An AI chatbot can be useful when visitors have recurring questions and the business has reliable information to answer them. It can help someone find a service, understand a process or reach the right team. It can also become an expensive extra layer between a customer and a simple answer.
The sensible starting point is the problem. If visitors cannot find your opening hours, first make the opening hours easy to find. If a product range generates detailed questions that do not fit a short page, a conversational route may deserve a closer look.
Find out what people actually ask
Review a sensible sample of enquiries, support requests and calls, using information you are authorised to access. Group the questions by purpose rather than copying customer details into a new tool. You might find that most concern delivery, compatibility, appointment availability or the same few product differences.
Then ask which answers are stable and which depend on live information. A general description of your process is different from confirming that an engineer can arrive tomorrow. The second answer requires access to a dependable scheduling system and rules about what the assistant may promise.
This exercise can improve your website content and page journeys even if you decide against a chatbot. Better service pages, clearer product information and a shorter enquiry form may address the same issue more directly.
Choose a useful first job
A focused assistant is easier to assess than one told to help with everything. For example, a supplier could start with navigation: ask what the visitor is trying to achieve and point them towards relevant product information. It could link to a specification page without inventing a personalised technical recommendation.
A service business might use an assistant to explain the information needed for an enquiry and then hand over to its team. That is a different scope from quoting a final price, diagnosing a problem or confirming a booking. Agree the boundary in ordinary business language before connecting any tools.
- What question or task is the assistant there to handle?
- Which approved information may it use?
- When should it stop and involve a person?
- What actions, if any, may it carry out?
- Who will check and update it after launch?
A natural answer can still be wrong
Generative systems can produce plausible but incorrect statements. NIST’s Generative AI Profile treats confabulation as a risk to manage. For a business chatbot, a polished tone is therefore a poor substitute for checking whether the answer is supported by current information.
Consider a fictional retailer whose delivery policy changes before Christmas. If the assistant continues to promise the old cutoff, it creates work and disappointment. A useful implementation needs a maintained information source, a clear way to retire outdated material and tests for the questions most likely to matter.
The cost of a wrong answer should influence the scope. An incorrect link is usually easier to recover from than an invented safety instruction or a promise about a refund. Sensitive or consequential questions need an appropriate human route rather than a more confident-sounding answer.
Make the handover part of the design
Customers should be able to reach your team without negotiating with the bot. Give the assistant a clear handover route and explain what happens outside staffed hours. If information is passed to a person, make the summary useful and avoid asking the visitor to repeat everything unnecessarily.
The assistant should also be identifiable as automated. A visitor should understand whether they are talking to a person and whether a response confirms an action. “I can help you request an appointment” is different from “your appointment is booked”. The wording must match the actual workflow.
Our AI and automation work looks at that whole journey. The interface, underlying information and people handling the next step need to work together.
Understand the information being collected
A chat box encourages people to type freely, including details you may not need. Consider what the provider stores, how access works and what users are told. The ICO’s AI guidance is an important starting point when personal information is involved; the right arrangements depend on your particular use.
A practical design might steer someone towards a secure account area for an order query, rather than inviting payment information into an open conversation. It might also keep general product guidance separate from private customer records. Ask your supplier to explain the data flow, rather than relying on an assurance that a tool is simply “secure”.
Measure useful outcomes, including the extra work
The number of conversations started tells you that people clicked a feature. It does not tell you whether they got a correct answer or became a good enquiry. Decide what a useful outcome looks like before a pilot begins.
For example, you could review whether visitors found an appropriate product page, whether handed-over enquiries contained useful information and how many answers needed correction. Compare similar periods and account for changes in traffic or seasonal demand. Keep customer experience and the quality of the next step alongside simple volume.
Include the time spent reviewing answers and maintaining information. A subscription that appears inexpensive may require considerable staff attention. Equally, an assistant that reliably handles repetitive questions could be worthwhile even without an immediate increase in sales. The business case should reflect your actual workload.
Run a small pilot before expanding
Begin with approved information and a limited task. Test straightforward questions, ambiguous wording, outdated information and requests outside the scope. Ask colleagues who were not involved in building it to try realistic scenarios. Include mobile and keyboard use in the review.
A connected feature may need development and integration work as well as a chat interface. Keep a way to switch it off while leaving normal contact routes available, and agree who responds if the provider or an integration fails.
Good ongoing website care includes checking that the feature still serves its purpose. If you want to assess a chatbot idea, bring us the questions your customers ask. We can help establish whether a conversational tool is useful, or whether a clearer page would do the job better.