Quick answer
A business AI chatbot should answer a defined set of questions from approved information and pass other cases to a person. Decide whether it only answers or can also change bookings or records. Test its limits before giving it wider access.
Choose the questions and information it may use
List approved documents and who keeps them current. Prepare common questions, unclear questions and questions the documents cannot answer. Sometimes the right response is to ask for details or say it does not have an answer.
If it uses customer records, agree permission, consent, data storage time and staff access. Ask how submitted information is used and what the contract allows.
Separate answers from actions
Explaining a policy is different from changing a booking, approving a refund or opening a private order. List each allowed action and the approval it needs.
Require separate permission checks and confirmation before an action. Define when the chatbot must stop or hand the conversation to a person.
Run a small, useful trial
- Use approved documents with a named update owner.
- Test supported, unclear and unsupported questions.
- Include outdated information and private-data requests.
- Record wrong answers and whether staff handovers work.
- Keep unnecessary customer details out of logs and examples.
- Compare model, hosting, monitoring and support costs.
Decide what success means
Measure useful results, such as resolved questions, correct handovers and incorrect answers. Where relevant, check staff time spent. A fluent reply is not proof that the answer or action is right.
Use sample or properly authorised test material. Do not put confidential customer enquiries into a public demonstration.
Your planning worksheet
Use this table to prepare the same brief for each provider. Add the person responsible and the current status for each item. Enter costs only from written quotes, in HKD, with the billing period.
| Item to plan | Details to give the provider | What to check |
|---|---|---|
| Answerable question | Approved source and expected answer | Answer is supported by the approved material |
| Unclear or unsupported question | Missing information and escalation rules | Chatbot asks for details, declines or passes the question to staff |
| Private action | Allowed operation, permission and approval | Unauthorized data/actions remain inaccessible |
Example project scenario
Illustrative planning example; not a completed customer project.
A business tests a chatbot using approved service FAQs. It asks a supported question, one with no answer, one based on an outdated document and one requesting another user’s data. The trial checks whether it recognises its limits and passes cases to staff. No accuracy result is claimed.
Common questions
Should the chatbot answer everything?
No. Agree when it should ask for more information, decline or pass the question to a person. Test those responses as well as helpful answers.
How do we compare proposals?
Give each provider the same approved information and questions. Compare supported answers, limits, staff handover, fees and responsibility for updates.
How is this different from using a public AI chat tool?
A business service needs approved sources, access rules, staff handover, data controls and ongoing checks. Ask how each of these will work for your team.
Download the planning worksheet (CSV)
Find a provider through TechQuotations
TechQuotations connects Hong Kong buyers with IT service providers. Use your brief to discuss the work and compare proposals. You choose the provider and agree the terms directly. A request does not guarantee a quotation.
Explore suitable providers through TechQuotations