From recurring tickets to dependable answers
A website chatbot for customer service needs reliable sources for the answers it gives. A retailer gets delivery questions every week. The shipping page lists normal lead times. Freight deliveries, remote locations and public holidays have different conditions. A website assistant becomes useful when it recognises these distinctions and asks for missing details. Start with a set of resolved enquiries. Remove personal information and group the questions by what customers actually want to know. Ten versions of a Saturday delivery question may need one reliable source. A question about a specific order requires a different kind of access and dependable identification.
For an initial Chatbase trial, select one subject area and appoint a content owner. Agree on a model answer for the common question, a clarifying question for exceptions and a handover for an individual enquiry. Test a question based on an incorrect assumption as well. If someone claims free express delivery was promised, the assistant should refer to the applicable terms. A friendly response must avoid creating fresh commitments. Record which version of the shipping conditions the answer relies on.
Set a useful scope for the assistant
Start where answers are frequent, stable and supported by evidence. Opening hours, basic shipping conditions or preparation for a consultation may be suitable. Suitability depends on your actual sources. If hours vary by branch, the assistant needs a location first. Where enterprise accounts have separate terms, a general answer must avoid implying it covers every contract. Before configuring anything, identify one question within the initial scope and one that requires personal review.
Expand scope once you can observe how visitors phrase their questions. Internal terminology often differs from customer language. Someone may ask about “the large machine” while your data sheets contain only model numbers. A short clarifying question may therefore help more than loading additional documents. Review whether frequently unanswered questions belong within the assistant’s purpose. Sometimes clearer website navigation is the more appropriate improvement.
Maintain answers responsibly
Maintain a small set of recurring review questions. Cover an ordinary case, an exception, a contradiction and a question without a reliable source. After a material content change, ask the same questions again. Compare the answers with the approved position. Identical wording is unnecessary. The correct rule, relevant qualification and a useful next action are what matter.
Investigate the basis of an incorrect answer first. Is the source obsolete, a condition missing or the wording ambiguous? Polishing one answer seldom fixes these causes permanently. Agree who may change a source and how quickly the updated information reaches the assistant. Then review actual handovers to support. A shorter conversation alone cannot distinguish a resolved issue from a visitor who gave up.
How does an unresolved question reach customer service?
Walk through the handover with someone who actually handles support. A visitor needs to know whether a parcel will arrive by Friday. General delivery times do not answer their question. Which contact route will you offer, what order reference does support require and when will someone read the request? Follow the journey on a phone. If it ends at a form asking for the same explanation again, you have found a specific improvement to make.
Consider what happens after office hours. A reply promising immediate help when nobody reads enquiries until the next working day shifts the problem into an inbox. Write a short handover that sets a realistic expectation. Check with support whether context can be passed on and whether the visitor understands which information the next person receives. Judge the quality of the whole conversation only after you have seen the transition work.
Three tests for your pilot
Phrase the same question differently
Try “Can my parcel arrive on Saturday?” and “Do you deliver at weekends?” against the same approved terms. Both replies should express the same rule. Check whether regional exceptions survive. Record differences alongside the source to distinguish a knowledge issue from a problem in the answer.
Expose a conflict between sources
In an isolated test setup, supply two sample texts with different return periods. Observe whether the conflict becomes visible. Decide which content is authoritative and who removes obsolete text. Repeat the original question after cleanup. A plausible answer is difficult to control without a dependable source.
Hand over an individual enquiry
Ask about the whereabouts of a particular order without connecting order data. A useful response identifies the support route and explains which details are needed there. Check with support that the handover reaches a monitored queue. Sending people to an unattended mailbox merely shifts the waiting time.
Preparing your team to use the tool
Prepare the knowledge behind the workflow
Choose a question regularly answered in two parts of the business. Ask the responsible people independently for an answer and its supporting source. This small comparison reveals whether the bottleneck is writing or knowledge. If one person remembers the current rule and another relies on an obsolete PDF, faster writing would propagate uncertainty. Designate an authoritative source and someone accountable for changes. Record the date from which a new version applies.
For a first trial, a small maintained collection is often easier to evaluate than the entire document estate. Choose material with a clear origin and intended audience. Identify older versions so their statements cannot be mistaken for current approval. Ask a second person to derive the answer using only the selected material. Any assumptions they have to add belong on the issue list. A fluent answer does not establish that the underlying information was present. Approval should cover the text, its provenance and the circumstances in which it may be used. Keep these together when handing the work to another team.
Design a conversation that knows its limits
Choose a question customers often ask just before buying. For a fictional machinery business, it might concern whether a spare part fits a particular model. The website lists product ranges, but some older machines have been modified. A recommendation based on the original model number alone misses the very case that could cost a buyer money. First ask the service team what evidence permits a definite answer and at what point a person needs to take over.
Write three enquiries in customers’ words. One includes the necessary details, one gives only the informal name of the machine and the third mentions a modification. Read the replies as someone unfamiliar with the product. Does the assistant explain why it needs a detail? Can the customer locate the number on the machine? A follow-up question helps only if the visitor can actually answer it. Otherwise they need a different route, such as a clearly identified technical service contact.
For the difficult case, design a practical handover. The conversation should establish what is already known and what remains uncertain. The buyer needs to understand the next action expected of them. The service desk needs enough context to ask a focused question. Check whether your actual combination of website, assistant and ticket system can transfer those details. If the visitor needs to contact someone separately, explain that step in language a customer can use.
Check the assistant after a product range changes
Imagine that a product line is replaced. The website has an updated product page, while an older manual remains available to existing customers. Both documents can be correct and still suggest different answers to the same question. The year of manufacture or serial number determines which statement applies. When maintaining sources, discuss who still needs each document. Simply removing everything old would make the experience worse for owners of earlier machines.
After the change, repeat the questions you used before launch and add an enquiry with a deliberately misspelled product name. For each attempt record the answer, the source quoted and the next action offered. A broken reference calls for a different correction from a failure to distinguish two models. Also establish who notices an incorrect recommendation on the public site and how you will handle that enquiry until the underlying issue is fixed.
Conversation counts alone tell you little. A long exchange might represent careful assistance or an assistant repeatedly asking the same question. Where your procedures permit it, follow a few anonymised exchanges from the opening message to handover. Can the customer now make the relevant decision or reach somebody qualified to help? That is a practical basis for deciding which questions belong on the website and which should lead to personal advice earlier.
Learn from unanswered questions
Pick five questions that should lead to a handover. Read the final exchange in each conversation and identify what stopped a useful answer. A missing public specification may justify a better product page. A question about the progress of a personal order belongs with customer service. Two conflicting delivery times need a purchasing decision before anyone edits the help centre. Assign each case to someone who can resolve that particular obstacle. After the change, repeat the original question in the customer's own words and try a shorter version with a spelling mistake. Record whether the response now points to the correct information or offers an appropriate way to reach a person. Both can be successful outcomes.
Turning internal documents into public answers
Many support answers start life in an internal document. Alongside useful customer information it may contain goodwill limits, personal contacts or purchasing prices. Produce an approved public version for the website. Agree with the document owner when a change to the internal original should trigger a review of the public answer. This keeps knowledge maintenance connected to the chatbot after launch.
Read the Google Workspace guide