Build the context, choose the boundaries and test the experience before customers speak to your assistant.
A clear path from your first workspace to your first customer conversation.
Link the business systems your assistant needs.
Set its knowledge, permissions and greeting.
Test your setup before enabling live calls.
Learn from conversations and give your team time back.
This example shows the intended experience for an order enquiry once the live voice and store integrations are available.
Explore the integration roadmapA little structure helps your assistant give a consistent response.
Write a concise description of your business, the services you offer and the kinds of enquiries you want to handle. Include opening hours and the public contact route your callers should use when they need a person. Keep information consistent across your website and approved knowledge.
Give each draft agent one recognisable purpose. Decide how it introduces itself, how much detail it should provide and when it should ask a clarifying question. An agent should be able to explain the limits of its role without confusing the caller.
Start with common questions and current policies. Use short articles with descriptive titles, and review them before approval. Do not include provider secrets or information that the assistant should never reveal during a conversation.
Real callers do not always have every detail ready. They may ask several questions at once, give an incorrect reference or change their request halfway through. Preparing those examples helps you evaluate whether the instructions are clear.
The current application supports draft configuration. Live audio testing, authorised tool execution and human transfer still need the voice service. Treat the checklist as preparation for that testing, rather than evidence that saving a draft activates a telephone number.
Use a representative question with all the information the workflow expects. Check that the intended answer is useful, understandable and grounded in approved knowledge.
Consider missing references, incomplete identity checks and unclear requests. The intended behaviour should ask for the right detail or explain a next step, rather than guess.
Include requests beyond the agent’s role. A safe outcome is a clear explanation and a human route; live permissions must prevent the tool from performing an unauthorised task.
Build an assistant around the way your business works.