Assistive AI helps a person by suggesting, drafting or flagging, while the person makes the decision and takes the action. Autonomous AI carries out tasks itself without waiting for approval, within rules set by the business. In practice it is a sliding scale rather than two boxes.
The right level depends on the risk of the task. Low-risk, easily reversed tasks such as sending a booking confirmation suit autonomy. Higher-risk tasks such as large refunds, pricing changes or public replies to complaints call for an approval step or a person making the decision.
In retail and hospitality, a common path is to begin with assistive tools such as drafted emails or suggested stock orders, then allow autonomy for narrow, proven tasks like answering opening-hours questions or confirming bookings. Over time, the autonomy boundary can be widened as results are tracked.
Problems arise when autonomy is granted before the AI has been tested or without a way to see what it did. Ask vendors what limits can be set, whether actions can be reversed and how you will be alerted to anything unusual. Set spending, refund and message caps.
It matters because choosing the right level is how a business gets benefit from AI without handing over decisions it is not ready to stand behind.