Before you start
Choose a task, not just a tool
Start with something employees already do. Describe the input, the work, and the result someone needs. A recurring draft, a summary of authorized information, or preparation for a business conversation may be easier to evaluate than an open-ended instruction to “use more AI.” Ask the people doing the task where the friction really sits.
Write down how the task works today
Understand the current process before changing it. Who supplies the information? Where does it live? Who decides whether the work is correct? Note time spent, common corrections, and exceptions. This creates a practical comparison for the pilot and may reveal that a simpler process or integration is enough.
Prepare the information and the account
Confirm the approved business tool, account or workspace, licensing, and permissions. Decide which information can be used for the pilot and what must be excluded. Review connected services and data sources as well as the assistant itself. A product name is not a substitute for reviewing the specific setup.
Train a small group with a real example
Choose a limited group and an authorized example of their actual work. Teach them to explain the task, provide useful context, and inspect the result. Discuss what an incomplete or incorrect response looks like. Make the route for questions clear rather than expecting every employee to work it out alone.
Keep approval in the workflow
Decide who reviews the output and what they check before it is sent, published, or used to take an action. Start with assistance that a person can inspect. If a later workflow includes automation, agree on permissions, limits, exception handling, and the actions that require human approval.
Review what happened before expanding
Compare the pilot with the original process. Did it reduce work, improve consistency, or simply move effort into checking and correction? Ask employees which parts were helpful. Expand only after the task, information boundaries, training, and support are understood. Training, workflow changes, and a custom app can then be scoped around the remaining need.