A website AI chatbot should help a visitor complete a specific job. That might be understanding a product, checking a published policy, preparing an enquiry or reaching the right person. The right design depends on which of those outcomes your website currently makes difficult.
Before comparing interfaces or subscription plans, look at the questions that arrive through your contact form and support inbox. They reveal what visitors could not find, did not understand or could not do by themselves.
Separate four different chatbot jobs
| Job | Useful result | Required connection |
|---|---|---|
| Answer a question | A clear answer supported by current material | Approved knowledge sources |
| Guide a choice | A relevant shortlist with the limits explained | Product facts and selection rules |
| Prepare an enquiry | A concise brief for the right person | Contact routing and an agreed intake |
| Take an action | A confirmed result in the business system | Authorized, verifiable application actions |
One assistant can eventually handle several jobs. Giving the first release a primary job makes its value easier to judge. A chatbot built to answer questions should not be assessed as though it already manages the whole sales process.
Identify what the visitor already knows
A visitor on a product page has different context from someone landing on a general article. Use the page context where it is reliable: which item is open, which market the page serves and which language the visitor is using.
Do not treat page context as proof of identity. Looking at an order-help page does not authorize access to an order. When personal account information is needed, the application must establish the appropriate access before returning it.
Also decide when the assistant should stay quiet. A visible invitation to ask a question can be enough. Interrupting every visitor with a long scripted conversation adds friction before the person has expressed a need.
Make the knowledge source part of the specification
Uploading a folder is only the beginning. The assistant needs a way to distinguish a current policy from an old draft, a retail condition from a trade agreement, and an actual product attribute from promotional wording.
Choose an owner for each important source and a process for corrections. If the support team finds an inaccurate answer, they should be able to trace it to the underlying material. The same source-management questions apply to AI search for company documents.
Make the handoff worth receiving
A human handoff should preserve the customer’s question, information already supplied and the reason the assistant could not finish. The next person should not need to ask the customer to start again.
Decide what the visitor sees outside staffed hours. A clear expectation about the next response is more useful than an assistant repeatedly offering to transfer to someone who is unavailable.
I build and operate Conviro, a customer communication platform that brings web chat and messaging channels together with AI assistance and human handoff. It is one implementation route to consider when those are the jobs you need; the website’s requirements should still determine the setup.
Choose a platform or a custom integration
A hosted product can fit a familiar support workflow with standard channels. A custom integration becomes more relevant when the assistant must respect unusual permissions, retrieve proprietary business records or work inside an existing interface.
Compare the complete workflow, including the operator’s view and the data you would need to export later. My custom AI development versus no-code comparison gives you a way to make that decision without starting from a vendor list.
Define what a successful conversation looks like
Use real question examples to agree on acceptable outcomes. A correct answer, a useful clarification and a well-prepared handoff can all be successful. A conversation ending quickly is not sufficient evidence that the customer was helped.
Review unsupported claims, repeated questions and failed transfers separately. They point to different changes: better source material, a clearer interface or a repaired integration.
If the missing piece is the connection between your website, knowledge and support operation, see my assistant implementation and integration work. Send the website and a few recurring questions to define the first useful job.
Updated 5 October 2026.