A customer asks whether a product is available, what a service includes or when an order will arrive. The assistant needs more than a persuasive reply: it needs the right source, permission to use it and a safe next step when the answer is missing. Start your project with these decisions.
01 — Define one useful job
Choose an initial task with a clear owner, such as answering questions about an approved service catalogue. Write down what the assistant may explain, which details require an authenticated customer and which decisions remain with staff. A product enquiry and a refund approval should not share the same permissions.
02 — Give every answer a dependable source
Gather current service descriptions, product records, operating hours and approved policies. Assign an owner and a review date to each source. Resolve conflicting versions before connecting them. If the information is missing or outdated, the assistant should say so and offer a human handover; a fluent answer is not evidence of accuracy.
03 — Test the conversations your customers actually have
Build a small test set in Arabic and English, including short questions, spelling mistakes, mixed-language product names and questions outside the agreed scope. Check factual accuracy and access boundaries separately. Try requests for another customer’s order and instructions to ignore company rules. Review failures before widening access.
04 — Design the handover before launch
Decide who receives unresolved enquiries, which details travel with the request and how the customer receives a reference. Agree a review routine for unsuccessful conversations. Begin with a limited pilot and keep a way to disable actions while preserving a contact route. Risk management and human oversight are addressed in the NIST reference below; this checklist is our practical project-planning approach.
Bring these to a discovery meeting
- The first customer task and the person responsible for it
- Approved knowledge sources and any restricted data
- Ten representative Arabic and English questions
- The team and channel for human handover
Reference & further reading: NIST: Generative Artificial Intelligence Profile