An AI assistant on a website or WhatsApp line can answer routine questions at any hour. It can also deliver an incorrect answer with complete assurance. The following five considerations determine which outcome a business obtains.

1. The scope of what it may answer

Define the scope in writing: opening hours, published prices, booking procedures, parking. Any question outside that scope should be referred to a member of staff. An assistant with a narrowly defined remit is considerably more reliable than one instructed simply to assist customers.

2. The source of its answers

A sound assistant draws its answers from the business’s own documents: the price list, the policies, the catalogue. Where no answer can be found there, it should say so and refer the matter onward. It should not speculate.

3. Its behaviour when it lacks an answer

There must be a defined handover to a member of staff, and the customer should be informed that it is taking place. Determine who receives these handovers and how promptly they are expected to respond.

4. Approval of what is sent in the business’s name

Answering a question about opening hours carries little risk. Issuing a quotation, confirming a refund or altering a booking carries considerably more. For any action that commits the business, the assistant should prepare the response and a member of staff should approve it.

5. The handling of customer data

Establish which provider processes the conversations, where the data is stored, and whether it is used to train that provider’s models. A clear answer should be available to all three questions. Businesses serving customers in Australia or India should assess the arrangement against the Privacy Act or the DPDP Act.

Testing before release

Before the assistant is released, arrange for people outside the team to test it as customers would, including with questions it ought not to answer. Review the transcripts. Continue to review them during the first weeks of operation, and adjust the scope defined in the first step as the questions customers actually ask become clear.