01
Choose a real task
Begin with something specific: finding information across approved documents, organising a daily plan or preparing a draft for review. A narrow starting point makes it easier to judge whether the assistant helps.
02
Map what it needs to know
List the information sources it needs and who controls them. Identify what is confidential, how access will be granted and what should never be stored. Treat memory as a deliberate product choice.
03
Decide what it may do
Answering a question, drafting an email and sending that email are different capabilities. Define where the assistant can suggest, where it can prepare and where a person must approve an action.
04
Agree how to evaluate it
Use representative tasks and difficult examples. Review accuracy, usefulness and failure behaviour. A convincing demo is only a starting point; testing shows where the boundaries are.
05
Make the controls understandable
People should be able to understand connected tools, review relevant history and change permissions. The interface should explain failures in a way that helps someone recover.
06
Expand with evidence
Start with the working scope, gather feedback and add capabilities deliberately. Voice, additional integrations and more autonomous actions each introduce new design and operational questions.