AI assistants
Does your website need an AI chatbot?
An AI chatbot can be useful when customers regularly need help finding or applying information, reliable source material exists and someone owns the handoff when the system cannot help. Its value depends on the questions it resolves and the customer experience it creates.
Start by examining the questions customers already ask.
Find out what is creating the questions
Review enquiry forms, support messages, sales calls and search behaviour where available. Group questions by topic, frequency and the consequences of getting an answer wrong.
Different problems call for different improvements:
| Customer need | Useful first response |
|---|---|
| “How much is delivery?” | Make delivery information visible near the buying decision |
| “Which product fits these requirements?” | Provide a comparison guide or guided selection |
| “Where is my order?” | Offer a secure route to order-specific information |
| “Can you make an exception?” | Route the request to an authorised person |
| Several questions across a large knowledge base | Consider a conversational interface using approved sources |
This exercise may identify immediate improvements to product pages and navigation, alongside a legitimate role for a chatbot.
Check the quality of the source material
A chatbot needs dependable information about the products or services it discusses.
For each source, identify its owner, when it was reviewed and what happens when information changes. Resolve contradictions between delivery pages, promotional banners, product descriptions and sales documents.
Some answers require live information. Availability, order status and account-specific details cannot reliably come from a general knowledge document. Those functions need suitable integrations and access controls.
Define the boundaries in ordinary language
A useful scope might say:
The assistant helps visitors understand published services, find relevant guides and prepare an enquiry. It can explain approved delivery information. Requests for contractual commitments, bespoke prices or exceptions go to a person.
Boundaries should shape the behaviour users actually see.
When information is unavailable, the assistant should say so and offer a sensible route forward. Where practical, it should link to the source behind an answer.
Using approved sources reduces avoidable errors, but does not guarantee that every generated answer is correct.
Design the human handoff
A handoff should preserve the useful context already provided.
Tell the visitor what will happen next and pass the relevant question and conversation summary to the responsible team. Ask only for the information needed to handle that request.
Avoid a loop in which the assistant repeatedly suggests help but never provides a usable contact route.
Test the difficult questions
Include:
- A question with an explicit answer in the sources.
- An ambiguous request requiring clarification.
- Conflicting or outdated information.
- A product or service the business does not offer.
- A request for an unauthorised promise.
- A question requiring private customer information.
- Instructions asking the assistant to disregard its boundaries.
Assess accuracy, source support, handoff quality and whether the visitor reaches a useful outcome.
Measure customer value
Useful measures include reviewed answer accuracy, successful resolution, escalation quality, customer feedback and total operating cost.
A conversation ending does not establish that the visitor received a satisfactory answer. Equally, fewer support messages can be misleading if customers are struggling to reach a person.
Peach & Stone’s Answer Layer concept starts with the questions customers need answered, the sources available and the limits of what the system should handle.
Next step: Gather twenty real customer questions. They provide a practical test set for deciding where improved content, guided navigation or a chatbot would help.