Start with the people asking for help, the decision they need to make and the information available. A clearer page, ordinary search or a fixed rule may solve the need without generated text. Where AI is useful, discovery identifies a bounded task and a suitable first audience. We review approved source samples and ordinary questions before proposing a pilot or broader implementation.
The selected work can include source preparation, retrieval or conversation design, an interface, provider connection, expressly authorised system actions and evaluation. The scope states the actual information boundary, users, supported source types and human route. Provider accounts, costs, data permissions and operating access are reviewed with their owners. A demonstration using fixtures and a provider-backed check answer different questions and are reported separately.
Quality review includes the task’s limits. Some questions can be supported fully, some only in part, and others not at all. We agree how that distinction appears and what the person can do next. The implementation should not treat answering a question as permission to make a booking, change a record or commit the business. Those actions, if required, need their own reviewed tools, rules and checks.
