Website chat assistants that actually know your business
By Keith Hamilton · · 4 min read
A website chat assistant answers customers' questions in their own words, using your own information rather than the AI's general knowledge. That is a big step on from older chatbots, which worked from fixed menus and keyword matching and could only answer the questions someone had anticipated. This article explains what makes an AI chatbot for a small business website genuinely useful, what it should not do, and how to keep it accurate.
What does an AI chatbot trained on your website actually do?
An AI chatbot trained on your website answers from the information you have given it, not from general knowledge. A general AI model knows a great deal about the world and nothing reliable about your prices, delivery areas or opening hours. Left to answer from general knowledge, it makes things up that sound right.
A chat assistant that knows your business works differently. When a customer asks a question, it first looks in your website pages, service descriptions and policies, and answers from those. If the answer is not there, it says so and offers another route.
Two features to look for:
- Answers from your sources: the assistant draws on your content, not its general knowledge
- Links to the source: it points the customer to the page its answer came from, so they can check
The same approach works for staff. Our article on internal knowledge assistants covers the internal version.
What information does a chat assistant need?
A chat assistant needs the same information your team uses to answer customers, because it is only as good as the information behind it. Before you start, gather:
- Your website pages, including services, prices where you publish them and frequently asked questions
- Delivery, returns, cancellation and complaints policies
- Opening hours, locations and service areas
- Any guidance your team already uses to answer common questions
Read through it first. If two pages disagree, or a policy is out of date, the assistant will repeat the problem. Tidying your content is one of the most valuable parts of the project.
What should a website chat assistant not do?
A website chat assistant must not go beyond what your policies say or guess when it does not know. Agree firm limits, and test that they hold:
- It must not promise prices, discounts, refunds or delivery dates beyond what your policies say
- It must not give legal, medical or financial advice
- It must not guess when it does not know
- It must not collect more personal data than it needs, such as payment details in a chat window
Can a chat assistant connect to your systems?
Yes, through a secure connection to the relevant system and a way of confirming who the customer is. An assistant that only reads your website answers general questions. To answer personal ones, such as "where is my order?" or "can I move my appointment?", it needs those connections.
Start with general questions, then add these connections once the basics are working well.
How does chatbot handover to a human work?
Chatbot handover to a human depends on your opening hours, and customers must always be able to reach a person:
- During working hours: pass the chat to a colleague, with the conversation so far visible so the customer does not have to repeat themselves
- Out of hours: take the customer's details and question, tell them when they will hear back, and send the conversation to your team's inbox
If those messages land in a shared inbox, AI inbox and ticket triage sorts them by topic and urgency.
How do you stay open with customers and protect their data?
Tell customers they are chatting with an AI assistant, and treat chat transcripts as personal data under UK GDPR. Decide how long you keep transcripts and who can see them, and update your privacy information. Check your supplier's terms on whether your conversations are used to train their models.
How do you keep a chat assistant accurate?
Give one person ownership of the content it relies on, because prices change, policies are updated and new questions appear. A chat assistant is not a one-off project. That owner should:
- Update the assistant whenever you change a policy or a published price
- Read a sample of conversations every week, especially early on
- Note the questions it could not answer and add the missing information
- Look for answers that were technically correct but unhelpful
How do you test a chat assistant before launch?
Test it with 30 to 50 real questions collected from your inbox, phone notes and reviews. Ask the assistant each one and check the answers with someone who knows the business well. Include awkward ones: complaints, vague questions, questions about competitors and attempts to get it to promise something it should not.
How do you measure AI customer service chat?
Record your starting point before launch, then track:
- The share of chats answered without a hand-over
- The reasons chats were passed to a person
- Questions it could not answer, grouped by topic
- Calls and emails on the same topics, to see whether they fall
- A sample of conversations reviewed for accuracy and tone
Where Forwardcycle fits
Forwardcycle helps UK businesses build chat assistants that answer from their own information and hand over to people cleanly. The same principle underpins a proof of concept for a district council, where answers come from the council's own website rather than the AI's general knowledge. No results are claimed yet. Find out more about our customer experience services or take the free AI Readiness Assessment.