01
Choose a specific starting task
Identify a repeated question or workflow that is narrow enough to examine. Describe the information the assistant needs and the kind of response that would help. An internal knowledge assistant and a customer-support assistant may require different access, interfaces, and review rules.
02
Connect the right source material
When an answer depends on approved information, make the source collection part of the design. Consider how documents are selected, updated, retrieved, and shown to the user. Retrieval-augmented generation can connect a language feature to relevant context, but the retrieval and the answer still need evaluation.
03
Separate an answer from an action
Preparing a suggested response is different from sending it or changing a record. Define which actions are permitted, which require approval, and which should be escalated. Give users a clear way to review, correct, or decline the proposed next step.
04
Evaluate the boundaries as well as the happy path
Use representative examples, missing-context cases, and questions the assistant should not answer. Check what happens when a source is unavailable or the input is unclear. Agree on human handover and fallback behavior before connecting the workflow to a production system.