting read. You can also upload files or paste text. Use that capacity selectively: start with the pages relevant to the bot’s job. Hybrid search is designed to retrieve exact codes, SKUs and prices, while full-text mode can suit a small FAQ or menu. Test both ordinary wording and a specific product code in the knowledge-base test panel, then review missed questions. For configured API actions, the visual builder includes a Test tab, Health tab, confirmation before changing data and a per-call “What ran” log; connection permissions and successful end-to-end tests still matter.
Write instructions the bot can actually follow
Specify the role, audience, tone, approved sources, and next step. Keep the instruction concrete: answer from the supplied business information, ask one necessary clarification at a time, and acknowledge when the information is missing.
For a home-services example, instruct the bot to identify the requested service and location before suggesting the next step. If a visitor asks for a price that is not in the approved material, the bot should explain that a quote is needed. It should not infer a price from another business or present an estimate as an agreed charge.
Give the bot examples of a good answer, an unsupported question, and a request that needs a human. Make the fallback useful by explaining what happens next and which contact route is available.
Connect actions only after answers work
Answering a question and changing another system are different responsibilities. Start with an informational pilot. Add lead capture, scheduling, or CRM updates only when the corresponding integration is configured and tested.
For every action, define required information, permission checks, success confirmation, and failure behavior. A bot should say an appointment is booked only after the booking system confirms it. If the tool times out, it should acknowledge the uncertainty rather than claim success or repeatedly create appointments.
With VAW, review the CRM integration workflow for connecting conversation information to client records. Test the exact fields and permissions in your selected setup. Do not assume that all CRM connections expose the same actions.
Test with a repeatable question set
Create a checklist of realistic visitor questions before launch. Include ordinary requests, misspellings, vague questions, unavailable services, missing information, attempts to obtain private data, and instructions that conflict with the bot’s rules.
Record the expected behavior for each test. Review factual accuracy, helpfulness, unnecessary data collection, and whether the correct handoff occurs. Repeat tests after changing instructions or knowledge. Automated evaluation can support this process, but manual review is still valuable for ambiguous cases; see ElevenLabs’ agent-testing documentation.
Test the website experience on desktop and mobile, including opening and closing the widget, entering text, reading long answers, and reaching the human contact route. Check failure states as well as successful conversations.
Launch narrowly and maintain it
Begin on a limited set of pages with a clearly described purpose. Tell visitors they are interacting with an AI assistant. Provide appropriate privacy information and avoid requesting more personal information than the task needs.
Review early conversations for unanswered questions and misleading answers. Update the underlying information instead of adding increasingly complicated instructions around incorrect source material. Track useful completed tasks, handoffs, and unresolved questions alongside conversation volume.
A chatbot is ready to grow when it handles its initial job reliably and someone owns its maintenance. Branding and resale can follow a dependable working service; they do not replace the work of making the bot useful.