What does AI do when it does not know the answer?

A language model is trained to write likely text, not to check whether that text is true. When a customer asks about something the model does not know, such as the delivery time of a specific product or an exception in your returns policy, it will often still write a convincing reply. This is called hallucination, and it is not a rare fault but a property of how these models work. You can reduce it by letting the model work only with information you provide, such as your own knowledge base and the details of an order. You cannot rule it out completely. For customer service this means that an AI that sends replies on its own will, sooner or later, send something wrong in your name.

Linedesk365 bases a draft on your team's earlier replies and the data in your webshop. Numbers such as amounts and order details are read from WooCommerce, read-only, never generated by the model.

A draft can therefore still contain a mistake when there was no good source. That is why nothing reaches a customer before a team member has read and approved it. That check is not optional: automatic sending is off.

Checking a draft takes less time than writing a reply from scratch, especially for questions that come in every day. The work shifts from typing to reviewing. It helps when the email that really needs attention is at the front. In Linedesk365 every customer email gets a priority and, where possible, a topic, and urgent complaints go to the top of the queue.

Automating a whole category should only be considered once its quality has been measured. In Linedesk365 automatic sending is off, and the software only proposes it for a category once your team demonstrably sends its drafts there almost unchanged.