Skip to content

Insights / The practical edit

AI chatbots for business websites: what should yours actually do?

Plan an AI chatbot around real customer questions. Define its scope, prepare reliable answers, design a human handover and test it before adding it to your website.

Back to all insights
AI chatbots for business websites: what should yours actually do?

A visitor asks whether you can finish a project before an event. Your chatbot gives an enthusiastic yes. Your team has never seen the brief. That is the sort of problem a chatbot plan should prevent before anyone chooses a tool.

Give it one useful job first

An AI chatbot can help people find information or explain an answer in everyday language. It should not start with permission to speak for every part of the business. Choose a narrow first job, such as explaining service options, answering common delivery questions or collecting the details needed for a follow-up.

Review a sample of real enquiries, removing personal information before sharing them outside the business. Group the repeated questions. Which have stable, approved answers? Which depend on availability, account details or a person’s judgement? The first group is a more practical starting point.

Prepare the answers before the interface

A good-looking chat window will not repair contradictory source material. Put the approved information in one maintained place: service descriptions, policies, contact routes and any clear limits. Assign an owner who can update it when the business changes.

Keep a short record for each topic: the source, the person responsible and the last review date. For pricing that depends on a brief, the approved answer should explain the factors and offer a next step. It should not invite the model to invent an estimate.

Design the handover as carefully as the answer

Tell visitors they are talking to an automated assistant. Give them a visible way to contact a person, and say when a reply can be expected. A handover should carry the useful context so the customer does not have to start the conversation again.

Here is an illustrative exchange for a service business:

Customer Can you launch our store in three weeks?

Assistant The team needs to check the scope before confirming a launch date. How many products do you expect to sell, and is there an existing store? I can include those details in your enquiry.

The answer is useful because it moves the conversation forward without making a commitment the business has not approved.

Keep permissions narrow

A public information assistant usually does not need access to private customer records. If you later connect accounts or internal systems, enforce access checks in those systems and limit which actions the assistant can take.

Messages and retrieved documents can contain instructions intended to manipulate a model. OWASP’s prompt injection guidance explains why instructions alone are not a complete defence. Treat outside content as untrusted, check outputs and keep consequential actions behind appropriate controls.

Test the difficult questions

Build a small test set before launch. Include questions with clear answers, vague requests, misspellings, unsupported languages, outdated information and requests outside the agreed scope. Add questions that ask it to reveal private information or ignore its rules.

For each test, define what a useful result looks like. Sometimes the correct result is an answer; sometimes it is an honest limitation and a handover. Also test keyboard access, the mobile layout and whether people can close the chat easily.

Measure useful outcomes

Review answer accuracy, unresolved questions, successful handovers and customer feedback. Do not treat fewer conversations with staff as success if customers simply give up. Keep a sample review routine and a way to disable the assistant if something goes wrong.

If the questions are few and predictable, improve the page content or use a simpler guided form first. A chatbot earns its place when it makes finding help easier. For the work that happens after an enquiry arrives, see our business automation guide.

Share this perspective LinkedIn ↗ Facebook ↗ WhatsApp ↗

Have a project in mind?

Let’s talk it through

Ideas worth exploring

A little perspective.

All insights