Let’s talk
GUIDE

What does a website chatbot do for a small business, and when is it worth it?

A website chatbot answers visitors on the page from information the business has approved: services, opening hours, prices, how to book. When a question needs a person, or the visitor is ready to book, it collects the details staff need and passes the conversation on with a short summary. It is worth it when the same questions arrive every day, including outside office hours, and not worth it when visitors rarely write.

The value is in the handover, not in the chat window. A chatbot that answers well but drops the enquiry into an inbox nobody reads saves nothing, and one that guesses an answer it was never given does harm. In a clinic, anything about symptoms or treatment goes to a person.

Updated Sep 28, 2026

How it works

  1. Write down the approved answers

    Services, hours, prices, policies and booking rules, in plain text the business signs off. Add a short list of things it must never answer, such as medical, legal or payment decisions.

  2. Decide what it collects

    Only the fields staff need to follow up: name, what the person needs, a preferred time, and how to reach them. Nothing more.

  3. Set the handover rules

    Health questions, complaints, refunds, requests for a person and anything outside the approved text go to staff. Decide where the summary lands (an inbox, WhatsApp or the helpdesk) and who reads it.

  4. Connect booking honestly

    If the business already books online, the chatbot answers the questions that come before booking and links to that page. If it is not connected to the calendar, it takes a request that staff confirm and never promises a time.

  5. Test with messages nobody scripted

    Typos, another language, a question it should refuse, "are you a person?", and questions the approved text does not cover. Each should get a correct answer, an honest "I don't know" or a handover.

  6. Read the conversations every week

    Questions it could not answer show what is missing from the approved text. Add the answer there, not in a one-off reply.

What it costs to run

Running a chatbot costs either the model that writes the answers, billed per token by the model provider, or a helpdesk vendor's own chatbot, billed per resolved conversation on top of its seats. These are list prices from the vendors' pages; the website, hosting and staff time are not included.

ItemList priceSource
Claude Haiku 4.5 through the Anthropic API (a smaller, cheaper model)US$1 per million input tokens, US$5 per million output tokensAnthropic, Claude API Docs, PricingRead on Sep 28, 2026
Claude Sonnet 5 through the Anthropic APIUS$2 per million input tokens, US$10 per million output tokensAnthropic, Claude API Docs, PricingRead on Sep 28, 2026
Intercom Fin AI AgentFrom US$0.99 per outcome, charged once per conversation, plus Intercom seats; on another helpdesk, a monthly minimum applies (Intercom gives 50 outcomes as an example)Intercom, Intercom Pricing | Plans for every team sizeRead on Sep 28, 2026
Help Scout AI Answers (add-on to a Help Scout plan)US$0.75 per resolution (a conversation resolved without a person); a monthly spending cap can be setHelp Scout, Help Scout Pricing: Affordable Customer Support SoftwareRead on Sep 28, 2026

Third-party list prices, read on the date shown. They are not Betterlane prices.

What we learned building it

From our own fictional demos, September 2026: an English clinic on demo.betterlaneagency.com, a physio assistant called Fisio Sol and a coworking-space chat called Cedar Workspace. None is a client system.

Two passing questions prove little
On our fictional clinic demo, an opening-hours question and a price question behaved as intended, and then a different message we typed ourselves got no answer. Checking a few scripted questions does not show that arbitrary messages work.
Health questions get a holding reply, not advice
In our fictional physio assistant, a message about knee pain, a service the clinic does not offer and a billing complaint each got a short holding reply with no health advice, and a one-line summary went to the test "reception" number.
Show where the answer came from
The Cedar Workspace demo takes each answer from a written document and highlights the passage it used. A test edits the document and checks that the next answer changes. That demo uses simple pattern matching, not a language model.
Missing facts need a planned reply
The same demo asks for clarification when a question is ambiguous and says plainly when the document does not cover something, instead of filling the gap.

When it is not worth it

  • If visitors rarely write, a contact form and a quick reply from a person cost less and need no setup.
  • If most visitors already book through an online booking page that works, a chatbot adds little beyond the questions that come before booking.
  • If the answers are not written down, or depend on live information the chatbot cannot read (stock, today's availability), it will either guess or hand over almost every conversation.

Questions

Will it make up answers?

It can, if nothing stops it. Limit it to the approved text, instruct it to say when it does not know, and test questions outside that text before it goes live.

Can it book appointments?

Only if it is connected to the calendar with permission and the booking rules are clear. Otherwise it takes a request that staff confirm, or links to the booking page the business already uses.

Should visitors be told it is automated?

Yes. Say it plainly in the first message and keep a visible way to reach a person.

What happens outside office hours?

It answers what the approved text covers and records the rest with the contact details, so staff start the day with summaries instead of an empty inbox. Do not promise a reply time the team cannot keep.

Want a website chatbot that answers from your own information and hands the rest to your team?

See the service: Website chatbotsLet’s talk

All guides