Practical AI · 3-minute read
A customer asks whether a late return fee can be waived. The chatbot replies, ‘Absolutely!’ Somewhere in the office, a policy document has begun quietly screaming.
A helpful tone is useful. An invented promise is expensive. A customer-facing chatbot needs a clear boundary between answering a question, asking for clarification and bringing in a person.
Give it a small job with approved sources
Copperline Equipment Rental is fictional. Its proposed chatbot helps customers find opening hours, public collection instructions and the correct contact route. It does not negotiate prices, waive fees, confirm live availability or provide equipment safety advice.
For each allowed topic, identify the approved source and its owner. Check which version is current and how changes reach the chatbot. If the system cannot access a current price or availability record, it should explain that limitation and offer the right contact route.
Source links help a reviewer trace an answer. They do not, by themselves, prove that the answer accurately represents the source.
Decide when to answer and when to hand off
Answer when current approved information directly supports the response.
Clarify when a missing detail, such as the collection branch, changes the answer.
Hand off when sources conflict, information is unavailable, an exception is requested or the topic is outside scope.
Using these approved public FAQs and topic boundaries, draft customer questions that test answer, clarify and handoff behavior. Include conflicting opening hours, an outdated price, an exception request and a question outside scope. For each test, identify the supporting source and expected response type. Do not invent policy, account information or a promise that staff will respond within an unapproved time.
Use the prompt to prepare tests, then have the business owner confirm the expected answers. A model grading its own guesses is not an independent quality department.
Make the handoff a real route
Explain that customers are speaking with an automated assistant. Offer a visible way to reach a person, with accurate hours and expectations. If a conversation summary is transferred, have the customer confirm it and collect only information needed for the request.
The UK Data and AI Ethics Framework emphasizes transparency, meaningful human control and ways to raise concerns. Those principles support a service design in which customers can understand the tool and challenge an unsatisfactory result. [1]
Have someone test the actual path from an unanswered question to the receiving team. Include keyboard use, plain-language questions, spelling variations and different communication needs. ‘Contact support’ is not a handoff if the link leads back to the same chatbot.
Track unsupported promises, failed handoffs and whether questions were actually resolved. Review examples using an approved privacy process. A low transfer rate is not a success if customers simply give up.
Your turn: Which question should your business’s chatbot immediately take to a person?
Source [1]: UK Data and AI Ethics Framework. Supports transparency, human control and routes for concerns. The rental business, conversation boundaries and tests are original editorial guidance, not a verified chatbot configuration.
